Adaptive deformable mirror- dynamics and modular control
Adaptive deformable mirror- dynamics and modular control
Adaptive deformable mirror- dynamics and modular control
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
ÔØÚÓÖÑÐÑÖÖÓÖ<br />
ÝÒÑ×ÒÑÓÙÐÖÓÒØÖÓÐ<br />
ÊÅÄÐÐÒÖÓ
ÔØÚÓÖÑÐÑÖÖÓÖ<br />
ÝÒÑ×ÒÑÓÙÐÖÓÒØÖÓÐ<br />
Proefschrift<br />
ter verkrijging van de graad van doctor<br />
aan de Technische Universiteit Delft,<br />
op gezag van de Rector Magnificus prof.ir. K.C.A.M. Luyben,<br />
voorzitter van het College voor Promoties,<br />
in het openbaar te verdedigen op dinsdag 8 november 2011 om 15.00 uur<br />
door Rogier Martin Lambert ELLENBROEK,<br />
werktuigbouwkundig ingenieur,<br />
geboren te Heerlen.
Dit proefschrift is goedgekeurd door de promotor:<br />
Prof.dr.ir. M. Verhaegen<br />
Samenstelling promotiecommissie:<br />
Rector Magnificus voorzitter<br />
Prof.dr.ir. M. Verhaegen Technische Universiteit Delft, promotor<br />
Prof.dr.ir. M. Steinbuch Technische Universiteit Eindhoven<br />
Prof.dr.ir. J.M.A. Scherpen Rijksuniversiteit Groningen<br />
Prof.ir. R.M. Schmidt Technische Universiteit Delft<br />
Prof.dr. C. Keller Universiteit Utrecht<br />
Dr.ir. P.R. Fraanje Technische Universiteit Delft<br />
Dr.ir. N.J. Doelman TNO<br />
Prof.dr.ir. J.A. Mulder Technische Universiteit Delft, reservelid<br />
Dr.ir. N.J. Doelman heeft als begeleider in belangrijke mate aan de totst<strong>and</strong>koming van het<br />
proefschrift bijgedragen.<br />
This thesis has been completed in partial fulfillment of the requirements of the Dutch Institute<br />
for Systems <strong>and</strong> Control (DISC) for graduate studies. The research described in this<br />
thesis was supported by the Dutch Innovative Research Project (IOP) Precision Technology.<br />
ISBN 978-90-9026363-2<br />
Copyright c○ 2011 by R.M.L. Ellenbroek<br />
All rights reserved. No part of the material protected by this copyright notice may be reproduced<br />
or utilized in any form or by any means, electronic or mechanical, including photocopying,<br />
recording or by any information storage <strong>and</strong> retrieval system, without written<br />
permission of the author.<br />
Printed by CPI Wöhrmann Print Service
ÒÛÓÓÖ<br />
Het was al laat in de avond toen de telefoon ging en Roger me vroeg of ik misschien<br />
geïnteresseerd zou zijn om als een tweede promovendus binnen zijn adaptieve optica<br />
project te werken. Er was echter een kleine maar: dit deel van het project zou worden<br />
uitgevoerd aan de technische universiteit van Delft in plaats van die in Eindhoven, waar ik<br />
op dat moment nog werkte. Hij had al een tijd lang zitten opscheppen over zijn project, dus<br />
het was onmogelijk om het aanbod af te wijzen, al moest ik nog wel even wennen aan het<br />
Delft-stad-van-de-koorballen concept. De uiteindelijke keuze was natuurlijk niet aan hem,<br />
dus ik wil hierbij mijn promotor Michel Verhaegen en Pieter Kappelhof (de toenmalige<br />
leider van het project) hartelijk bedanken voor het vertrouwen om mij voor deze positie<br />
aan te nemen. Bovendien wil ik de mensen van het IOP Precisie Technologie hartelijk<br />
bedanken voor het organiseren van de r<strong>and</strong>voorwaarden waaronder dit project mogelijk<br />
was. Tenslotte zijn er nog een hele sloot mensen aan wie ik voor het bereiken van deze<br />
mijlpaal dank verschuldigd ben en waarvan ik er een paar nog even expliciet wil noemen:<br />
Al snel bleek dat het begeleiden van mij als promovendus niet vanzelf gaat en veel tijd<br />
kost. Ik wil daarom Niek en Michel van harte bedanken voor hun tijd en moeite om mij<br />
bij te sturen en op het rechte pad te houden. Niek, bedankt ook om zelfs na de officiële<br />
eind-datum van het project deze tijd te blijven vinden.<br />
In het project moest er door TNO ook elektronica worden ontwikkeld, hetgeen deels<br />
aan Emdes (later onderdeel van QPI) werd uitbesteed. Ik wil bij deze Paul Keijzer, Ton<br />
Lommen, Wout van de Maden en Gerhard Hein bedanken voor de inzet en samenwerking<br />
die uiteindelijk heeft geleid tot een werkend prototype van een complex systeem.<br />
Verder wil ik Rudolf Mak, Kees van Berkel en William de Bruijn bedanken voor hun<br />
doorzettingsvermogen in het nog efficiënter maken van de ontwikkelde elektronica.<br />
Kitty, bedankt voor je gezelligheid en hulp en dat je tot het bittere eind mijn rots in de<br />
br<strong>and</strong>ing van DCSC was. Jelmer, ik weet niet hoe ik je morele ondersteuning, gezelligheid,<br />
etc. als kamergenoot voor 4 jaar in deze lofzang voldoende uit de verf kan laten komen.<br />
Ik ben blij dat we elkaar nu ’bijna buren’ mogen noemen en nog op allerlei manieren meer<br />
van elkaar gaan zien. Karel, bedankt dat je me – in ruil voor je bijdehante antwoorden op<br />
al mijn werk-gerelateerde vragen – bij onze wekelijkste squash sessie zo vaak liet winnen.<br />
De kaasfondue zal ik nooit vergeten en hopelijk komen daar in de toekomst nog allerlei<br />
dingen bij. Sjoerd, je was een onverwachte vriend die op de belangrijkste momenten voor<br />
me klaar stond en toen zover mogelijk weg ging wonen. Vooral die laatste actie blijf ik<br />
jammer vinden.<br />
Roger, is er niks te vertellen dat je niet al weet, maar dat belet me niet om hier een<br />
v
vi Dankwoord<br />
paar woorden neer te zetten. Ondanks de woede-uitbarstingen, de scheld-kanonnades, de<br />
haren-trekkerij, de machtsstrijd, de afgunst en het haantjesgedrag van <strong>and</strong>eren, zijn we<br />
goeie vrienden gebleven en hebben daar zelfs het ’bijna-familie’ en het ’collega’ gevoel<br />
aan toegevoegd. Ik vind het bijzonder dat je in zoveel aspecten een deel van mijn leven<br />
uitmaakt en ik ben dan ook blij dat we elkaar bij Mapper nog steeds in al die hoedanigheden<br />
tegenkomen.<br />
Pap en mam, voor jullie was mijn vertrek waarschijnlijk best een beetje plotseling en was<br />
mijn promotie project zowel inhoudelijk als praktisch een beetje een ver-van-jullie-bedshow.<br />
Toch heeft jullie warmte en mentale steun er veel aan bijgedragen dat dit boek er<br />
uiteindelijk toch is gekomen.<br />
Martin, dankjewel voor je zonnige gezicht dat alles altijd weer goedmaakt.<br />
En Mirjam, de laatste en dikste knuffels zijn voor jou. Voor alle keren dat ik ’nee’ zei,<br />
gevolgd door een zin met het woord ’proefschrift’. Voor alle keren dat je zei ’e finiscila!’<br />
of ’basta col perfezionismo’. Voor alles en voor altijd: ti amo.
ÓÒØÒØ×<br />
Dankwoord v<br />
Contents x<br />
Summary xi<br />
Samenvatting xiv<br />
Nomenclature xvii<br />
Acronyms xxiv<br />
1 Introduction 1<br />
1.1 Notes on notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2<br />
1.2 Astronomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2<br />
1.3 <strong>Adaptive</strong> optics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5<br />
1.3.1 Atmospheric turbulence . . . . . . . . . . . . . . . . . . . . . . . 5<br />
1.3.2 The wavefront sensor . . . . . . . . . . . . . . . . . . . . . . . . . 6<br />
1.3.3 The wavefront corrector . . . . . . . . . . . . . . . . . . . . . . . 8<br />
1.3.4 Optical configurations . . . . . . . . . . . . . . . . . . . . . . . . 11<br />
1.3.5 The <strong>control</strong> system . . . . . . . . . . . . . . . . . . . . . . . . . . 12<br />
1.4 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14<br />
1.4.1 The wavefront corrector . . . . . . . . . . . . . . . . . . . . . . . 14<br />
1.4.2 The <strong>control</strong> system . . . . . . . . . . . . . . . . . . . . . . . . . . 15<br />
1.5 Distributed <strong>control</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17<br />
1.5.1 Distributed <strong>control</strong> for AO . . . . . . . . . . . . . . . . . . . . . . 20<br />
1.6 Problem formulation <strong>and</strong> organization of this thesis . . . . . . . . . . . . . 21<br />
1.7 Scientific contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22<br />
2 Design requirements <strong>and</strong> design concept 25<br />
2.1 Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26<br />
2.1.1 Atmospheric turbulence . . . . . . . . . . . . . . . . . . . . . . . 26<br />
2.1.2 The Kolmogorov turbulence model . . . . . . . . . . . . . . . . . 26<br />
2.2 Error budget . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30<br />
2.2.1 The fitting error . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31<br />
2.2.2 The temporal error . . . . . . . . . . . . . . . . . . . . . . . . . . 31<br />
2.2.3 Error budget division . . . . . . . . . . . . . . . . . . . . . . . . . 32<br />
2.3 Actuator requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33<br />
2.4 Control system <strong>and</strong> electronics requirements . . . . . . . . . . . . . . . . . 35<br />
vii
viii Contents<br />
2.5 The design concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38<br />
2.5.1 The <strong>mirror</strong> facesheet . . . . . . . . . . . . . . . . . . . . . . . . . 38<br />
2.5.2 The actuator modules . . . . . . . . . . . . . . . . . . . . . . . . . 39<br />
2.5.3 The <strong>control</strong> system <strong>and</strong> electronics . . . . . . . . . . . . . . . . . . 40<br />
2.5.4 The base frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41<br />
3 Efficient <strong>control</strong> for AO: concepts <strong>and</strong> challenges 43<br />
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44<br />
3.2 Existing <strong>control</strong> methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 44<br />
3.2.1 Generic problem statement . . . . . . . . . . . . . . . . . . . . . . 44<br />
3.2.2 Traditional approach . . . . . . . . . . . . . . . . . . . . . . . . . 45<br />
3.2.3 More generic <strong>control</strong>ler designs . . . . . . . . . . . . . . . . . . . 48<br />
3.3 Scaling problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49<br />
3.3.1 Computational dem<strong>and</strong> . . . . . . . . . . . . . . . . . . . . . . . . 49<br />
3.3.2 Practical problems . . . . . . . . . . . . . . . . . . . . . . . . . . 54<br />
3.4 Distributed <strong>control</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54<br />
3.4.1 Hardware considerations . . . . . . . . . . . . . . . . . . . . . . . 55<br />
3.4.2 Control considerations . . . . . . . . . . . . . . . . . . . . . . . . 56<br />
3.5 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56<br />
3.6 Possible solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58<br />
3.6.1 Phase reconstruction through analog electronics . . . . . . . . . . . 58<br />
3.6.2 A distributed disturbance model . . . . . . . . . . . . . . . . . . . 58<br />
3.6.3 Iterative distributed phase reconstruction . . . . . . . . . . . . . . 59<br />
3.6.4 Recursive adaptive distributed reconstruction <strong>and</strong> prediction . . . . 62<br />
3.6.5 Local, identical influence functions . . . . . . . . . . . . . . . . . 66<br />
3.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66<br />
4 Data driven distributed <strong>control</strong> 69<br />
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70<br />
4.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71<br />
4.3 Problem description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71<br />
4.4 Design approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72<br />
4.4.1 Parametrization of the distributed <strong>control</strong>ler . . . . . . . . . . . . . 73<br />
4.4.2 Internal model <strong>control</strong> . . . . . . . . . . . . . . . . . . . . . . . . 74<br />
4.5 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75<br />
4.6 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77<br />
4.6.1 Gershgorin’s circle theorem . . . . . . . . . . . . . . . . . . . . . 78<br />
4.7 Identification procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79<br />
4.7.1 Optimization criterion <strong>and</strong> approach . . . . . . . . . . . . . . . . . 79<br />
4.7.2 A two-stage approach . . . . . . . . . . . . . . . . . . . . . . . . 79<br />
4.7.3 Algorithm summary . . . . . . . . . . . . . . . . . . . . . . . . . 84<br />
4.7.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84<br />
4.8 Simulation <strong>and</strong> breadboard results . . . . . . . . . . . . . . . . . . . . . . 85<br />
4.8.1 Performance measures . . . . . . . . . . . . . . . . . . . . . . . . 85<br />
4.8.2 Breadboard data . . . . . . . . . . . . . . . . . . . . . . . . . . . 86<br />
4.8.3 Artificial data set . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Contents ix<br />
4.9 Conclusions <strong>and</strong> future work . . . . . . . . . . . . . . . . . . . . . . . . . 89<br />
5 The variable reluctance actuator 91<br />
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92<br />
5.2 The single actuator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94<br />
5.2.1 The actuator membrane suspension . . . . . . . . . . . . . . . . . 94<br />
5.2.2 The electromagnetic force . . . . . . . . . . . . . . . . . . . . . . 96<br />
5.2.3 A static actuator model . . . . . . . . . . . . . . . . . . . . . . . . 101<br />
5.2.4 A dynamic actuator model . . . . . . . . . . . . . . . . . . . . . . 104<br />
5.2.5 Measurements <strong>and</strong> validation . . . . . . . . . . . . . . . . . . . . 109<br />
5.2.6 Sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 111<br />
5.2.7 Lessons learned . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116<br />
5.3 The actuator module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117<br />
5.3.1 Measurement results . . . . . . . . . . . . . . . . . . . . . . . . . 120<br />
5.3.2 Power dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . 121<br />
5.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122<br />
5.5 Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123<br />
6 Electronics 125<br />
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126<br />
6.2 Driver electronics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126<br />
6.2.1 Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126<br />
6.2.2 Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127<br />
6.3 Communication electronics . . . . . . . . . . . . . . . . . . . . . . . . . . 130<br />
6.4 Implementation <strong>and</strong> realization . . . . . . . . . . . . . . . . . . . . . . . . 131<br />
6.4.1 Pulse Width Modulation (PWM) implementation . . . . . . . . . . 131<br />
6.4.2 Field Programmable Gate Array (FPGA) implementation . . . . . . 134<br />
6.4.3 The ethernet to Low Voltage Differential Signalling (LVDS) bridge 136<br />
6.5 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136<br />
6.6 Evaluation of <strong>control</strong> aspects . . . . . . . . . . . . . . . . . . . . . . . . . 138<br />
6.7 Testing <strong>and</strong> validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141<br />
6.7.1 Communications tests . . . . . . . . . . . . . . . . . . . . . . . . 141<br />
6.7.2 Parasitic resistance measurements . . . . . . . . . . . . . . . . . . 142<br />
6.7.3 Actuator system validation . . . . . . . . . . . . . . . . . . . . . . 142<br />
6.7.4 Nonlinear behavior . . . . . . . . . . . . . . . . . . . . . . . . . . 147<br />
6.8 Power dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148<br />
6.8.1 Optimizing the FPGA power efficiency . . . . . . . . . . . . . . . 149<br />
6.8.2 Cooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151<br />
6.9 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152<br />
7 System modeling <strong>and</strong> characterization 155<br />
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156<br />
7.2 Deformable Mirror (DM) integration . . . . . . . . . . . . . . . . . . . . . 156<br />
7.2.1 Integration of the 61 actuator <strong>mirror</strong> . . . . . . . . . . . . . . . . . 156<br />
7.2.2 Integration of the 427 actuator <strong>mirror</strong> . . . . . . . . . . . . . . . . 156<br />
7.3 Static system validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
x Contents<br />
7.3.1 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159<br />
7.3.2 Measurements <strong>and</strong> results . . . . . . . . . . . . . . . . . . . . . . 164<br />
7.3.3 Power dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . 173<br />
7.4 Dynamic system validation . . . . . . . . . . . . . . . . . . . . . . . . . . 176<br />
7.4.1 Dynamic modeling . . . . . . . . . . . . . . . . . . . . . . . . . . 176<br />
7.4.2 System identification . . . . . . . . . . . . . . . . . . . . . . . . . 180<br />
7.4.3 Modal analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182<br />
7.5 Discrete time <strong>control</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188<br />
7.5.1 A note on distributed <strong>control</strong> . . . . . . . . . . . . . . . . . . . . . 190<br />
7.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190<br />
8 Conclusions <strong>and</strong> recommendations 193<br />
8.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194<br />
8.2 Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198<br />
Appendices 200<br />
A Shack-Hartmann spot positioning . . . . . . . . . . . . . . . . . . . . . . . 201<br />
B On using local passivity to enforce global stability . . . . . . . . . . . . . . 203<br />
C Fourier series of a PWM signal . . . . . . . . . . . . . . . . . . . . . . . . 206<br />
D The LVDS protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208<br />
E The UDP protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211<br />
F Spatial variation of actuator properties . . . . . . . . . . . . . . . . . . . . 213<br />
G Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214<br />
Bibliography 216<br />
Curriculum vitae 229
ËÙÑÑÖÝ<br />
ÔØÚÓÖÑÐÑÖÖÓÖ<br />
ÝÒÑ×ÒÑÓÙÐÖÓÒØÖÓÐ<br />
The refractive index of air varies a.o. with temperature, humidity, pressure <strong>and</strong> the CO2<br />
concentration. Due to atmospheric turbulence this refractive index varies both in space <strong>and</strong><br />
in time, leading to aberrations in images of light having passed though it. These aberrations<br />
limit the achievable resolution of optical telescopes such that the quality of their images<br />
is no longer diffraction limited. An <strong>Adaptive</strong> Optics (AO) system is a means to recover<br />
the diffraction limited quality of the images. This can be achieved a.o. by reflecting the<br />
incoming light on a Deformable Mirror (DM) that adapts its shape to the wavefront of this<br />
light such that some norm of the residual wavefront after reflection is minimal.<br />
Such a reflection based AO system consists of three main components: a WaveFront<br />
Sensor (WFS), a DM <strong>and</strong> a <strong>control</strong> system. In this thesis novel designs are considered for<br />
the latter two components, primarily aimed at the 8m class of telescopes in visible light.<br />
The WFS is assumed to be of the Shack-Hartmann type <strong>and</strong> not further investigated. As<br />
astronomers want images of ever fainter celestial objects, larger <strong>and</strong> larger telescopes are<br />
foreseen to increase both resolution <strong>and</strong> light gathering power. Therefore, an important<br />
design driver is the extendibility to a larger number of Degrees Of Freedom (DOF) in<br />
combination with a low power consumption to prevent the need for active cooling systems.<br />
The main requirements for the DM <strong>and</strong> <strong>control</strong> system are derived for atmospheric<br />
conditions that are typical for telescope sites. The spatial <strong>and</strong> temporal properties of the<br />
atmosphere are modeled by the spatial <strong>and</strong> temporal spectra of the Kolmogorov turbulence<br />
model <strong>and</strong> the frozen flow assumption. The main sources for the residual wavefront<br />
aberrations – important for the DM design – are identified as the fitting error <strong>and</strong> the<br />
temporal error that are caused by a limited number of actuators <strong>and</strong> a limited <strong>control</strong><br />
b<strong>and</strong>width respectively. A choice for the number of actuators <strong>and</strong> the <strong>control</strong> b<strong>and</strong>width is<br />
made that leads to a desired optical quality after correction in terms of a Strehl ratio of 0.85<br />
<strong>and</strong> for which the specifications are considered feasible for both areas of expertise.<br />
The requirements on the DM actuators are based on a visible wavelengthλ of550nm. This<br />
leads to a required pitch of 3 to 6mm, a total actuator stroke of ±5.6µm, an inter-actuator<br />
stroke of ±0.36µm <strong>and</strong> a resolution of 5nm. From the resulting actuator density <strong>and</strong> the<br />
desire for active cooling follows a maximum allowed power dissipation per actuator of<br />
1mW. The corresponding requirement on the <strong>control</strong> b<strong>and</strong>width is 200Hz, from which<br />
additional requirements are derived for the sampling frequency <strong>and</strong> the first DM resonance<br />
frequency. From a budgeting of the phase margin both are required to be higher than 1kHz.<br />
The phase delay <strong>and</strong> latency of the driver <strong>and</strong> communication electronics at the b<strong>and</strong>width<br />
frequency must be small compared to the sampling time.<br />
When designing <strong>control</strong>lers for AO systems with increasingly large numbers of DOFs<br />
xi
xii Summary<br />
this will lead to a high computational load that must properly h<strong>and</strong>led. Without efficient<br />
algorithms, the dem<strong>and</strong>ed computational power increases quadratically with the number<br />
of degrees of freedom <strong>and</strong> thus to the fourth power in the telescope’s aperture diameter.<br />
A <strong>modular</strong>ly distributed <strong>control</strong> system is proposed to extend the <strong>control</strong> system towards<br />
large numbers DOF without problems with the computational loads. It is shown that<br />
the distributed <strong>control</strong>ler architecture complicates the wavefront reconstruction step that<br />
is commonly present in most AO <strong>control</strong> systems due to the spatial <strong>dynamics</strong> of the<br />
Shack-Hartmann WFS used. The performance of a distributed algorithm that combines the<br />
reconstructor with an adaptive prediction algorithm is shown in simulation to be able to<br />
better that of the traditional integrator <strong>control</strong>ler. Moreover, other simulation results show<br />
that under the frozen flow assumption this performance does not deteriorate as the number<br />
of DOF increases.<br />
The distributed algorithm is also investigated in a more general, non-adaptive context to<br />
better underst<strong>and</strong> the limitations caused by the enforced structure <strong>and</strong> study its feasibility<br />
for AO systems for large telescopes. It is shown which correlations between wavefront<br />
measurement signals can <strong>and</strong> cannot be exploited by the distributed algorithm <strong>and</strong> how this<br />
translates into assumptions on the data generating system. An algorithm is proposed to<br />
identify unknown <strong>control</strong>ler coefficients from open-loop measurement data, but despite the<br />
fact that the algorithm is applied off-line its high computational load makes it an unsuitable<br />
c<strong>and</strong>idate for high-DOF systems. But nevertheless it shows the feasibility of a <strong>modular</strong>ly<br />
distributed <strong>control</strong> architecture for large AO systems under the assumptions of Kolmogorov<br />
statistics <strong>and</strong> frozen flow <strong>and</strong> while neglecting the DM system <strong>dynamics</strong>.<br />
The requirements on the DM system are realized by using electromagnetic reluctance<br />
actuators that connect to a continuous reflective facesheet via a thin rod <strong>and</strong> are driven by<br />
a 16-bit Pulse Width Modulation (PWM)-based voltage source implemented in Field Programmable<br />
Gate Arrays (FPGAs). The facesheet stretches over all actuators to minimize<br />
the introduced optical aberrations, but all other components are <strong>modular</strong>. The actuators are<br />
manufactured in modules of 61 in a hexagonal grid <strong>and</strong> consist of layers to reduce cost <strong>and</strong><br />
allow accurate assembly. A driver Printed Circuit Board (PCB) is developed that contains<br />
driver electronics for one actuator module <strong>and</strong> consists of a PWM generator implemented<br />
in FPGAs <strong>and</strong> a low-pass filter for each actuator. The PCB is placed behind the actuator<br />
module to preserve the <strong>modular</strong> concept <strong>and</strong> the FPGAs also implement an Low Voltage<br />
Differential Signalling (LVDS) based communication protocol that allows to <strong>control</strong> a DM<br />
system containing up to 32 modules of 61 actuators. An ethernet-to-LVDS communications<br />
bridge is developed such that the system can be connected to a st<strong>and</strong>ard computer or laptop.<br />
A mathematical model is derived for the DM system that includes the behavior of the reflective<br />
facesheet, the electromagnetic reluctance actuator <strong>and</strong> the driver <strong>and</strong> communication<br />
electronics. The model is verified on several levels: for single actuators driven by a current<br />
source, for actuator modules driven by the PWM voltage sources <strong>and</strong> on the level of the<br />
complete system. On each level both static <strong>and</strong> dynamic (i.e. white noise excitation) measurements<br />
are performed that are compared to the model. Differences found between the<br />
single actuator measurements <strong>and</strong> the model are analyzed based on a sensitivity analysis,<br />
leading to minor improvements of the grid module design <strong>and</strong> providing a basis for future<br />
improvements. Seven manufactured actuator modules are found to behave in accordance<br />
with the derived model <strong>and</strong> satisfy the requirements for displacement, force <strong>and</strong> power con-
Summary xiii<br />
sumption. This proves that the manufacturing <strong>and</strong> assembly process is robust <strong>and</strong> allows the<br />
production of reliable actuator modules.<br />
Finally, the developed actuator modules <strong>and</strong> electronics are integrated with the reflective<br />
facesheet to form a complete DM system. A model for this facesheet is derived based on an<br />
analytical solution to the biharmonic plate equation for the surface shape of the facesheet<br />
due to a point force. The model is used to derive the actuator influence functions <strong>and</strong> extended<br />
with lumped masses to include the dynamic behavior. From the model, the transfer<br />
functions, impulse response functions <strong>and</strong> mode shapes are derived. The verification of<br />
the static behavior of the DM system is done using an interferometer. Verification of the<br />
modeled <strong>dynamics</strong> is performed by applying white noise excitation to the actuators <strong>and</strong><br />
measuring both the displacement as well as the velocity response of the <strong>mirror</strong> facesheet<br />
by means of a laser vibrometer. System identification is performed on the measurement<br />
data using the Multivariable Output-Error State-sPace (MOESP) algorithm to arrive at a<br />
state-space system description from which transfer functions, impulse response functions<br />
<strong>and</strong> mode shapes are derived that are compared to those of the model. Finally, the system is<br />
evaluated on <strong>control</strong> aspects, showing that the low damping of the resonances may lead to a<br />
reduction of the achievable b<strong>and</strong>width or to dependence of this b<strong>and</strong>width on the sampling<br />
frequency.
ËÑÒÚØØÒ<br />
ÔØÓÖÑÖÖ×ÔÐ<br />
ÝÒÑÒÑÓÙÐÖÒ×ØÙÖÒ<br />
De brekingsindex van lucht varieert onder <strong>and</strong>ere met temperatuur, luchtvochtigheid,<br />
druk en CO2 concentratie. Door atmosferische turbulentie varieert deze brekingsindex<br />
bovendien in zowel tijd als plaats, hetgeen leidt tot verstoringen in afbeeldingen van licht<br />
dat er doorheen schijnt. Voor optische telescopen beperken de aberraties die deze verstoringen<br />
veroorzaken de haalbare resolutie, zodat de kwaliteit van de afbeeldingen niet langer is<br />
begrensd door diffractie in de telescoop. Een <strong>Adaptive</strong> Optics (AO) systeem is een middel<br />
om deze diffractie begrensde kwaliteit te herwinnen. Dit kan onder <strong>and</strong>ere worden bereikt<br />
door het invallende licht op een deformeerbare spiegel te laten reflecteren die zijn vorm<br />
aanpast aan het golffront van dit licht zodanig dat het golffront na reflectie minimaal is in<br />
een bepaalde norm.<br />
Een dergelijk op reflectie gebaseerd AO systeem bestaat uit drie hoofdcomponenten: een<br />
golffront sensor, een deformeerbare spiegel en een regelsysteem. In dit proefschrift worden<br />
nieuwe ontwerpen beschouwd voor de laatste twee componenten met 8m klasse telescopen<br />
in zichtbaar licht als richt-applicatie. De golffront sensor wordt niet beschouwd; hiervoor<br />
wordt een sensor van het Shack-Hartmann type aangenomen. Omdat astronomen blijven<br />
hongeren voor afbeeldingen van steeds zwakkere hemellichamen worden er telescopen ontworpen<br />
van steeds grotere afmetingen om zowel de afbeeldingsresolutie als de lichtsterkte te<br />
verhogen. Daarom vormt uitbreidbaarheid naar een groter aantal vrijheidsgraden een belangrijk<br />
uitgangspunt voor de gepresenteerde ontwerpen. Daarnaast is een laag energiegebruik<br />
belangrijk opdat er geen actief koelsysteem nodig is.<br />
De belangrijkste eisen voor de deformeerbare spiegel en het regelsysteem worden afgeleid<br />
voor condities van de atmosfeer zoals die typisch op telescoop locaties voorkomen. De<br />
spatiële en temporele eigenschappen van de atmosfeer worden gemodelleerd door de<br />
spatiële en temporele spectra van het Kolmogorov turbulentie model in combinatie met<br />
de aanname dat de wind de atmosferische aberratie voortbeweegt, maar dat deze in de tijd<br />
stationair is (zgn. frozen flow). De belangrijkste bronnen voor de residuele golffront onvlakheid<br />
die van belang zijn voor het ontwerp van de deformeerbare spiegel zijn de fit-fout<br />
en de temporele fout die respectievelijk worden veroorzaakt door het beperkte aantal actuatoren<br />
en de beperkte regelb<strong>and</strong>breedte. Er wordt een keuze voor het aantal actuatoren<br />
en de benodigde regelb<strong>and</strong>breedte gemaakt die tot een gewenste optische kwaliteit leidt na<br />
reflectie uitgedrukt in een Strehl-ratio van 0.85 en waarvoor de bijbehorende specificaties<br />
voor beide vakgebieden haalbaar zijn.<br />
De eisen voor het ontwerp worden gebaseerd op licht met een golflengteλ van 550nm. Voor<br />
de correctie van de genoemde verstoringen leidt dit tot een inter-actuator afst<strong>and</strong> van 3 tot<br />
6mm, een totale actuator slag van ±5.6µm, een inter-actuator slag van ±0.36µm, een resolutie<br />
van 5nm en een maximum dissipatie in warmte van 1mW per actuator. De vereiste<br />
xiv
Samenvatting xv<br />
regelb<strong>and</strong>breedte is 200Hz, waaruit de additionele eisen volgen dat zowel de bemonsteringsfrequentie<br />
als de eerste eigenfrequentie van de deformeerbare spiegel beide boven 1kHz<br />
moeten liggen. Echter, bij het ontwerp van regelsystemen voor AO systemen met steeds<br />
meer graden van vrijheid zal de benodigde rekenkracht van de processoren van het regelsysteem<br />
snel toenemen. Zonder efficiënte algoritmen neemt de benodigde rekenkracht toe<br />
met het kwadraat van het aantal vrijheidsgraden en dus met de 4 e macht in de diameter<br />
van de telescoop. Er wordt daarom een modulair gedistribueerde architectuur van het regelsysteem<br />
voorgesteld om het zonder problemen wat betreft rekenkracht te kunnen toepassen<br />
voor grote aantallen vrijheidsgraden. Deze gedistribueerde architectuur bemoeilijkt helaas<br />
de golffront reconstructie stap die onderdeel is van het regel algoritme om te compenseren<br />
voor de spatiële dynamica van de gebruikte Shack-Hartmann golffront sensor. Er wordt<br />
echter in simulatie getoond dat een gedistribueerd algoritme waarin een adaptieve predictor<br />
wordt gecombineerd met een iteratieve reconstructor in staat is om de prestaties van<br />
de traditionele integrerende regelaar te overtreffen. Bovendien laten simulaties zien dat<br />
deze prestatie niet afneemt indien het aantal vrijheidsgraden toeneemt bij gelijkblijvende<br />
rekenkracht per vrijheidsgraad.<br />
Het gedistribueerde algoritme wordt ook in een meer generieke, niet adaptieve context onderzocht<br />
om tot een beter begrip van de beperkingen van de opgelegde structuur te komen.<br />
Er wordt getoond welke correlaties die mogelijk bestaan tussen golffront-metingen door<br />
een gedistribueerd algoritme wel en niet kunnen worden benut om de prestatie te verbeteren<br />
en welke aannames de structuur impliceert voor het data genererende systeem. De onbekende<br />
coëfficiënten van de regelaar worden uit open lus meetdata geschat met behulp van<br />
een specifiek ontwikkeld algoritme. Dit algoritme werkt niet in real-time werkt, maar desondanks<br />
maakt de benodigde rekenkracht het tot een ongeschikte k<strong>and</strong>idaat voor AO systemen<br />
met een groot aantal vrijheidsgraden. Desalniettemin laat het zien dat een modulair<br />
gedistribueerd regelsysteem tot goede prestaties in staat is indien de atmosfeer zich zoals<br />
aangenomen gedraagt.<br />
De eisen gesteld aan het deformeerbare spiegel systeem worden gerealiseerd met een ontwerp<br />
op basis van elektromagnetische reluctantie actuatoren die via sprieten zijn verbonden<br />
met een reflecterend membraan en worden bekrachtigd door spanningsbronnen op basis van<br />
het Pulse Width Modulation (PWM) principe. Het reflecterende membraan strekt zich over<br />
alle actuatoren uit om zo de optische aberraties die door de spiegel worden geïntroduceerd<br />
te minimaliseren. De actuatoren worden gemaakt in modules met 61 actuatoren geplaatst<br />
in een hexagonaal raster die uit goedkoop te fabriceren lagen bestaan om zo de kosten laag<br />
te houden en accuraat te kunnen assembleren. Er is een Printed Circuit Board (PCB) ontwikkeld<br />
die de aanstuur-elektronica bevat voor een gehele actuator module en voor iedere<br />
actuator een in Field Programmable Gate Arrays (FPGAs) geïmplementeerde PWM generator<br />
en een analoog laag-doorlaat filter bevat. Het PCB wordt achter de actuator module<br />
geplaatst in overeenstemming met het modulaire concept. De FPGAs implementeren naast<br />
de PWM generatoren ook een Low Voltage Differential Signalling (LVDS) gebaseerd communicatie<br />
protocol dat het mogelijk maakt om een deformeerbare spiegel system besta<strong>and</strong>e<br />
uit 32 modules aan te sturen via een enkele kabel. Om deze aansturing vanuit een st<strong>and</strong>aard<br />
computer of laptop te kunnen doen is bovendien een systeem ontwikkeld dat als brug<br />
fungeert tussen ethernet en LVDS.<br />
Er is een wiskundig model afgeleid dat zowel het statisch als dynamisch gedrag van het<br />
reflecterende membraan, de reluctantie actuator en de aansturings- en communicatie elek-
xvi Samenvatting<br />
tronica van het deformeerbare spiegel systeem beschrijft. Dit model is geverifieerd op verschillende<br />
niveaus: voor losse actuatoren bekrachtigd met een stroombron, voor actuator<br />
modules bekrachtigd met de ontwikkelde PWM spanningsbronnen en op het niveau van het<br />
volledige systeem. Op elk niveau zijn zowel statische als dynamische metingen gedaan die<br />
zijn vergeleken met het model. Gevonden verschillen tussen het gedrag van de losse actuator<br />
en het model worden verklaard door middel van een gevoeligheidsanalyse, hetgeen<br />
leidt tot kleine verbeteringen aan de actuator module en zicht geeft op mogelijke toekomstige<br />
verbeteringen van het ontwerp. Een zevental gerealiseerde actuator modules blijken<br />
zich conform het afgeleide model te gedragen en voldoen aan de eisen zoals die gesteld zijn<br />
wat betreft slag, benodigde kracht en vermogensdissipatie. Dit toont aan dat het maak- en<br />
assemblageproces robust is en de productie van betrouwbare modules mogelijk en haalbaar<br />
is. Zeven ontwikkelde actuator modules worden geïntegreerd met een reflecterend membraan<br />
tot een deformeerbare spiegel met 427 actuatoren. Bij het aanbrengen van een ring<br />
die beschadiging moet voorkomen breekt het membraan helaas, zodat er slechts een spiegel<br />
prototype op basis van een enkele actuator module resteert voor verder onderzoek.<br />
Het reflecterende membraan wordt gemodelleerd op basis van de biharmonische plaatvergelijking,<br />
waarvoor een analytische oplossing bestaat bij belasting met een punt-kracht.<br />
Uit dit model kunnen de invloeds-functies van de actuatoren worden bepaald en door het<br />
model uit te breiden met puntmassa’s die de massa van het membraan representeren biedt<br />
het bovendien inzicht in het dynamische gedrag van het systeem in termen van overdrachtsen<br />
impuls responsie functies en modale vormen. De verificatie van het statische gedrag van<br />
het systeem wordt gedaan op basis van interferometer metingen. Het dynamische gedrag<br />
wordt geverifieerd door de actuatoren aan te sturen met witte ruis signalen en de snelheid<br />
en verplaatsing van het reflecterende membraan te meten met een laser vibrometer. Uit<br />
de excitatie- en meetsignalen wordt met behulp van het Multivariable Output-Error StatesPace<br />
(MOESP) algoritme een state-space model geschat waarvan de overdrachts- en impuls<br />
responsie functies en modale vormen kwalitatief worden vergeleken met die van het model.<br />
Tenslotte worden de regeltechnische aspecten van het systeem geëvalueerd, hetgeen toont<br />
dat de lage demping van de resonanties mogelijk de haalbare regelb<strong>and</strong>breedte beperkt,<br />
danwel afhankelijk maakt van de gebruikte bemonsteringsfrequentie.
Symbols<br />
Roman uppercase<br />
ÆÓÑÒÐØÙÖ<br />
Symbol Description Unit<br />
0 vector or matrix whose elements are all zero<br />
1 vector or matrix whose elements are all unity<br />
Aga cross section of the axial air gap [m 2 ]<br />
Agr cross section of the radial air gap [m 2 ]<br />
Am cross section of the actuator membrane suspension [m 2 ]<br />
Aw cross section of the coil winding [m 2 ]<br />
B magnetic field density [T]<br />
Bs magnetic saturation [T]<br />
Bρ influence matrix that links the Pulse Width Modulation (PWM)<br />
voltages to the facesheet deflection at the actuator locations<br />
Bf,w influence matrix that links the PWM voltages to the facesheet<br />
deflection on the measurement grid of the Wyko interferometer<br />
˜Bf,w measured, zero piston, influence matrix that links the PWM<br />
voltages to the facesheet deflection on the measurement grid<br />
of the Wyko interferometer<br />
Bf<br />
influence matrix that links the PWM voltages to the facesheet<br />
deflections at an arbitrary grid of points on the facesheet<br />
C <strong>control</strong> system<br />
C(s) continuous time <strong>control</strong>ler<br />
C(z) discrete time <strong>control</strong>ler<br />
C1 linear stiffness coefficient [-]<br />
C2 nonlinear stiffness coefficient [-]<br />
Ca<br />
Caf<br />
diagonal matrix whose i th diagonal element is the stiffness ca<br />
of actuator i<br />
stiffness matrix comprehending both the facesheet <strong>and</strong> actuator<br />
stiffnesses<br />
[m/V]<br />
[m/V]<br />
[m/V]<br />
[m/V]<br />
[N/m]<br />
[N/m]<br />
CFPGA capacitance of the Field Programmable Gate Array (FPGA) [F]<br />
Cl capacitance used in the analog low pass filter [F]<br />
C 2 N Atmospheric turbulence strength [m −2<br />
3 ]<br />
D diameter [m]<br />
Df flexural rigidity [Nm]<br />
Dn index of refraction structure function [-]<br />
Ds diameter of the connection struts [m]<br />
Dt diameter of the telescopes primary <strong>mirror</strong> [m]<br />
DDM diameter of the Deformable Mirror (DM) [m]<br />
Dφ phase structure function [-]<br />
xvii
xviii Nomenclature<br />
Symbol Description Unit<br />
E Young’s modulus or elastic modulus [N/m 2 ]<br />
Ef Young’s modulus of the <strong>mirror</strong> facesheet [N/m 2 ]<br />
Em Young’s modulus of the actuator membrane suspension [N/m 2 ]<br />
Es Young’s modulus of the connection strut [N/m 2 ]<br />
Fa actuator force [N]<br />
Fa vector of actuator forces [N]<br />
Fρ net force acting on the facesheet at the actuator location [N]<br />
Fm magnetic force [N]<br />
Fres actuator force resolution [N]<br />
Fs spring force [N]<br />
Fρ<br />
vector of net forces acting on the facesheet at the actuator locations<br />
Fi magnetomotive force in the flux path with index i [A]<br />
G WaveFront Sensor (WFS) system<br />
G WFS geometry matrix [-]<br />
GF Fried geometry matrix characteristic block [-]<br />
GH Hudgin geometry matrix characteristic block [-]<br />
H DM system<br />
H(s) transfer function from voltage to position [m/V]<br />
HI(s) transfer function from current to position [m/A]<br />
Hm(s) transfer function from force to position [m/N]<br />
H ∗ p,Ts(z,θ) discretized transfer function from voltage to position [m/V]<br />
H ∗ v,Ts(z,θ) discretized transfer function from voltage to speed [m/sV]<br />
ˆHp,Ts(z,θ) estimated transfer function from voltage to position [m/V]<br />
ˆHv,Ts(z,θ) estimated transfer function from voltage to speed [m/sV]<br />
ˆH ∗ p,Ts(z,θ) estimated <strong>and</strong> discretized transfer function from voltage to position<br />
[m/V]<br />
ˆH ∗ v,Ts(z,θ) estimated <strong>and</strong> discretized transfer function from voltage to<br />
speed<br />
[m/sV]<br />
Hτc Transfer function for the communication latency [-]<br />
HZOH(s) Transfer function of the zero order hold operation [-]<br />
Hbc magnetic field intensity in the coil core [A/m]<br />
Hcm coercivity of the Permanent Magnet (PM) [A/m]<br />
Hga magnetic field intensity in the axial air gap [A/m]<br />
Hgr magnetic field intensity in the radial air gap [A/m]<br />
Hm magnetic field intensity in the PM [A/m]<br />
Hr magnetic field intensity in the core <strong>and</strong> baseplate [A/m]<br />
I current [A]<br />
I identity matrix<br />
Ia current through the actuator coil [A]<br />
ICl current through the capacitance Cl [A]<br />
If current through the fictitious winding [A]<br />
IRl current through the resistance Rl [A]<br />
J1(·) Bessel function of the first kind [-]<br />
Ja current density in the actuator coil [A/m 2 ]<br />
Ka motor constant [N/A]<br />
Ka diagonal matrix, whose i th diagonal element is the motor constant<br />
ka of actuator i<br />
[N/A]<br />
Km facesheet stiffness matrix [N/m]<br />
[N]
Nomenclature xix<br />
Symbol Description Unit<br />
L length [m]<br />
L0 atmospheric outer scale [m]<br />
L11, L22 self inductance [H]<br />
L12, L21 mutual inductance [H]<br />
La actuator inductance [H]<br />
Ll inductance of the low pass filter [H]<br />
Ls length of the connection strut [m]<br />
Lw length of coil wire [m]<br />
M moment [Nm]<br />
Maf diagonal matrix whose i th diagonal element is the sum of the<br />
moving actuator mass <strong>and</strong> the lumped facesheet mass at its location<br />
[kg]<br />
N number of windings [-]<br />
Na number of actuators [-]<br />
Nav number of actuators [-]<br />
Nb number of counter bits [-]<br />
Nd demagnetization factor [-]<br />
Nf number of fictitious windings [-]<br />
Nm number of actuator modules [-]<br />
Nn number of distributed <strong>control</strong>ler nodes [-]<br />
Ns number of WFS lenselets [-]<br />
Nw number of pixels used in the Wyko interferometer [-]<br />
P plant to be <strong>control</strong>led<br />
P pressure [N/m 2 ]<br />
temporal power spectrum [J/Hz]<br />
Power dissipation [W]<br />
P projection matrix to remove the ’piston’ term [-]<br />
P0 light intensity [lm/m 2 ]<br />
Pa power dissipation in the actuator [W]<br />
Pdyn(fclk) dynamic power dissipation in an FPGA as function of the clock<br />
frequency<br />
[W]<br />
Pe electrical power dissipation [W]<br />
Pload power dissipation due to the load [W]<br />
Psc short circuit power dissipation [W]<br />
Ptot total power dissipation [W]<br />
Qn Hadamard matrix of size nxn [-]<br />
R open-loop <strong>control</strong> system<br />
R open-loop, minimum variance wavefront reconstruction matrix [-]<br />
Ra electrical resistance of the actuator coil [Ω]<br />
Rc electrical resistance applied for the fine PWM signal [Ω]<br />
Rl electrical resistance of the analog low pass filter [Ω]<br />
ℜbc magnetic reluctance of the part of the baseplate that forms the<br />
core of the actuator coil<br />
[1/H]<br />
ℜc magnetic reluctance of the actuator moving core [1/H]<br />
ℜflc magnetic reluctance of leakage flux path of the coil [1/H]<br />
ℜflm magnetic reluctance of leakage flux path of the PM [1/H]<br />
ℜga magnetic reluctance of the actuator axial air gap [1/H]<br />
ℜgr magnetic reluctance of the actuator radial air gap [1/H]<br />
ℜm magnetic reluctance of the PM [1/H]
xx Nomenclature<br />
Symbol Description Unit<br />
S Strehl ratio [-]<br />
S(s) transfer function from disturbance to residual error [-]<br />
T temperature [K]<br />
Te <strong>control</strong> loop delay [s]<br />
Ts sampling time [s]<br />
TPWM time periode for the PWM base frequency [s]<br />
Ur three-column matrix containing piston, tip <strong>and</strong> tilt modes evaluated<br />
on an arbitrary grid<br />
[-]<br />
Uρ three-column matrix containing piston, tip <strong>and</strong> tilt modes evaluated<br />
on the actuator grid<br />
[-]<br />
V voltage [V]<br />
V matrix of actuator comm<strong>and</strong> voltage vectors for identification [V]<br />
Va voltage over the actuator coil [V]<br />
Vcc supply voltage [V]<br />
VC l voltage over the capacitor Cl [V]<br />
vi actuator comm<strong>and</strong> voltage vector with indexifor identification [V]<br />
Vm volume of the PM [m 3 ]<br />
Vw coil volume [m 3 ]<br />
VRa voltage over the resistance Ra [V]<br />
W magnetic coenergy [J]<br />
Roman lowercase<br />
Symbol Description Unit<br />
ba mechanical damping in the actuator [Ns/m]<br />
c speed of light in vacuum [m/s]<br />
cD compression factor (=Dt/DDM) [-]<br />
ca actuator stiffness [N/m]<br />
d inter actuator spacing [m]<br />
dt inter actuator spacing projected on the telescope aperture [m]<br />
f frequency of light [Hz]<br />
fc <strong>control</strong> b<strong>and</strong>width [Hz]<br />
fclk FPGA clock frequency [Hz]<br />
fe undamped mechanical actuator resonance frequency [Hz]<br />
fFPGA FPGA clock frequency [Hz]<br />
fG Greenwood frequency [Hz]<br />
fPWM PWM base frequency [Hz]<br />
fN Nyquist frequency [Hz]<br />
fs sampling frequency [Hz]<br />
g gravitation acceleration [m/s 2 ]<br />
h distance between the core in the undeflected membrane suspension<br />
core <strong>and</strong> the PM<br />
[m]<br />
h height [m]<br />
hn heat transfer coefficient [W/m 2 ]<br />
l0 atmospheric inner scale [m]<br />
lb magnetic flux path length through the baseplate [m]<br />
lc magnetic flux path length through the moving core [m]<br />
lga magnetic flux path length through the axial air gap [m]
Nomenclature xxi<br />
Symbol Description Unit<br />
lgr magnetic flux path length through the radial air gap [m]<br />
lm magnetic flux path length through the PM [m]<br />
m mass [kg]<br />
mac mass of the moving core in the actuator [kg]<br />
maf mass of the <strong>mirror</strong> facesheet per actuator [kg]<br />
mf mass of the <strong>mirror</strong> facesheet [kg]<br />
ms mass of the actuator strut [kg]<br />
n white noise vector<br />
n index of refraction [-]<br />
nair index of refraction of air [-]<br />
r spatial coordinate [m]<br />
r0 Fried parameter [m]<br />
rc communication radius [-]<br />
ri normalized spatial coordinate with indexiin the complex plane [-]<br />
r vector of normalized coordinates in the complex plane [-]<br />
rf <strong>mirror</strong> facesheet radius [m]<br />
rm actuator membrane suspension radius [m]<br />
s Laplace variable (s = jω) [rad/s]<br />
the number of block-rows used in the Multivariable Output-<br />
Error State-sPace (MOESP) algorithm<br />
[-]<br />
s(t) Open-loop wavefront disturbance measurement vector [rad]<br />
t time [s]<br />
t thickness [m]<br />
tm actuator membrane suspension thickness [m]<br />
tf <strong>mirror</strong> facesheet thickness [m]<br />
v(h) wind speed at altitudeh [m/s]<br />
v speed of light [m/s]<br />
vw wind speed [m/s]<br />
wx rigid body rotation around the x-axis (tip) [m]<br />
wp rigid body displacement in z-direction (piston) [m]<br />
wy rigid body rotation around the y-axis (tilt) [m]<br />
y(t) Closed-loop wavefront disturbance measurement vector [rad]<br />
za actuator displacement [m]<br />
˙za actuator velocity [m/s]<br />
¨za actuator acceleration [m/s 2 ]<br />
zf facesheet deflection [m]<br />
zf,0 unactuated facesheet deflection [m]<br />
ˆzf measured facesheet deflection [m]<br />
z z-transform variable (z = e jω ) [-]<br />
zia inter actuator stroke [m]<br />
z0 initial axial air gap height [m]<br />
zs suspension membrane deflection [m]<br />
Greek symbols<br />
Symbol Description Unit<br />
α linear coefficient of expansion [m/m/K]<br />
switching activity in an FPGA [-]
xxii Nomenclature<br />
Symbol Description Unit<br />
integrator gain [-]<br />
β integrator leak factor [-]<br />
γw comm<strong>and</strong> vector scaling constant [m]<br />
Γv diagonal scaling matrix on comm<strong>and</strong> voltages [-]<br />
Γz diagonal scaling matrix on measured displacements [-]<br />
δij Kronecker delta [-]<br />
ǫ(t) vector of wavefront phase at time t [-]<br />
ζ telescope angle w.r.t. Zenith [ ◦ ]<br />
η actuator coupling [-]<br />
θ rotation around the z-axis [rad]<br />
θ angular coordinate [-]<br />
θa angular distance between object <strong>and</strong> reference star [rad]<br />
Θ angle of the chief ray w.r.t. the optical axis [rad]<br />
κ spatial frequency [1/m]<br />
κx spatial frequency in x direction [1/m]<br />
κy spatial frequency in y direction [1/m]<br />
κz spatial frequency in z direction [1/m]<br />
κf fitting error coefficient [-]<br />
λ wavelength of light [m]<br />
thermal conductivity [W/mK]<br />
flux linkage [Wb]<br />
Λ diagonal scaling matrix on comm<strong>and</strong> voltages [-]<br />
µ0 magnetic permeability of vacuum [N/A 2 ]<br />
µr relative magnetic permeability [-]<br />
µrm relative magnetic permeability of the PM [-]<br />
µr b relative magnetic permeability of the baseplate [-]<br />
ν Poisson ratio [-]<br />
νm Poisson ratio of the actuator membrane suspension material [-]<br />
νf Poisson ratio of the <strong>mirror</strong> facesheet material [-]<br />
ρ material density [kg/m 3 ]<br />
ρm density of the actuator membrane suspension material [kg/m 3 ]<br />
ρa density of air [kg/m 3 ]<br />
ρf density of the <strong>mirror</strong> facesheet material [kg/m 3 ]<br />
ρi complex coordinate with index i [m]+j[m]<br />
ρs density of the connection strut material [kg/m 3 ]<br />
σ 2 angle wavefront variance due to off axis observation angle [nm 2 ]<br />
σ 2 cal wavefront variance due to calibration errors [nm 2 ]<br />
σ 2 ctrl wavefront variance due to a <strong>control</strong> error [nm 2 ]<br />
σ 2 delay wavefront variance due to a delay [nm 2 ]<br />
σ 2 fit wavefront variance due to limited number of actuators [nm 2 ]<br />
σ 2 meas wavefront variance due to measurement errors [nm 2 ]<br />
σ 2 n wavefront variance due to measurement noise [nm 2 ]<br />
σ 2 temp wavefront variance due to limited b<strong>and</strong>width [nm 2 ]<br />
σ 2 total total wavefront variance [nm 2 ]<br />
σ(v,i) expected Root Mean Square (RMS) actuator voltage based on a<br />
Von Karmann spatial power spectrum<br />
σ 2 wf wavefront variance [nm 2 ]<br />
τ delay [s]<br />
τe Sensor integration time [s]<br />
[V]
Nomenclature xxiii<br />
Symbol Description Unit<br />
τc communication latency [s]<br />
τUDP delay caused by the User Datagram Protocol (UDP) packet<br />
transfer<br />
[s]<br />
τLV DS delay caused by the Low Voltage Differential Signalling<br />
(LVDS) packet transfer<br />
[s]<br />
φ optical phase [rad]<br />
φ(t) vector containing wavefront phase disturbance values [rad]<br />
φi magnetic flux in a circuit with index i [Wb]<br />
Φ spatial PSD [rad 2 ]<br />
magnetic flux [Vs]<br />
Ωρρ facesheet compliance matrix w.r.t. the actuator grid [m/N]<br />
Ωrρ facesheet compliance matrix mapping forces at the actuator locations<br />
ρ to displacements at an arbitrary gridr<br />
[m/N]<br />
Operators <strong>and</strong> sets<br />
.<br />
Symbol Description<br />
∇ Laplacian operator (partial derivative)<br />
⊗ Kronecker product<br />
◦ Hadamard (element-wise) product<br />
Tr(·) Trace of the dotted matrix<br />
\ set difference operator<br />
〈·〉 expected value of the dotted expression<br />
Nn(m,C) set of ergodic white noise signals s(t) ∈ R n with mean m <strong>and</strong> covariance<br />
(matrix) C<br />
n<br />
S(i) Union of the sets S(i), equivalent toS(1) ∪...∪S(n)<br />
i=1<br />
·F<br />
δij<br />
Frobenius norm of<br />
<br />
the dotted expression<br />
1 for i = j,<br />
Kronecker delta:<br />
0 for i = j.
Ê Ç Ê ÊÅ ËÁ<br />
Analog to Digital Convertor<br />
Aliased Frequency Response<br />
ËÅ<br />
Function<br />
Æ<br />
<strong>Adaptive</strong> Optics<br />
<br />
Auto-Regressive<br />
ÀÌ<br />
Auto-Regressive Moving<br />
Average<br />
ÈÍ<br />
Application-Specific Integrated<br />
ÅÇË<br />
Circuit<br />
<strong>Adaptive</strong> Secondary Mirror<br />
Æ<br />
Controller Area Network<br />
Ë<br />
Charge Coupled Device<br />
<br />
Canada France Hawaii<br />
ÁË<br />
Telescope<br />
Central Processing Unit<br />
Å<br />
Complementary Metal Oxide<br />
Ç<br />
Semiconductor<br />
<br />
Computer Numerical Control<br />
Å<br />
Curvature Sensor<br />
ÄÌ<br />
Digital to Analog Convertor<br />
Dutch Institute for Systems <strong>and</strong><br />
ÄÌ<br />
Control<br />
Ê<br />
Deformable Mirror<br />
Degrees Of Freedom<br />
ËÇ<br />
Direct Current<br />
Electrical Discharge Machining<br />
Å<br />
European Extremely Large<br />
Ì<br />
Telescope<br />
ÁÊ<br />
Extremely Large Telescope<br />
ÇÎ<br />
Eigensystem Realization<br />
Algorithm<br />
European Southern<br />
Observatory<br />
Finite Element Model<br />
Field Effect Transistor<br />
Finite Impulse Response<br />
Field Of View<br />
xxiv<br />
ÖÓÒÝÑ×<br />
È Ê ÏÀÅ ÄÇ ÀËÌ Á ÁÈ<br />
Field Programmable Gate<br />
ÁÊ<br />
Array<br />
ÂÏËÌ<br />
Frequency Response Function<br />
ÄÌ<br />
Full Width Half Maximum<br />
ÄÅÁ<br />
Ground Layer <strong>Adaptive</strong> Optics<br />
ÄÅË<br />
Hubble Space Telescope<br />
<br />
Integrated Circuit<br />
ÄÈÎ<br />
Internet Protocol<br />
ÄÌÁ<br />
Infra Red<br />
ÄÎË<br />
James Webb Space Telescope<br />
Large Binocular Telescope<br />
ÄÉ<br />
Linear Matrix Inequality<br />
ÄË<br />
Least Mean Squares<br />
Å<br />
Computer Aided Design<br />
Å<br />
Linear Parameter Varying<br />
ÅÇ<br />
Linear Time-Invariant<br />
Low Voltage Differential<br />
ÅÇÇ<br />
Signalling<br />
Linear Quadratic Gaussian<br />
Least Significant Bit<br />
ÅÅË<br />
Moving Average<br />
Media Access Control<br />
ÅÁÅÇ<br />
Multi-Conjugate <strong>Adaptive</strong><br />
ÅÅÌ<br />
Optics<br />
ÅÌ<br />
Multi-Object <strong>Adaptive</strong> Optics<br />
ÅÇËÈMultivariable<br />
ÅË<br />
Output-Error<br />
ÆÈ<br />
State-sPace<br />
Micro-Electro-Mechanical<br />
ÇÈ<br />
Systems<br />
ÇÏÄ<br />
Multi-Input Multi-Output<br />
Multiple Mirror Telescope<br />
Mean Time Before Failure<br />
Most Significant Bit<br />
Non Common Path Aberration<br />
Optical Path Difference<br />
Overwhelmingly Large
ÈËÁ È È Ë Acronyms<br />
ÈÀ ËÀË<br />
xxv<br />
ÈÅÆ ËÁËÇ<br />
ÈÅ ËÆÊ<br />
Telescope<br />
ÈË ËÎ<br />
Predictor Based Subspace<br />
ÈË ÌÈ<br />
IDentification<br />
ÈË Ì<br />
Ranging<br />
Printed Circuit Board<br />
ÈÌÎ ÌÅÌ<br />
Steepest Descent<br />
Partial Differential Equation<br />
ÈÏÅ ÍÈ<br />
Shack Hartmann sensor<br />
PHYsical layer (ethernet)<br />
ÉÈ ÍÄ<br />
Single-Input Single-Output<br />
Lead Manganese Niobate<br />
ÊÅ ÍË<br />
Signal to Noise Ratio<br />
Permanent Magnet<br />
ÊÄË Î<br />
Singular Value Decomposition<br />
Power Spectral Density<br />
ÊÅË ÎÄÌ<br />
Transmission Control Protocol<br />
Point Spread Function<br />
ÊÌ ÏË<br />
Transfer Function<br />
ÏÀÌ<br />
Thirty Meter Telescope<br />
ÏËË<br />
User Datagram Protocol<br />
Pyramid Sensor<br />
ÇÀ<br />
Ultra Low Expansion<br />
Peak To Valley<br />
ÍÈ<br />
Universal Serial Bus<br />
Pulse Width Modulation<br />
Variance Accounted For<br />
Quadratic Programming<br />
Very Large Telescope<br />
R<strong>and</strong>om Access Memory<br />
WaveFront Sensor<br />
Recursive Least Squares<br />
William Herschel Telescope<br />
Root Mean Square<br />
Wide Sense Stationary<br />
Real-Time Control<br />
Zero Order Hold<br />
ËÁÊSCIntillation Detection <strong>and</strong><br />
Xylinx University Program
ÔØÖÓÒ<br />
ÁÒØÖÓÙØÓÒ<br />
The field of astronomy is briefly introduced to show the relevance of <strong>Adaptive</strong><br />
Optics (AO) systems. The components of such a system are described <strong>and</strong><br />
challenges for future AO systems are discussed. Finally, an overview of this<br />
thesis is given.<br />
1
1<br />
2 1 Introduction<br />
1.1 Notes on notation<br />
Before introducing the subject of adaptive optics, note that the mathematical symbols <strong>and</strong><br />
operators that are frequently used throughout this thesis are listed in the nomenclature on<br />
page xvii. Scalar values are generally denoted by italic lower case symbols, whereas vectors<br />
are denoted by boldface symbols. Matrices are denoted by capital, uppercase symbols <strong>and</strong><br />
both sets <strong>and</strong> systems are denoted by calligraphical symbols. The (i,j) th elements of a<br />
matrixAis denoted by a subscript between square brackets: A [i,j] <strong>and</strong> thei th element of a<br />
vectorxis denotedx [i]. A subscript with round brackets (e.g. a (i)) denotes an enumeration<br />
of the subscripted symbol, whereas subscripted symbols without brackets such asav denote<br />
specific unique symbols.<br />
Further, Ia denotes an identity matrix of size a × a, where the subscript dimension will<br />
be omitted if this is clear from the context. The complex-valued z-transform variable z is<br />
used to indicate discrete-time systems <strong>and</strong> denotes the forward temporal shift operator. The<br />
corresponding continuous time Laplace operator is denoted by the symbol s. The notation<br />
〈·〉 denotes the statistical expected value of the dotted expression over time t. In case that<br />
the expected value is not taken w.r.t. time this is made clear in the context. The average<br />
value over a specific time range is denoted as:<br />
〈·〉 t1<br />
t0 = 1<br />
t1−t0+1<br />
t1<br />
·(t).<br />
Further, the notation · F denotes the Frobenius norm of the dotted expression <strong>and</strong> δij<br />
denotes the Kronecker delta s.t.<br />
<br />
1 for i = j,<br />
δij =<br />
t=t0<br />
0 for i = j.<br />
The set N n (m,C) comprehends ergodic, normally distributed white noise signals x(t) ∈<br />
R n with mean m = 〈x〉 <strong>and</strong> covariance C = xx T for which x(t)x T (t−p) = 0 for<br />
p = 0. The union of two setsS1 <strong>and</strong>S2 will be denoted asS1∪S2 <strong>and</strong> the union of multiple<br />
setsS (i) fori = 1...m will be denoted:<br />
m<br />
S (i) = S (1) ∪S (2) ∪...∪S (m).<br />
i=1<br />
The set S that consists of all elements in S1 that are not present in the set S2 is denoted<br />
S = S1 \S2.<br />
Further, letG [S1,S2] denote the sub-matrix obtained from a matrixGby extracting all rows<br />
i ∈ S1 <strong>and</strong> columnsj ∈ S2 thereof in lexicographical order. Similarly,p [S] denotes a vector<br />
obtained by stacking the elementsp [i] fori ∈ S.<br />
1.2 Astronomy<br />
The invention of the telescope is commonly attributed to Lipperhey <strong>and</strong> Janssen around<br />
1608, although this is a historically debated subject. The first astronomers used transmissive
1.2 Astronomy 3<br />
Flat, undistorted wavefront<br />
Average refractive index<br />
Lower than average refractive index<br />
e.g. higher than average temperature<br />
Higher than average refractive index<br />
e.g. lower than average temperature<br />
Distorted wavefront<br />
t0<br />
t0+t<br />
Figure 1.1: Wavefront distortions caused by variations in the index of refraction of air. Light enters<br />
at the top <strong>and</strong> passes through hotter <strong>and</strong> colder than average air bubbles. As a result, the<br />
corresponding rays are retarded <strong>and</strong> advanced respectively.<br />
lenses with satisfying results on small telescopes, but they soon realized that to obtain more<br />
detailed images a larger aperture diameter was required. The quality of glass castings was<br />
limited, hence the reflectivity of polished metals like copper <strong>and</strong> tin was used for larger<br />
telescopes. On the other h<strong>and</strong>, polishing large <strong>mirror</strong>s was specialized <strong>and</strong> tedious work<br />
<strong>and</strong> it took more than 80 years [98] before it was realized that the limited image detail of<br />
1m+ telescopes was not always caused by polishing errors. Although Christian Huygens<br />
proposed around 1656 that the atmosphere was to blame, this was recognized by the<br />
majority of astronomers only at the end of the 19th century. Ever since, the astronomers’<br />
desire for more <strong>and</strong> more detailed images of the sky has been obstructed by our turbulent<br />
atmosphere.<br />
This introduces quickly changing aberrations to the wavefront of the incoming light that<br />
limit the spatial resolution of the images that can be recorded. The wavefront is the surface<br />
formed by all light particles emitted by an object at the same time, which – in the absence<br />
of disturbances – can be regarded as an exp<strong>and</strong>ing sphere. Most celestial objects are so<br />
distant, that when their light arrives at the earth, the wavefront sphere is so large that the<br />
small fraction that will penetrate the atmosphere <strong>and</strong> fall onto an observing telescope can<br />
be considered flat.<br />
The atmospheric wavefront distortion (unflatness) is the result of variations in the index<br />
of refraction of air. This index represents the speed of light in vacuum relative to that in<br />
air <strong>and</strong> mainly depends on temperature, but also on humidity <strong>and</strong> pressure. Since these<br />
parameters vary spatially in the atmosphere, some parts of the wavefront are advanced,<br />
whereas others are retarded (figure 1.1).<br />
With an ideal circular telescope – one whose components do not introduce aberrations<br />
– <strong>and</strong> without the presence of the atmosphere a point source is shaped by diffraction on its<br />
finite components <strong>and</strong> is described by the Airy function:<br />
P0(θ) = πD2<br />
4λ 2<br />
⎡<br />
⎣ 2J1<br />
<br />
πD|θ|<br />
λ<br />
<br />
πD|θ|<br />
λ<br />
Here P0 is the light intensity as function of the angular coordinate θ, λ the wavelength,<br />
D the telescope diameter <strong>and</strong> J1(·) the Bessel function of the first kind. The Airy function<br />
⎤<br />
⎦<br />
2<br />
1
1<br />
4 1 Introduction<br />
Figure 1.2: The Airy function. Figure 1.3: The Airy function (top view).<br />
is shown graphically in figures 1.2 <strong>and</strong> 1.3. The first dark ring is at an angular distance of<br />
1.22 λ<br />
D <strong>and</strong> is called the resolution of the ideal telescope. An astronomical object can be seen<br />
as a number of point sources, each of which spreads according to the Airy function. Mathematically,<br />
the Airy function is an optical transfer function, such that the object’s image is<br />
obtained by convolution of the objects source function with the Airy function. This results<br />
in the image with the least degradation possible <strong>and</strong> the resolving power of the telescope is<br />
called diffraction limited. In practice, the image is never diffraction limited, but further degraded<br />
due to wavefront distortions introduced by the atmosphere <strong>and</strong> telescope optics. The<br />
image is now obtained through convolution with a distorted Airy function, which is referred<br />
to with the more general term Point Spread Function (PSF). The PSF is thus determined by<br />
wavefront aberrations <strong>and</strong> forms the basis for various measures of optical quality:<br />
• Strehl ratio: S(θ) = P(θ)<br />
· 100%. This is the actual central peak intensity (P(θ))<br />
P0(θ)<br />
relative to the diffraction limited central peak intensity of the PSF (P0(θ)).<br />
• The Full Width Half Maximum (FWHM) of the PSF, which is the diameter around<br />
the central peak at which the PSF reaches half the peak value.<br />
• Encircled energy. This measures the energy distribution in the PSF as the energy<br />
fraction contained within a circle of radiusr.<br />
It is only in the last fraction of a second when the light passes through the atmosphere to<br />
arrive at a telescope that its wavefront gets distorted. However, for a long time no solution<br />
to this problem could be found. To obtain the best image of the actual object the wavefront<br />
distortion should either be compensated or the image should be taken before the light hits<br />
the atmosphere. The first ideas towards wavefront compensation were published by Horace<br />
W. Babcock in 1953 [12], which is considered to be the start of the field of <strong>Adaptive</strong><br />
Optics (AO). However, no actual compensation was achieved until 20 years later.<br />
Atmospheric wavefront distortions can also be evaded by placing the telescope in space.<br />
Although this was no realistic option for the early astronomers, by now this has been successfully<br />
proven by the clear images of the Hubble Space Telescope (HST). Its successor –<br />
the James Webb Space Telescope (JWST) – will soon be launched, having a much larger primary<br />
<strong>mirror</strong>. Nevertheless, space telescopes have several important drawbacks: the space<br />
launch is expensive <strong>and</strong> requires extreme precautions for the fragile telescope; space is a<br />
harsh environment <strong>and</strong> maintenance is nearly impossible. Consequently, the primary <strong>mirror</strong><br />
area of space telescopes is much more difficult to extend than that of their earth based counterparts.<br />
Especially, since AO systems have now reached the level of maturity that allows
1.3 <strong>Adaptive</strong> optics 5<br />
Object<br />
Science<br />
image<br />
Atmospheric<br />
turbulence<br />
Telescope<br />
Deformable Mirror<br />
Controller<br />
Wavefront<br />
sensor<br />
Figure 1.4: Schematic of the<br />
adaptive optics system in a<br />
ground based telescope<br />
the construction of large telescopes with almost diffraction limited seeing. Recently, even<br />
the power of using several wavefront correctors simultaneously was demonstrated on-sky<br />
using the MAD demonstrator [126].<br />
AO also finds applications in high power laser systems, optical communication <strong>and</strong><br />
medicine with each their own specific challenges. This thesis will focus on astronomy,<br />
starting with a description of the system components that are relevant for this application<br />
field in the next section.<br />
1.3 <strong>Adaptive</strong> optics<br />
A schematic of an astronomical telescope with an AO system is shown in figure (1.4). After<br />
reflecting on the primary <strong>and</strong> several other <strong>mirror</strong>s of the telescope, the distorted light reflects<br />
on the wavefront corrector, which is here assumed to be a Deformable Mirror (DM).<br />
In practice several other optical components may be present before <strong>and</strong> after the corrector,<br />
but eventually a dichroic beam splitter splits the wavefront into two parts. One part proceeds<br />
to the WaveFront Sensor (WFS) <strong>and</strong> the other to the science camera that records the image.<br />
Based on current <strong>and</strong> past WFS measurements the Real-Time Control (RTC) calculates new<br />
actuator signals for the DM.<br />
As mentioned, the intended application field for the DM developed within the context of<br />
this thesis is astronomy. In the coming sections the system components relevant for this<br />
field will be described in more detail.<br />
1.3.1 Atmospheric turbulence<br />
As mentioned, the index of refraction of air depends a.o. on its temperature <strong>and</strong> humidity.<br />
The atmosphere is often represented as a set of different size air bubbles for which these<br />
properties are different, giving each their own refractive index n. Wind carries these air<br />
bubbles over the telescopes aperture, while only very slowly changing these properties. The<br />
1
1<br />
6 1 Introduction<br />
Figure 1.5: Schematic of<br />
a Shack-Hartmann wavefront<br />
sensor.<br />
plane<br />
wavefront<br />
aberrated<br />
wavefront<br />
focal length<br />
CCD<br />
detector<br />
latter is called the frozen flow assumption or Taylor hypothesis. At most telescope sites a<br />
large part of the turbulence is at the lower altitude, where a temperature gradient between air<br />
<strong>and</strong> ground exists [178]. This is called the ground layer <strong>and</strong> typically 75% of the distortion<br />
is caused by the lower 2km of the atmosphere.<br />
In 1941 Kolmogorov laid the foundation for currently used atmospheric turbulence models<br />
[112]. Kolmogorov concluded that in a turbulent flow the kinetic energy decreases with<br />
the −5<br />
3 power of the spatial frequency. From this, Tatarski [173] <strong>and</strong> Fried [63] developed<br />
the st<strong>and</strong>ard model for astronomical seeing. In chapter 2 a more detailed description of<br />
atmospheric turbulence <strong>and</strong> its consequences for the DM design is given.<br />
1.3.2 The wavefront sensor<br />
Wavefront sensors for AO systems come in many types. Examples are Shearing interferometers,<br />
Curvature Sensor (CS) [148], Pyramid Sensor (PS) [145] <strong>and</strong> Shack Hartmann<br />
sensor (SHS) [160]. All mentioned sensors use an indirect method to measure the actual<br />
wavefront shape (phase) <strong>and</strong> require a mathematical transformation to derive this information<br />
from the measurements.<br />
The CS uses an array of lenses to focus the wavefront into a multitude of spots. A detector<br />
measures the intensities before <strong>and</strong> after the focal plane. If there is a local curvature in the<br />
wavefront, the focus position of the spot is changed, leading to variation of the measured<br />
intensities. In the PS the wavefront falls onto many small pyramids. The facets of each<br />
pyramid split the light into a number of beams that are imaged onto a detector. If the<br />
wavefront is flat the result is an equal amount of light in all beams, but otherwise this<br />
distribution changes. Extra resolution can be gained by mechanically moving the pyramids<br />
[145].<br />
For various reasons the SHS is currently the most widely used in AO systems for optical<br />
telescopes <strong>and</strong> will be considered throughout this thesis. It consist of a two dimensional<br />
array of lenselets of which each lens images incident light onto a Charge Coupled<br />
Device (CCD) detector or quad-cell array (figure 1.5). The position offset of this image<br />
with respect to a reference is a measure for the first spatial derivative of the incoming<br />
wavefront at the location of a lenselet.<br />
Due to this design of the sensor, its <strong>dynamics</strong> consist of both a spatial <strong>and</strong> a temporal<br />
displacement
1.3 <strong>Adaptive</strong> optics 7<br />
ǫ3 ǫ6 ǫ9<br />
ǫ2<br />
ǫ1<br />
y2<br />
y1<br />
ǫ5<br />
ǫ4<br />
y4<br />
y3<br />
ǫ8<br />
ǫ7<br />
Figure 1.6: The Fried geometry. The measured<br />
gradients (arrows) are related to the values of<br />
the four surrounding phase points (circles) via<br />
the definition in (1.1). This definition enables<br />
the unobservable waffle mode indicated by the<br />
checkerboard coloring of the phase points.<br />
ǫ3 ǫ6 ǫ8<br />
ǫ2<br />
ǫ1<br />
y2<br />
y1<br />
ǫ5<br />
ǫ4<br />
y4<br />
y3<br />
ǫ7<br />
Figure 1.7: The Hudgin geometry. The measured<br />
gradients (arrows) are related to the values<br />
of three neighboring phase points (circles)<br />
via the definition in (1.2).<br />
part. The temporal part of the <strong>dynamics</strong> of the wavefront sensor arises from the working<br />
principle of the underlying CCD camera or quad-cell array. These detectors collect incoming<br />
photons, effectively integrating light energy over the exposure time. The integrating<br />
behavior corresponds to a low-pass filter characteristic. Since these photons must be used<br />
as efficiently as possible, this exposure time is usually equal to the sampling time of the<br />
AO system. This implies an average measurement delay of half a sampling time, to which<br />
read-out <strong>and</strong> post-processing times must be added, justifying the assumption of a full<br />
sample delay that will be made throughout this thesis.<br />
In practice, the spatial SHS <strong>dynamics</strong> are modeled via either modal or zonal approaches.<br />
For modal approaches a set of orthogonal, differentiable basis functions is defined such as<br />
the Zernike polynomials [136]. Both the polynomials themselves as well as their spatial<br />
derivatives are analytic functions that can be discretized over the sensor grid. For wavefront<br />
reconstruction, these discretized functions can be used to estimate the modal coefficients for<br />
a given set of gradient measurements. The actual wavefront shape can then be determined<br />
on an arbitrary grid as a weighted sum of the Zernike polynomials discretized on this grid.<br />
For zonal approaches such as the Fried <strong>and</strong> Hudgin geometries [179], the wavefront gradients<br />
are modeled as a function of the values of a few surrounding phase grid points. Let<br />
the residual wavefront phase distortion that determines the SHS measurements at time t be<br />
denoted by a vector ǫ(t) ∈ R Nǫ . For the Fried geometry it is assumed that the wavefront<br />
phase is discretized over a virtual phase grid as defined in figure 1.6. When including one<br />
full sample measurement delay <strong>and</strong> measurement noisew (i)(t), each gradient measurement<br />
1
1<br />
8 1 Introduction<br />
vectory (j)(t) ∈ R 2 is expressed in terms of the four surrounding phase values as:<br />
y (j)(t) = 1<br />
⎡<br />
ǫ<br />
<br />
1 1 −1 −1 ⎢<br />
2 1 −1 1 −1 ⎣<br />
<br />
GF<br />
j<br />
tr(t−1)<br />
ǫ j<br />
br (t−1)<br />
ǫ j<br />
tl (t−1)<br />
ǫ j<br />
bl (t−1)<br />
⎤<br />
⎥<br />
⎦+w<br />
(i)(t), (1.1)<br />
<br />
ǫS (t−1)<br />
(j)<br />
where the notationǫ j<br />
tl (t) refers to the residual wavefront disturbance at the phase grid point<br />
that lies top-left (tl) adjacent to sensor grid point j. Similarly, bl, tr <strong>and</strong> br refer to the<br />
bottom-left, top-right <strong>and</strong> bottom-right points respectively. The indices of the corresponding<br />
nodes form the four elements of the setS (j). For instance, in figure 1.6 this implies that<br />
S (2) = {6,5,3,2} <strong>and</strong>S (3) = {8,7,5,4} .<br />
For the Hudgin geometry the spatial gradients are defined as illustrated in figure 1.7. Correspondingly,<br />
the measurementsy (j)(t) are expressed as:<br />
y (j)(t) = 1<br />
⎡<br />
ǫ<br />
1 −1 0<br />
⎣<br />
2 0 1 −1<br />
<br />
GH<br />
j<br />
tr(t−1)<br />
ǫ j<br />
tl (t−1)<br />
ǫ j<br />
bl (t−1)<br />
⎤<br />
⎦+w<br />
(i)(t), (1.2)<br />
<br />
ǫS (t−1)<br />
(j)<br />
When y (j)(t), ǫ (i)(t) <strong>and</strong> w (i)(t) are stacked in the vectorsy(t) ∈ R 2Ns , ǫ(t) ∈ R Nǫ <strong>and</strong><br />
w(t) ∈ R 2Ns respectively, then for both geometries the measurements can be expressed as:<br />
y(t) = Gǫ(t−1)+w(t).<br />
For the Fried geometry, each 2×Nǫ block-row of the generally tall matrix G ∈ R 2Ns×Nǫ<br />
contains only the four non-zero columns of GF at the columns corresponding to the definition<br />
of ǫS (j) in (1.1). Similarly, for the Hudgin geometry each 2×Nǫ block-row of the<br />
matrix G contains the three non-zero columns of GH at the columns corresponding to the<br />
definition ofǫS (j) in (1.2).<br />
The geometry matrix is rank deficient for both geometries. The Hudgin geometry has one<br />
unseen wavefront shape that yields a zero measurement, which is called the piston mode.<br />
For this mode, all phase values are equal, i.e. ǫ (1)(t) = ... = ǫ (Nǫ)(t).<br />
The Fried geometry has two unseen modes: besides the piston mode it also has the waf-<br />
fle mode. The wavefront shape corresponding to the latter has phase values such that<br />
ǫ j<br />
tl<br />
(t) = ǫj<br />
br<br />
(t) <strong>and</strong> ǫj<br />
bl<br />
(t) = ǫj tr(t) ∀ j = 1...Ns. For this shape, the values of<br />
diagonally adjacent phase points are thus equal, which corresponds to the checkerboardlike<br />
phase point coloring in figure 1.6. The implications of these unseen modes for <strong>control</strong><br />
will be addressed in the following chapters.<br />
1.3.3 The wavefront corrector<br />
The wavefront corrector performs the physical correction of the wavefront. A wide variety<br />
of wavefront correctors exists. The goal of this section is to show the diversity <strong>and</strong> to point<br />
out the main properties of the different correctors.
1.3 <strong>Adaptive</strong> optics 9<br />
Probably the oldest wavefront corrector is the segmented <strong>mirror</strong>. This <strong>mirror</strong> consists of a<br />
number of small, closely packed <strong>mirror</strong> segments that can move in one or three Degrees<br />
Of Freedom. In the first case the individual <strong>mirror</strong> elements can only move up <strong>and</strong> down<br />
(piston) along the optical axis. In the second case each <strong>mirror</strong> segment can rotate over<br />
two orthogonal axes of tilt as well. Piezoelectric actuators <strong>and</strong> strain gauges are most<br />
commonly used to move the segments <strong>and</strong> to provide position feedback. One example<br />
is the segmented <strong>mirror</strong> from ThermoTrex Corporation, which is used on the 4.2 meter<br />
William Herschel Telescope (WHT) [198]. This <strong>mirror</strong> has 76 <strong>mirror</strong> segments, each of<br />
which have tip/tilt <strong>and</strong> piston actuation giving a total of 228 Degrees Of Freedoms (DOFs).<br />
Other examples can be found in [37], [104], [97] <strong>and</strong> [76]. By having separated segments<br />
there is no cross coupling <strong>and</strong> <strong>mirror</strong> parts can be relatively easily replaced. Furthermore,<br />
as will be discussed in the following chapters, the lack of cross coupling between actuators<br />
simplifies the <strong>control</strong> design. These advantages come at the cost of small gaps between the<br />
segments that act as an optical grating <strong>and</strong> cause diffraction.<br />
Another type of wavefront corrector is the DM. Most DM’s have continuous facesheets<br />
that are deformed out-of-plane by stacks of piezoelectric actuators placed perpendicularly<br />
under the reflective surface. This type of DM has been under development since 1974 <strong>and</strong><br />
was first built for high energy laser systems [179]. At the end of the ’70s these <strong>mirror</strong>s were<br />
also developed for infrared systems [46, 47]. Current development of this type of <strong>mirror</strong><br />
is driven by miniaturization [164], increasing actuator linearity, stroke [154] <strong>and</strong> position<br />
accuracy, decreasing operating voltages, drift [28, 154] <strong>and</strong> hysteresis [165]. Piezo stacked<br />
<strong>deformable</strong> <strong>mirror</strong>s are made a.o. by Xinetics, CILAS <strong>and</strong> OKO Technologies. One of<br />
the largest AO-systems with a piezo stacked <strong>mirror</strong> is in the 10-meter Keck telescopes.<br />
Here a 349-channel piezoelectric <strong>mirror</strong> from Xinetics is implemented [194]. Besides<br />
piezoelectric materials also Lead Manganese Niobate (PMN) or magnetostrictive actuators<br />
are used.<br />
A separate class of continuous facesheet DM’s are the bimorph <strong>mirror</strong>s. Unlike the DM’s<br />
with stacked piezo actuators, bimorph <strong>mirror</strong>s have actuators placed parallel to the reflective<br />
surface. A bimorph <strong>mirror</strong> usually consists of a glass or metal facesheet that is bonded<br />
to a sheet of piezoelectric ceramic. There’s a conductive electrode in the bond between<br />
the piezoelectric material <strong>and</strong> the facesheet. On the backside of the ceramic a series of<br />
electrodes is attached. When a voltage is applied between the front <strong>and</strong> back electrode the<br />
dimensions of the piezoelectric material change <strong>and</strong> a local radius of curvature is forced<br />
into the <strong>mirror</strong>. Bimorph <strong>mirror</strong> were first used in astronomy in the beginning of the ’90s on<br />
the Canada France Hawaii Telescope (CFHT) [149]. One of the largest bimorph <strong>mirror</strong>s is<br />
a 188-element bimorph <strong>mirror</strong>, developed by CILAS, <strong>and</strong> currently used in the AO-system<br />
for the 8.2-meter SUBARU telescope. This <strong>mirror</strong> is 130mm across, but only the inner<br />
90mm is illuminated [172]. The remaining 40 electrodes outside this diameter are needed<br />
to enforce the proper boundary conditions [140].<br />
In bimorph <strong>mirror</strong>s the local curvature is proportional with the voltage <strong>and</strong> the coefficient<br />
of the dielectric tensor <strong>and</strong> inversely proportional with the square of the thickness. The<br />
maximum voltage is given by the breakdown voltage. This also determines the gap between<br />
the electrodes <strong>and</strong> thereby sets a limit for the actuator density. Since the mechanical<br />
resonance frequency is mainly determined by the diameter-thickness ratio it is clear that a<br />
trade-off between <strong>mirror</strong> size, resonance frequency <strong>and</strong> stroke (curvature) is to be made.<br />
Critical in the design are the bonds between the different layers. Bimorph <strong>mirror</strong>s suitable<br />
1
1<br />
10 1 Introduction<br />
for high power lasers with integrated cooling have also been developed [7, 156, 190].<br />
Besides piezo stacked <strong>and</strong> bimorph <strong>mirror</strong>s a few implementations exist with actuators that<br />
introduce bending moments at the edge of the <strong>mirror</strong> [67].<br />
To reduce the background emissivity from surfaces added by the AO system the number<br />
of reflective surfaces in astronomical telescopes should be kept to a minimum. This is<br />
especially the case for Infra Red (IR) observations. From this thought the idea for an <strong>Adaptive</strong><br />
Secondary Mirror (ASM) was born in the ’90s [155]. In contrast with the previously<br />
discussed correctors, secondary <strong>mirror</strong>s in a telescope are usually strongly curved, giving<br />
additional challenges in making them adaptive. The first ASM was built for the 6.5m<br />
Multiple Mirror Telescope (MMT) in Arizona in the mid ’90s <strong>and</strong> has 336 actuators [120].<br />
The static secondary <strong>mirror</strong> was replaced by a thin <strong>deformable</strong> zerodur shell with a radius<br />
of curvature of 1795mm. The shell is 1.9mm thick <strong>and</strong> 640mm in diameter [127]. In the<br />
center of the shell a membrane suppresses the lateral DOF’s. A total of 336 small, radially<br />
magnetized permanent magnets are glued to the backside of the zerodur shell <strong>and</strong> together<br />
with voice coils that are fixed in a reference plate form actuators that push <strong>and</strong> pull at the<br />
shell. Capacitive sensors are placed concentrically with the actuators in between the Ultra<br />
Low Expansion (ULE) glass reference structure <strong>and</strong> the backside of the thin shell. They<br />
provide distance measurements for the local feedback loops. A 30mm thick aluminium<br />
plate with cooling channels is used to drain the produced heat [24, 25, 101, 181, 182].<br />
After the conversion at the MMT, two ASM’s were made for the Large Binocular<br />
Telescope (LBT). The ASMs for both of its telescopes have a radius of curvature of<br />
1974.2mm <strong>and</strong> measure 911mm across. To reduce the deformation forces <strong>and</strong> resulting<br />
power dissipation a 1.6mm thick zerodur shell was chosen. Each of these shells have 672<br />
electromagnetic actuators [68].<br />
One of the Very Large Telescope (VLT)’s will also be equipped with an ASM for which<br />
first light is foreseen in 2015. This <strong>mirror</strong> has a radius of curvature of 4553mm, measures<br />
1120mm across <strong>and</strong> is equipped with 1170 actuators [9].<br />
The above ASMs exhibit a few drawbacks, one of which is their high complexity. Due to<br />
the lack of mechanical stiffness in the thin shell, hundreds of eigenfrequencies lie below or<br />
around the desired <strong>control</strong> b<strong>and</strong>width <strong>and</strong> need to be dealt with by the <strong>control</strong> system. To<br />
achieve this, each actuator is equipped with a capacitive sensor <strong>and</strong> associated electronics<br />
<strong>and</strong> needs a significant amount of computational power for closed-loop <strong>control</strong> [193].<br />
Research on <strong>control</strong>ling this thin shell is still ongoing [151]. Since the power consumption<br />
is high (MMT:2kW [120], LBT:2.665kW [18], VLT:1.47kW [9]), active fluidic cooling is<br />
needed for which leakage is known to occur [68]. Furthermore the gap between the thin<br />
shell <strong>and</strong> the reference structure is only ∼ 50µm, rendering contamination a serious risk<br />
[68]. The assemblies have high masses (MMT:130kg, LBT:250kg, VLT:180kg), resulting<br />
in mechanical resonances at low frequencies. As a result, winds causes the assembly hub<br />
to resonate in its metering structure leading to optical degradation. At the MMT, these<br />
resonances lie at 14Hz for the rotation mode perpendicular to the optical axis <strong>and</strong> 19Hz<br />
for the mode along the optical axis. Additional measures had to be taken to reduce the<br />
detrimental effects hereof [167].<br />
The last type of continuous facesheet <strong>deformable</strong> <strong>mirror</strong> to be discussed here is the<br />
membrane <strong>mirror</strong>. A very thin (
1.3 <strong>Adaptive</strong> optics 11<br />
a voltage to the electrostatic electrode actuators it is possible to deform the membrane. In<br />
most cases a bias voltage is applied to all the electrodes, to make the membrane initially<br />
spherical. This way, the membrane can be moved in both directions. Probably the most<br />
widely spread example is the 37 actuator electrostatic <strong>deformable</strong> <strong>mirror</strong> from OKO<br />
Technologies [122, 180]. Due to the thickness of the membrane these <strong>mirror</strong>s are very<br />
fragile. Critical in the design is to avoid possible snap down <strong>and</strong> avoid dust in the very<br />
narrow gaps. Another actuation method on membrane <strong>mirror</strong>s can be found in [38, 39].<br />
Here a small magnet is suspended by the membrane <strong>and</strong> coils are used to exert a force <strong>and</strong><br />
deform it. Since no mechanical stiffness exists, scaling to large diameters is not possible<br />
while retaining inter actuator stroke <strong>and</strong> density as well as dynamic properties.<br />
Micro-Electro-Mechanical Systems (MEMS) devices form another class of small DMs.<br />
With the potential to be fabricated in large quantities <strong>and</strong> with large numbers of actuators<br />
this seems a promising technique. Nevertheless, most MEMS suffer from limited<br />
(inter)actuator stroke <strong>and</strong> poor surface quality. MEMS DMs are manufactured by Boston<br />
Micromachines <strong>and</strong> Iris AO.<br />
Finally, note that not all wavefront correctors are based on reflection. High-order transmission<br />
based correctors are also available [121]. Most of these correctors are based on<br />
liquid crystals <strong>and</strong> are limited in stroke <strong>and</strong> dynamic behavior. For astronomy this requires<br />
them to be used in woofer-tweeter configurations, where they are located after low-order<br />
DMs that first correct the large, slow part of the wavefront distortion. Such multi-<strong>mirror</strong><br />
configurations will be discussed in the next paragraph.<br />
1.3.4 Optical configurations<br />
Although most AO systems only have a single DM, more advanced configurations have<br />
been designed that use multiple wavefront correctors. Such configurations allow a larger<br />
Field Of View (FOV) <strong>and</strong>/or the observation of multiple stars simultaneously. The FOV is<br />
the area of the sky that can be observed with a certain resolving power at a single moment<br />
in time (i.e. without rotating the telescope). Using a single DM for wavefront correction the<br />
<strong>and</strong> the light of a single bright star (natural guide star) for wavefront sensing, the correction<br />
quality will degrade as the object under observation is located further away from the guide<br />
star. The area around the guide star for which the wavefront errors can be successfully<br />
compensated is called the isoplanatic patch <strong>and</strong> forms the telescope’s useful FOV. When<br />
laser beacons (laser guide stars) [98] are used to supply the light for the wavefront sensor,<br />
the location of the guide star can be chosen, thus reducing the problem of anisoplanatic<br />
errors. However, laser guide stars introduce errors of a different, more complex character<br />
that are outside the scope of this thesis. The multi-<strong>mirror</strong> configurations that will now be<br />
briefly discussed, share the concept of using not only multiple wavefront correctors, but<br />
also multiple wavefront sensors that are aimed at different guide stars.<br />
A first approach to allow observation of multiple stars simultaneously is Multi-Object<br />
<strong>Adaptive</strong> Optics (MOAO) [10]. Since the light of each object has a different path through<br />
the turbulent atmosphere, the wavefront disturbance slightly varies per object. In MOAO<br />
the incoming starlight is split in the focal plane into multiple areas that are each corrected<br />
by a separate <strong>mirror</strong>, such that multiple objects over a wider FOV can be observed with<br />
1
1<br />
12 1 Introduction<br />
AO on the same telescope. The wavefront disturbance information relevant for each area<br />
is extracted from the measurements of multiple wavefront sensors using atmospheric tomography<br />
[10, 146, 196]. By combining the measurements of sufficiently many wavefront<br />
sensors at known locations <strong>and</strong> orientations, atmospheric tomography allows to reconstruct<br />
the wavefront at any point – up to a certain resolution – in the atmosphere. For MOAO<br />
these points are chosen within the area of the object in the focal plane for which correction<br />
is desired.<br />
Tomography is also used for another approach that leads to an increased FOV, called Multi-<br />
Conjugate <strong>Adaptive</strong> Optics (MCAO). Here each wavefront corrector is optically conjugated<br />
to a specific atmospheric turbulence layer. The effect of these layers can be corrected<br />
based on information extracted from measurements of multiple wavefront sensor using atmospheric<br />
tomography.<br />
A last approach, which allows the observation of multiple objects while using only a single<br />
corrector is called Ground Layer <strong>Adaptive</strong> Optics (GLAO). This is based on the previously<br />
mentioned fact that the lowest layer in the earth’s atmosphere (the ground layer) is the most<br />
detrimental (figure 2.2). Correction of this layer leads to a significant increase of the optical<br />
quality over a very wide FOV. Therefore, the single corrector is optically conjugated to the<br />
ground layer <strong>and</strong> atmospheric tomography on measurements of multiple WFSs provides the<br />
necessary disturbance information.<br />
1.3.5 The <strong>control</strong> system<br />
A last, but vital part of the AO system is the <strong>control</strong> system. Based on mathematical formulae,<br />
this processes wavefront sensor measurements in real time to determine suitable<br />
comm<strong>and</strong> setpoints for the actuators. The formulae consist of several parts [98, 100]. As<br />
mentioned the WFS sensors are often CCD based <strong>and</strong> require both an image processing <strong>and</strong><br />
a mathematical transformation step to determine the wavefront phase shape. Subsequently,<br />
the formulae include models of the behavior of the wavefront corrector <strong>and</strong> if desired the<br />
spatial <strong>and</strong>/or temporal <strong>dynamics</strong> of the behavior of the (atmospheric) wavefront distortion.<br />
The ultimate goal of the <strong>control</strong> system is to compensate this distortion, the quality of which<br />
can e.g. be measured using the Strehl ratio (section 1.2). According to the Maréchal approximation<br />
[20, 179], this ratio is inversely proportional to the variance of the wavefront<br />
aberration (section 2.2). Therefore, the goal for the <strong>control</strong> system can be formulated more<br />
specifically as to minimize this variance. In other words, the AO <strong>control</strong> problem is to<br />
find the <strong>control</strong> law formulae that lead to the smallest residual wavefront aberration in a<br />
least squares sense. When assuming the noise <strong>and</strong> disturbance to be generated from Gaussian,<br />
white noise, this means that the <strong>control</strong> problem for AO fits the Linear Quadratic<br />
Gaussian (LQG) <strong>and</strong>H2 optimal <strong>control</strong> frameworks [100]. This will be further elaborated<br />
in chapter 3.<br />
Implementation<br />
At the time when the first AO system was built, computers were far too slow to suitably<br />
implement a <strong>control</strong> law <strong>and</strong> dedicated analog circuits were used instead [98]. But the computational<br />
hardware of AO systems has always been cutting edge, as required by the fast<br />
update rates, high numbers of actuators <strong>and</strong> sensors <strong>and</strong> the high cross-coupling required
1.3 <strong>Adaptive</strong> optics 13<br />
WFS<br />
RTC<br />
corrector<br />
incoming<br />
starlight<br />
beam<br />
splitter corrector<br />
science<br />
camera<br />
Figure 1.8: Schematic of an AO system configures<br />
in open-loop.<br />
incoming<br />
starlight<br />
RTC<br />
beamsplitter<br />
WFS<br />
science<br />
camera<br />
Figure 1.9: Schematic of an AO system configured<br />
in closed-loop.<br />
between <strong>control</strong>ler inputs <strong>and</strong> outputs. Nowadays, the formulae are implemented on stateof-the-art<br />
digital computing hardware. For instance, the <strong>control</strong> system for the MAD system<br />
[58] is implemented on four power-PC’s. For other, large AO systems dedicated Field Programmable<br />
Gate Array (FPGA) boards are often used that can perform many calculations<br />
in parallel [70]. The data processors obtain the measurements via a fast, usually digital<br />
communication link from the WFS <strong>and</strong> then apply the <strong>control</strong> law to determine the actuator<br />
comm<strong>and</strong>s. These are then communicated to Digital to Analog Convertors (DACs) <strong>and</strong><br />
applied to the actuators.<br />
Control configurations<br />
AO systems can be configured in both open-loop (figure 1.8) as well as closed-loop (figure<br />
1.9). Both configurations have their advantages <strong>and</strong> both are used in practice. In the openloop<br />
configuration, the measurements are not influenced by the shape of the corrector <strong>and</strong><br />
provide direct information on the wavefront distortion.<br />
However, in case of strong turbulence the wavefront distortion may exceed the range of<br />
the WFS, leading to poor performance. In closed-loop <strong>control</strong> the sensor measures the<br />
corrected wavefront, which requires a smaller sensor range. Further, as the effect of the<br />
<strong>control</strong> actions is not observed by the open-loop <strong>control</strong> system, this must completely rely<br />
on a model that accurately describes the behavior of the wavefront corrector. But for the<br />
same reason this model cannot be calibrated in this configuration. On the other h<strong>and</strong>, as<br />
long as the <strong>control</strong>ler is stable an inaccurate model cannot lead to instabilities. This is in<br />
contrast to the closed-loop case, in which model mismatch may lead to an unstable system.<br />
But this seems a reasonable price to pay for solving all previously mentioned issues of the<br />
open-loop configuration. Therefore, throughout this thesis the closed-loop <strong>control</strong> system<br />
of figure 1.9 will be considered.<br />
1
1<br />
14 1 Introduction<br />
1.4 Challenges<br />
At the moment, optical telescopes being built have an aperture diameter of at most 10m.<br />
Examples are the VLTs, the LBT, the Keck <strong>and</strong> SUBARU telescopes. AO systems for these<br />
telescopes are largely ready, but will initially only skim the surface of the full correction<br />
potential. Much is still to be expected from wavefront correctors with more DOFs, faster<br />
dynamic behavior <strong>and</strong> better optimized <strong>control</strong> laws.<br />
Nevertheless, the observation of even fainter <strong>and</strong> more distant celestial objects requires even<br />
larger apertures. Telescopes are currently being designed with aperture diameters ranging<br />
between 30 <strong>and</strong> 50m. A consortium in the USA has conceived the Thirty Meter Telescope<br />
(TMT) with an aperture diameter of 30m [171]. European Southern Observatory (ESO)<br />
has conceived a 42m aperture telescope called the European Extremely Large Telescope<br />
(E-ELT) [75], which is in fact a smaller version of the originally planned Overwhelmingly<br />
Large Telescope (OWL) 1 with a diameter of 100m [43]. Further, a 50m aperture telescope<br />
was being designed by a consortium around the Swedish Lund University <strong>and</strong> called the<br />
EURO-50 [6]. This is now superceded by the E-ELT.<br />
The design of AO systems for such large telescopes involves serious challenges for all parts<br />
of the AO system. In this research project these challenges have been investigated for both<br />
the wavefront corrector <strong>and</strong> the <strong>control</strong> system. In the next subsections, these challenges<br />
will be discussed more elaborately.<br />
1.4.1 The wavefront corrector<br />
As will be discussed in the next chapter, a constant optical quality corresponds to a constant<br />
actuator density. Therefore, the number of <strong>control</strong>lable degrees of freedom of DMs for the<br />
mentioned future large telescopes must be in the order of tens of thous<strong>and</strong>s. The highest<br />
number of actuators currently available for DMs is∼1000 costing around 1ke per actuator.<br />
For several reasons it is not trivial to extend current designs to larger numbers of actuator:<br />
• Extendability. Straightforward extension of many current DM designs leads to an<br />
increase in mass that cannot be matched by stiffness <strong>and</strong> thus leads to a severe reduction<br />
of the resonance frequencies. Low resonance frequencies reduce the achievable<br />
<strong>control</strong> b<strong>and</strong>width <strong>and</strong> thus the achievable wavefront correction performance. Extendability<br />
is not only needed for the mechanics but also for the <strong>control</strong> system <strong>and</strong><br />
electronics involved.<br />
• Scalability. DMs are needed with a wide range of actuator pitch. The first generation<br />
AO systems for the E-ELT will have∼30mm actuator pitch <strong>and</strong> around 8000 actuators<br />
[103]. Later generations will have an actuator pitch down to 1mm with over a 100.000<br />
actuators. No current design is available that matches these requirements.<br />
• Power dissipation. Most DM designs involve substantial power dissipation. As a<br />
consequence, the temperature of the DM surface will rise with respect to its environment,<br />
leading to detrimental air flow in the path of light. To prevent this, an active<br />
1 N.B. OWL also refers to the bird for its keen night vision.
1.4 Challenges 15<br />
cooling system is required, which adds complexity <strong>and</strong> the risk of leakage. Moreover,<br />
the fluid flow will introduce vibrations on the nm level that affect the wavefront<br />
correction performance.<br />
• Failure probability. As the number of actuators increases, the probability of defective<br />
actuators also increases. When actuators have a high stiffness, a defect actuator fixes<br />
the DM position at its position. This creates a so-called hard point that will affect a<br />
large fraction of the <strong>mirror</strong> area <strong>and</strong> thus degrade its performance. Besides developing<br />
actuators with a high Mean Time Before Failure (MTBF), defective actuators should<br />
thus not cause a significant decrease in the optical surface quality.<br />
• The price per channel. Given the 500Me total budget of the Extremely Large<br />
Telescope (ELT) <strong>and</strong> the current cost per channel of 1ke, an AO system for these<br />
telescopes with in the order of 100.000 channels will not be affordable.<br />
Extendable <strong>and</strong> scalable <strong>mirror</strong> design is needed, in mechanics, electronics <strong>and</strong> <strong>control</strong> with<br />
lightweight construction with high resonance frequencies, low power dissipation <strong>and</strong> soft<br />
<strong>and</strong> cheap actuators. As a starting point for further requirements an 8m telescope on a<br />
representative astronomical site is chosen. In this research project a design was proposed<br />
that is driven by above-mentioned reasons. This design will be sketched in the next chapter<br />
<strong>and</strong> is described in more detail in [174]. The focus of this thesis will lie with the electronics<br />
<strong>and</strong> the <strong>control</strong> system, whose challenges will be elaborated in the next subsection.<br />
1.4.2 The <strong>control</strong> system<br />
One of the main challenges with early AO systems was that for the system to be able to<br />
correct for atmospheric wavefront disturbances the shape of the DM needed to be updated<br />
around 1000 times per second. The first functioning AO system was finished in 1974<br />
[98], having a 21 channel DM <strong>and</strong> a shearing interferometer wavefront sensor measuring<br />
32 slopes simultaneously. Similar to the Shack-Hartmann wavefront sensor described<br />
above, a shearing sensor also does not provide direct information of the wavefront phase,<br />
but via a spatial transformation. Inversion of this transformation (reconstruction) <strong>and</strong><br />
subsequent calculation of suitable actuator comm<strong>and</strong> signals are computationally costly<br />
operations, which would have taken more than a day on a contemporary computer [98].<br />
Instead, an analog electronic circuit was designed in which measurements were introduced<br />
as <strong>control</strong>led currents, yielding the actuator comm<strong>and</strong>s as measurable voltages. Although<br />
this <strong>control</strong>ler structure was very inflexible, it could perform the reconstruction step within<br />
microseconds, which is fast even for today’s st<strong>and</strong>ards.<br />
Other approaches towards fast update rates involved the use of a simple photodiode as a<br />
WFS. This was placed in a focal plane to measure the light intensity at the center of the<br />
guide star image, which is a measure for the Strehl ratio. The <strong>control</strong>ler would then quickly<br />
superpose a series of shapes to the wavefront corrector <strong>and</strong> measure the corresponding<br />
effects in light intensity. A suitable set of actuator comm<strong>and</strong>s was then calculated from<br />
the results using relatively simple computations. However, the length of the series of<br />
shapes <strong>and</strong> measurements required by this method is equal to the number of actuators of<br />
the corrector. As this number increases, so must the number of measurements <strong>and</strong> to keep<br />
the sampling rate constant the measurements need to be performed faster. The limits are<br />
1
1<br />
16 1 Introduction<br />
here not determined by the intensity measurements, but by the <strong>dynamics</strong> of the wavefront<br />
corrector, which limit the speed at which it can change shape.<br />
A final AO system design aimed at reducing <strong>control</strong>ler complexity is the combination of<br />
a CS with a bimorph <strong>mirror</strong>. As mentioned in section 1.3.3, the actuators of this type of<br />
<strong>mirror</strong> introduce local curvature to the surface. Consequently, the static mapping from<br />
actuator comm<strong>and</strong>s to measurements is almost identity, which means that a computationally<br />
expensive reconstruction step can be avoided in the <strong>control</strong> law [113, 179]. However, as<br />
discussed in the same section, the construction principle of bimorph <strong>mirror</strong>s does not allow<br />
the number of actuators to be increased without sacrificing stroke or dynamic behavior<br />
(first resonance frequency). This renders these <strong>mirror</strong>s unsuitable for future large telescopes.<br />
Currently, Shack-Hartmann sensors are the most widely used in AO systems <strong>and</strong> digital<br />
<strong>control</strong> systems have become sufficiently fast to do all required computations. However,<br />
for the future large telescopes being designed, the latter will not be trivial to maintain.<br />
Digital processors may continue to increase in computational power, but the last few years<br />
the gain in power has come mainly from parallel (multi-core) architectures instead of<br />
increased numbers of sequential computations per time unit. Without efficient algorithms,<br />
the required computational power for AO increases approximately with the square of the<br />
number of actuators <strong>and</strong> thus to the fourth power in the telescope aperture. This is plotted<br />
in figure 3.5 on page 51 based on a desired Strehl ratio of 0.87[-] <strong>and</strong> a sampling rate of<br />
1kHz. It shows that without efficient numerical algorithms an AO system for the 42m<br />
E-ELT with over 100.000 actuators would require almost 10.000 processors capable of<br />
10 giga-flops. Starting point of this graph is the traditional reconstructor plus integrator<br />
<strong>control</strong> law [98, 179]. It is yet unclear <strong>and</strong> difficult to predict how the computational<br />
dem<strong>and</strong> of recently proposed, more general optimal <strong>control</strong> laws will scale. This will<br />
require representative WFS measurement data sets from actual large telescopes that are yet<br />
unavailable.<br />
The traditional <strong>control</strong> approach will require a careful design of both hard- <strong>and</strong> software<br />
to achieve an efficient parallel computer system. In [53, 72, 187] <strong>control</strong> algorithms<br />
are shown with a computational complexity of O(N 3/2<br />
a ). But even such algorithms will<br />
require many processors to compute the setpoints for the 100.000 actuators at a rate of 1kHz.<br />
Besides computational problems, increasing the number of actuators yields many<br />
practical problems [21]. The actuation principle is usually based on electricity <strong>and</strong> requires<br />
each actuator to have at least two connection wires. In case of 100.000 actuators, this leads<br />
to 200.000 wires <strong>and</strong> thus a large probability of defects, disturbances, etc. To keep the<br />
lengths of these wires to a minimum <strong>and</strong> obtain a straightforward multi-processor hardware<br />
architecture, a <strong>modular</strong>, distributed <strong>control</strong> system is proposed.<br />
In this distributed <strong>control</strong> system each actuator or small group of actuators is driven<br />
by a separate hardware module that has direct communication links to a few neighboring<br />
modules. Each module receives a small fraction of the measurements available from the<br />
wavefront sensor <strong>and</strong> all modules are identical in hardware, but may differ in software.<br />
This allows cost-efficient production of the modules <strong>and</strong> enables both the straightforward<br />
construction of a <strong>control</strong> system hardware for large AO systems as well as quick replacement<br />
of defective modules.
1.5 Distributed <strong>control</strong> 17<br />
However, by assigning computational power per actuator, the total power increases only<br />
linearly with the number of actuators (figure 3.5), which is not sufficient for the currently<br />
available algorithms. Moreover, as will be discussed in detail in chapter 3, these algorithms<br />
are not suitable for the distributed architecture. This requires research into new algorithms<br />
whose prime design driver is the distributed structure. The performance achieved using<br />
such algorithms may be subject to the choice for specific properties of the structure, but<br />
should approximate that of traditional, centralized architectures. Such properties may be<br />
what neighbors the modules can communicate with, what information they exchange <strong>and</strong><br />
which measurements they receive. A suitable choice for these properties requires insight<br />
into their effect on the AO system’s performance.<br />
Since the AO case forms only a specific part of the spectrum of distributed <strong>control</strong>, a short<br />
literature review on this topic will be given in the next section.<br />
1.5 Distributed <strong>control</strong><br />
In the usual, centralized setting, a <strong>control</strong>ler receives all measurements available of the<br />
sensors of a system as its inputs <strong>and</strong> uses these to compute comm<strong>and</strong>s for all actuators<br />
of the plant. The opposite of this approach is called decentralized <strong>and</strong> is characterized<br />
by the fact that each actuator is <strong>control</strong>led by a separate <strong>control</strong>ler. In this setting it is<br />
commonly assumed that all <strong>control</strong>lers are uniquely associated with a sensor, such that<br />
the only coupling between the <strong>control</strong>lers is through the plant. A distributed <strong>control</strong>ler<br />
is a combination of these extremes in the sense that each actuator has its associated<br />
<strong>control</strong>ler, but now neighboring actuators are able to communicate. They are thus not only<br />
connected through the plant, but also through communication channels. In the absence of<br />
communication delays, a centralized <strong>control</strong>ler is thus equivalent to a distributed <strong>control</strong>ler<br />
for which direct communication is allowed between all pairs of <strong>control</strong>lers. In that case<br />
any centralized <strong>control</strong> law can be implemented without modification. These concepts are<br />
illustrated in figure 1.10.<br />
The reason why to use a distributed <strong>control</strong> system is always driven by requirements<br />
of an application. If a centralized <strong>control</strong> system is possible within these requirements,<br />
the complexity <strong>and</strong> limitations introduced by a distributed structure will not outweigh its<br />
benefits. Nevertheless, for quite some time already the distributed <strong>control</strong> field is receiving<br />
a lot of research attention. This is closely linked to the rise of application fields where a<br />
centralized <strong>control</strong>ler is not possible (1) or where benefits outweigh the added complexity<br />
(2). In such application fields the system to be <strong>control</strong>led often consists of multiple<br />
interacting subsystems that are to some extent physically separated <strong>and</strong> can all include<br />
an implementation of a local <strong>control</strong>ler. In application fields of the first kind it is either<br />
not practically possible to send the measurements of all sensors in the system to a central<br />
location or to send <strong>control</strong> comm<strong>and</strong>s to all actuators in the system from a central location.<br />
Reasons may be latency or unreliability of communication due to e.g. distances, problems<br />
with cabling because of the large numbers of sensors <strong>and</strong> actuators, etc.<br />
For application fields of the second kind, the distributed <strong>control</strong> benefits that outweigh the<br />
added complexity are often flexibility <strong>and</strong>/or scalability of the system. When the system<br />
itself is subject to change – e.g. the number of subsystems varies – a centralized <strong>control</strong>ler<br />
1
1<br />
18 1 Introduction<br />
Centralized Distributed Decentralized<br />
P<br />
C<br />
P<br />
C1<br />
C2<br />
Ci<br />
Cn-1<br />
Cn<br />
P<br />
C1<br />
Ci<br />
Cn-1<br />
Figure 1.10: The schematic representation of the centralized, distributed <strong>and</strong> decentralized <strong>control</strong><br />
concepts.<br />
must be dimensioned both in computational power <strong>and</strong> in communication capabilities to<br />
deal with the largest number of subsystems that is to be expected . For a distributed <strong>control</strong><br />
implementation the addition of subsystems implicitly leads to an increase in computational<br />
power <strong>and</strong> communication capabilities <strong>and</strong> thus provides flexibility. Potentially, this also<br />
leads to a scalable system – provided that the behavior <strong>and</strong> interaction of the subsystems<br />
allow for the same system performance with the same <strong>control</strong>ler structure.<br />
Examples of application fields where distributed <strong>control</strong> is developing a role are vehicle<br />
platooning [137] <strong>and</strong> automated highways, formation flying of aircrafts [200], spacecrafts<br />
[176] <strong>and</strong> satellites [128], the process industrial plants [8], optical telescope <strong>control</strong> systems<br />
[107], inflatable structures in space, paper machines [169], power networks [134], etc.<br />
For vehicle platooning or automatic highway systems the communication infrastructure<br />
that would allow all cars to send position <strong>and</strong> velocity information to a central server<br />
would be very complex <strong>and</strong> expensive <strong>and</strong> the computational power this server would<br />
require to process this information would be enormous. Since vehicles interact only locally,<br />
a distributed <strong>control</strong> system in which local vehicle <strong>control</strong>lers communicate only with<br />
nearby vehicles would greatly simplify the system implementation. Moreover, it would<br />
significantly reduce its cost, whereas the resulting system has now become flexible <strong>and</strong><br />
scalable w.r.t. the number of vehicles.<br />
The problem in formation flying of satellites is that for some applications the communication<br />
between a central server <strong>and</strong> all satellites cannot be guaranteed, regardless of the<br />
location of the server. Assuming that satellites will always be in range of at least one<br />
other, neighboring satellite, a distributed <strong>control</strong> architecture could solve this problem.<br />
An additional benefit is a reduction of the required communication range <strong>and</strong> thus power<br />
C2<br />
Cn
1.5 Distributed <strong>control</strong> 19<br />
usage.<br />
A application field where distributed <strong>control</strong> has always played a role is the <strong>control</strong> of<br />
power networks. Due to the recent rise of small, local power generators (wind-mills, solar<br />
cells, etc.), regulating the frequency <strong>and</strong> phase of the potential as well as the power through<br />
the network links has become much more complex. A centralized <strong>control</strong> system is here<br />
not an option for political reasons.<br />
Finally, distributed <strong>control</strong> finds application in small, highly specialized areas such as alignment<br />
systems for the tiles that form the primary <strong>mirror</strong>s of future large optical telescopes.<br />
Sensors measure the alignment of each tile in several DOFs w.r.t. their neighbors, based<br />
on which local <strong>control</strong>lers provide setpoints for the local actuators to achieve global shape<br />
<strong>control</strong>.<br />
The general approach towards <strong>control</strong>ler synthesis is through the minimization of a<br />
certain cost function. To enforce the so obtained <strong>control</strong>ler to have a distributed structure,<br />
constraints have to be applied, which for general plants renders the optimization problem<br />
non-convex. This implies that it is no longer known whether there is a single optimal<br />
solution or whether there a multiple or how good a c<strong>and</strong>idate solution is compared to<br />
it/them. The basis on which the currently available results on distributed <strong>control</strong> are<br />
founded is the exploiting of structure present in the system to be <strong>control</strong>led. The explicit<br />
constraints that have to be applied to general plants can be relaxed by making a priori<br />
assumptions on this structure, in some cases leading to efficient synthesis algorithms <strong>and</strong><br />
convex problems.<br />
For instance results have been shown for distributed systems satisfying certain spatial<br />
invariance properties [14, 41, 79, 129, 168]. In [13, 14] this spatial invariance is used in a<br />
Fourier domain approach, reducing the optimization to a family of problems over spatial<br />
frequency. It is moreover shown that the spatial invariance property is inherited by the<br />
<strong>control</strong>ler that is optimal w.r.t. generalLp induced norm performance criteria.<br />
In simulations, this approach has been applied for the alignment <strong>control</strong> of hexagonal<br />
segments of the primary <strong>mirror</strong> of a future large telescope [107]. First, a spatially invariant<br />
system model is here proposed for which a spatially invariant <strong>control</strong>ler is derived with<br />
known quadratic performance upper bounds. This is then truncated <strong>and</strong> applied to a 19<br />
segment <strong>mirror</strong> model. Although it is mentioned that stability may be lost in this truncation,<br />
no approach is given to recover this.<br />
In [168], a distributed <strong>control</strong>ler is sought for shape <strong>control</strong> of large two dimensional arrays.<br />
A distributed proportional+integrating (PI) <strong>control</strong>ler is proposed, where both sets of gains<br />
apply not only to local errors, but also to those of neighbors. An optimization problem<br />
is then formulated for these gains in terms of robust stability <strong>and</strong> frequency dependent<br />
norms on the shape error <strong>and</strong> <strong>control</strong> effort transfer functions. The assumption of spatial<br />
invariance then allows a transformation to the Fourier domain, where the optimal <strong>control</strong>ler<br />
parameters can be found using linear programming. This approach is further extended in<br />
[79] to include a derivative <strong>control</strong>ler action <strong>and</strong> a method to h<strong>and</strong>le the effects of finite<br />
boundaries including stability issues.<br />
In [129] the system to be <strong>control</strong>led is considered to be an interconnection of a number of<br />
identical subsystems. The interconnected systems are not necessarily spatially invariant<br />
<strong>and</strong> can have an arbitrary interconnection graph topology. An approach is presented to<br />
synthesize feedback <strong>control</strong>lers for this class of systems that retain the distributed structure.<br />
The synthesis problem is formulated as a multi-objective optimization problem with Linear<br />
1
1<br />
20 1 Introduction<br />
Matrix Inequality (LMI) constraints <strong>and</strong> system norms (e.g. H2 <strong>and</strong> H∞) as performance<br />
indices. Although this formulation involves some conservatism, the resulting optimization<br />
problem can be solved very efficiently. Moreover, the proposed approach is demonstrated<br />
– including simulation results – for classical distributed <strong>control</strong> applications such as the<br />
paper machine <strong>and</strong> satellite formation flying.<br />
A relatively new approach towards the design of distributed <strong>control</strong>lers is to apply game<br />
theory <strong>and</strong> consider the nodes <strong>and</strong> communication links as players that strive towards a<br />
common goal [147]. Synthesis of an optimal distributed <strong>control</strong>ler is then replaced by an<br />
iterative process, where the local <strong>control</strong>lers optimize their local cost functions until the<br />
point that none can be improved by adjusting local parameters alone. This point is called<br />
the Nash equilibrium, which in this approach replaces the global optimum (if this exists at<br />
all). In section 3.6.3 of this thesis a wavefront reconstruction algorithm will be presented<br />
that is based on overlapping, local optimization problems with the same limitation.<br />
A significant part of the distributed <strong>control</strong>ler synthesis problem is also present in optimal<br />
distributed filtering – i.e. distributed Kalman filtering. Also here, for specific systems<br />
satisfying spatial invariance properties, the Fourier domain approach has been shown to be<br />
of value [99]. For more general cases the approach based on weighted averaging proposed<br />
in [4] may be used. This approach has been experimentally validated for position tracking<br />
of remotely <strong>control</strong>led robots in [5]. However, a drawback of this approach is the high<br />
computational dem<strong>and</strong> arising as the system state increases. Other approaches towards<br />
distributed Kalman filtering may be found in [11, 138, 144, 166].<br />
Despite all progress <strong>and</strong> achievements, actual implementations of the above approaches are<br />
scarce. This may be explained by the fact that the successful synthesis approaches require<br />
plants to possess such particular structures that distributed <strong>control</strong> as a field remains a niche.<br />
More general assumptions on the plant structure for distributed <strong>control</strong>ler synthesis to be<br />
assimilated by industry. Therefore, future research is planned to extend such methods with<br />
Linear Parameter Varying (LPV) techniques, where certain plant parameters are allowed<br />
to vary spatially. Besides by its limitations on structure, the need for distributed <strong>control</strong> is<br />
attenuated by the fact that both the computational power of commercial processors as well<br />
as the communication capabilities of commercial devices is still growing steadily. This<br />
makes that centralized solutions remain feasible even for large systems.<br />
1.5.1 Distributed <strong>control</strong> for AO<br />
A specific problem that arises when applying available distributed <strong>control</strong> approaches to the<br />
case of AO, is that the typical AO system also has dynamic coupling through the sensor<br />
<strong>and</strong> not only through the plant. As mentioned, a SHS offers only indirect measurements –<br />
spatial gradients – of the desired quantity: wavefront phase. As will be shown in chapter 3,<br />
this forms the main challenge of the distributed <strong>control</strong> problem for AO. Moreover, these<br />
measurements are not collocated with the actuators. In fact, it is shown in appendix A that<br />
the best quality of SHS measurement is obtained when the spots are placed in between the<br />
actuators. As a result it becomes less trivial to associate sensors to specific <strong>control</strong>lers. In the<br />
approaches presented in chapters 3 <strong>and</strong> 4 of this thesis, the sensors are therefore assigned to<br />
multiple <strong>control</strong>lers. Chapter 3 also contains a more extensive survey on distributed <strong>control</strong><br />
literature for AO, although the available literature for this application is limited.
1.6 Problem formulation <strong>and</strong> organization of this thesis 21<br />
1.6 Problem formulation <strong>and</strong> organization of this thesis<br />
Existing large en future even larger telescopes can only be utilized fully, when they are<br />
equipped with AO systems that enhance the telescopes resolution to the diffraction limit.<br />
The development of new DM technology that meets these requirements is therefore essential.<br />
This thesis will focus on the design, testing <strong>and</strong> <strong>control</strong> of a new DM that is extendable<br />
<strong>and</strong> scalable in mechanics, electronics <strong>and</strong> <strong>control</strong>. Since this thesis is a result of a joint<br />
research project there is an accompanied thesis, by Roger Hamelinck [174], on the comprehensive<br />
design of the new AO system.<br />
This thesis is organized as follows. In the next chapter the design requirements for the new<br />
DM will be introduced, based on available knowledge of the wavefront distortion <strong>and</strong> a desired<br />
optical quality in terms of Strehl ratio. This leads to a conceptual design of the DM<br />
system with requirements for <strong>control</strong>, communication <strong>and</strong> driver electronics.<br />
In chapter 3 the problem of <strong>control</strong>ler design for AO systems will be analyzed w.r.t. the increasing<br />
size of optical telescopes. The advantages <strong>and</strong> limitations of (efficient) algorithms<br />
described in literature will be discussed <strong>and</strong> problems arising for future large telescopes<br />
will be described. The latter include cabling for many closely placed actuators, but also the<br />
computational power required to implement <strong>control</strong> laws for many thous<strong>and</strong>s of in- <strong>and</strong> outputs.<br />
At this point a <strong>modular</strong>, distributed <strong>control</strong> system architecture will be proposed. Such<br />
a hardware design can be built for a DM system with any number of actuators. Although<br />
there are no direct scaling problems, its main limitations are the computational power that<br />
increases linearly with the number of actuators <strong>and</strong> the fixed communication structure. The<br />
applicability of currently available efficient <strong>control</strong> algorithms for AO will be evaluated for<br />
this architecture, but they are found unsuitable.<br />
On the other h<strong>and</strong>, several design concepts <strong>and</strong> DM properties facilitate the use of a <strong>modular</strong>,<br />
distributed <strong>control</strong> system. One of these properties is that – when assuming a certain<br />
extent of frozen flow behavior – the disturbance to be suppressed can be well predicted over<br />
a short time horizon using only local information [45, 100]. At the end of the chapter an<br />
adaptive scheme is shown that exploits this fact to perform wavefront reconstruction <strong>and</strong><br />
prediction within the distributed framework.<br />
Further, in chapter 4 another approach towards distributed <strong>control</strong>ler design is proposed<br />
that also exploits local predictability of the disturbance. An ideal DM is considered whose<br />
transfer matrix is equal to the identity matrix <strong>and</strong> a SHS is considered with a full sample<br />
delay. The <strong>control</strong>ler structure is chosen a priori as a network of output interconnected<br />
Auto-Regressive Moving Average (ARMA) filters, whose coefficients are identified from<br />
open-loop measurement data. Several approaches are presented to guarantee the stability of<br />
this open-loop <strong>control</strong>ler via constrained optimization. Application results are shown both<br />
for data obtained from an AO breadboard at TNO Science <strong>and</strong> Industry as well as for synthetic<br />
data generated according to well known turbulence models.<br />
In chapter 5 the variable reluctance actuator design is introduced that forms the heart of the<br />
new DM. A detailed model is derived that comprehends the electromagnetic <strong>and</strong> mechanical<br />
domains of its behavior. Measurement results are presented of a realized prototype,<br />
which are compared to the model. Model parameters are identified from the measurement<br />
data <strong>and</strong> compared to first principle estimates. Some differences are found, but in general<br />
the measurements <strong>and</strong> parameters agree well with the derived model <strong>and</strong> first principle estimates.<br />
With minor modifications, the single actuator design is then transformed to a design<br />
1
1<br />
22 1 Introduction<br />
of actuator grid modules consisting of 61 hexagonally arranged actuators. Their design<br />
is presented together with measurement results obtained from seven realized prototypes.<br />
The found differences with the single actuators are explained via a sensitivity analysis on<br />
the derived model <strong>and</strong> the known design changes. The identified actuator properties such<br />
as its stiffness, resonance frequency, motor constant, inductance <strong>and</strong> viscous damping are<br />
presented for all actuators <strong>and</strong> their variation is analyzed <strong>and</strong> discussed. Finally, recommendations<br />
for future design improvements are given based on the mentioned sensitivity<br />
analysis.<br />
In chapter 6 the design, modeling <strong>and</strong> realization of the driver <strong>and</strong> communication electronics<br />
for the actuator is considered. First the requirements are derived, leading to a design<br />
that has been realized by QPI (the former EMDES). The driver electronics are realized<br />
in modules containing drivers for 61 actuators. Each driver is chosen as a Pulse Width<br />
Modulation (PWM) voltage source together with a second order analog low-pass filter.<br />
Three FPGAs implement the PWM generators for all 61 actuators together with the serial<br />
Low Voltage Differential Signalling (LVDS) communication protocol. The actuator model<br />
is extended with the electronic driver circuit <strong>and</strong> a delay to describe the serial communication<br />
link. The model thus describes the single actuator behavior from the its setpoint to its<br />
position, velocity <strong>and</strong> several electrical quantities. It is validated using measurements on<br />
the seven realized actuator modules <strong>and</strong> several model parameters will again be estimated<br />
from the results.<br />
In chapter 7 the actuator modules are connected to the reflective membrane to form a full<br />
AO system. A model of this full system is derived by coupling a number of single actuator<br />
instances through a model of the reflective membrane. This model is then analyzed w.r.t.<br />
influence functions, resonance frequencies <strong>and</strong> mode shapes, impulse response <strong>and</strong> transfer<br />
functions. Further, it is validated using measurements on a 61 actuator DM prototype. Static<br />
measurements (i.e. influence function <strong>and</strong> flatness) are obtained using interferometers <strong>and</strong><br />
dynamic measurements using white noise excitation <strong>and</strong> a laser vibrometer. An LTI model<br />
is identified from the measurement data <strong>and</strong> compared to the analytically derived model.<br />
The thesis finishes by stating general conclusions <strong>and</strong> recommendations for future research.<br />
1.7 Scientific contributions<br />
R. Ellenbroek, M. Verhaegen, R. Hamelinck, N. Doelman, M. Steinbuch, <strong>and</strong> N. Rosielle.<br />
Distributed <strong>control</strong> in <strong>Adaptive</strong> Optics - Deformable <strong>mirror</strong> <strong>and</strong> turbulence modeling. In<br />
B. L. Ellerbroek <strong>and</strong> D. C. B. Calia, editors, Proceedings of SPIE: Astronomical telescopes<br />
<strong>and</strong> instrumentation - Advances in <strong>Adaptive</strong> Optics, volume 6272, May 2006.<br />
R. Hamelinck, N. Rosielle, M. Steinbuch, R. Ellenbroek, M. Verhaegen, <strong>and</strong> N. Doelman.<br />
Actuator tests for a large <strong>deformable</strong> membrane <strong>mirror</strong>. In B. L. Ellerbroek <strong>and</strong> D. C. B.<br />
Calia, editors, Proceedings of SPIE: Astronomical telescopes <strong>and</strong> instrumentation -<br />
Advances in <strong>Adaptive</strong> Optics, volume 6272, May 2006.<br />
R. Hamelinck, R. Ellenbroek, N. Rosielle, M. Steinbuch, M. Verhaegen, <strong>and</strong> N. Doelman.<br />
Validation of a new adaptive <strong>deformable</strong> <strong>mirror</strong> concept. In N. Hubin, C. E. Max, <strong>and</strong> P. L.<br />
Wizinowich, editors, Proceedings of SPIE: Astronomical telescopes <strong>and</strong> instrumentation,
1.7 Scientific contributions 23<br />
volume 7015, Marseille, France, June 2008.<br />
R. Ellenbroek <strong>and</strong> R. Hamelinck. Adaptief deformeerbare spiegel voor telescopen. Precisietechnologie<br />
jaarboek, 16:112–118, 2009.<br />
1
ÔØÖØÛÓ<br />
×ÒÖÕÙÖÑÒØ×Ò×Ò<br />
ÓÒÔØ<br />
The main requirements for the adaptive <strong>deformable</strong> <strong>mirror</strong> <strong>and</strong> <strong>control</strong> system<br />
are derived for typical atmospheric conditions. The spatial <strong>and</strong> temporal<br />
properties of the atmosphere are covered by the spatial <strong>and</strong> temporal spectra<br />
of the Kolmogorov turbulence model <strong>and</strong> the frozen flow assumption. The<br />
main sources for the residual wavefront aberrations are identified. The fitting<br />
error, caused by a limited number of actuators <strong>and</strong> the temporal error, caused<br />
by a limited <strong>control</strong> b<strong>and</strong>width, are considered to be the most important for the<br />
<strong>mirror</strong> design. A balanced choice for the number of actuators <strong>and</strong> the <strong>control</strong><br />
b<strong>and</strong>width is made for a desired optical quality after correction. Then the actuator<br />
requirements are defined, such as the pitch, total stroke <strong>and</strong> inter-actuator<br />
stroke, resolution <strong>and</strong> power dissipation. Requirements are derived for the<br />
<strong>control</strong> system <strong>and</strong> the electronics. Finally, the full Deformable Mirror (DM)<br />
system design concept is presented, consisting of the thin <strong>mirror</strong> facesheet, the<br />
<strong>mirror</strong>-actuator connection, the actuators, the <strong>control</strong> system, the electronics<br />
<strong>and</strong> the base frame.<br />
Sections 2.1, 2.2, 2.3 <strong>and</strong> 2.5 are joint work with Roger Hamelinck [PhD]<br />
25
2<br />
26 2 Design requirements <strong>and</strong> design concept<br />
2.1 Requirements<br />
The goal is to make a DM that can correct a wavefront of an 8-meter telescope in visible<br />
light, which is aberrated by atmospheric turbulence to the diffraction limit. The <strong>mirror</strong>’s<br />
main requirements will be derived from the spatial <strong>and</strong> temporal properties of typical atmospheric<br />
conditions as they exist on astronomical sites such as Cerro Paranal in Chile. These<br />
conditions will be shown to determine the number of actuators, the (inter) actuator stroke<br />
<strong>and</strong> the <strong>control</strong> b<strong>and</strong>width. Further, the <strong>mirror</strong> should have low roughness <strong>and</strong> high reflection<br />
for the wavelengths utilized <strong>and</strong> be functional in a temperature range between−10 ◦ C<br />
<strong>and</strong>30 ◦ C [85].<br />
The <strong>mirror</strong> surface may not heat up more than 1K relative to the environment to prevent<br />
the <strong>deformable</strong> <strong>mirror</strong> itself to become a significant heat source. Finally, the number of<br />
sensors <strong>and</strong> actuators in the <strong>Adaptive</strong> Optics (AO) system will be of such order of magnitude<br />
that efficient <strong>control</strong> algorithms are required to prevent problems in the realization<br />
of suitable computation hardware. Known efficient <strong>control</strong> algorithms such as proposed in<br />
[53, 72, 129, 188] exploit the structure present in a system to obtain efficient implementations.<br />
For AO applications, such algorithms exploit sparsity or spatial invariance of the<br />
Deformable Mirrors (DMs) influence matrix <strong>and</strong> generally comprehend its temporal <strong>dynamics</strong><br />
only in terms of a number of samples delay. The DM to be designed should behave<br />
accordingly up to a sampling time scale defined in section 2.4.<br />
2.1.1 Atmospheric turbulence<br />
In section 1.2 it is explained that refractive index variations of the atmosphere cause wavefront<br />
aberrations. Based on the work of Edlén [49] several contributions have been made<br />
to describe the dependence of the refractive index nair on temperature, pressure, humidity<br />
<strong>and</strong> CO2-concentration [19, 32, 108, 139, 141]. Many different formulations exist, which<br />
are often aimed at specific wavelength of interest. Because of the weak dependence on the<br />
relative humidity (for vertical propagation through the atmosphere) <strong>and</strong> CO2-concentration,<br />
these are often neglected [98]. The dependence of the refraction index on pressure <strong>and</strong><br />
temperature is given by [36]:<br />
nair = 1+7.76·10 −5P<br />
T<br />
<br />
1+ 7.52·10−3<br />
λ2 <br />
Where P is the pressure in millibars, T the temperature in K <strong>and</strong> λ the wavelength in<br />
microns. As a result of the change in the refractive index some parts of the initially flat<br />
wavefront are advanced <strong>and</strong> some parts of the wavefront are retarded.<br />
2.1.2 The Kolmogorov turbulence model<br />
The work of Kolmogorov in 1941 [112] formed the basis for currently used atmospheric<br />
turbulence models . Kolmogorov concluded that in a turbulent flow the kinetic energy is fed<br />
into the system at the outer scaleL0 <strong>and</strong> decreases till it is dissipated in heat at the smallest,<br />
inner scale l0. The outer scale corresponds to the radius of the largest air bubbles <strong>and</strong> the<br />
inner scale to that of the smallest. Outside the outer scale the isotropic behavior of the
2.1 Requirements 27<br />
Solar energy<br />
outer scale<br />
Wind shear<br />
convection<br />
inner scale<br />
Figure 2.1: Schematic of Kolmogorov turbulence.<br />
Energy is fed into the system at the outer<br />
scale <strong>and</strong> cascades till dissipated in heat at the<br />
inner scale.<br />
Figure 2.2: A typical profile measured with<br />
a SCIDAR instrument at Mt. Graham (profile<br />
taken from S.E.Egner [50]).<br />
atmosphere is violated <strong>and</strong> inside the inner scale viscous effects are dominant <strong>and</strong> kinetic<br />
energy is dissipated in heat. This is schematically shown in figure 2.1.<br />
Spatial model of atmospheric turbulence<br />
Kolmogorov described the r<strong>and</strong>om movement of the wind with statistical quantities by<br />
means of structure functions. Structure functions describe the mean squared difference between<br />
two r<strong>and</strong>omly fluctuating values. With the assumption that the atmosphere is locally<br />
homogeneous, isotropic <strong>and</strong> incompressible he concluded from a dimensional analysis that<br />
the kinetic energy decreases with the spatial frequency to the power− 5<br />
3 . Tatarski [173] related<br />
Kolmogorov’s velocity structure function to the index of refraction structure function<br />
Dn(h,r) given by:<br />
Dn(h,r) = 〈|n(h,r ′ )−n(h,r ′ +r)| 2 〉<br />
= C 2 2<br />
N (h)r 3, for l0 ≪ r ≪ L0<br />
where〈〉 denotes the variance of the enclosed expression at heighth<strong>and</strong> distancer. C2 N (h)<br />
is used to take into account the atmospheric turbulence contributions from all altitudes above<br />
the telescope. figure 2.2 gives a typical C2 N (h) profile. From this refractive index structure<br />
function profile it becomes clear that the ground layer <strong>and</strong> the high wind speed at the jet<br />
stream at about 10 km height strongly contribute to the wavefront aberrations.<br />
To quantify the effect of variations in index of refractions in terms of wavefront phase,<br />
another structure function is used: the phase structure functionDφ(r). For the values of the<br />
phase φ at any two points in the wavefront that are separated by a distance r this structure<br />
function is given by [98]:<br />
Dφ(r) = 〈|φ(r ′ ,t)−φ(r ′ +r,t)| 2 〉,<br />
2 ∞<br />
2π 1 5<br />
= 2.91 r 3<br />
λ cos(ζ) 0<br />
C 2 N(h)dh,<br />
2
2<br />
28 2 Design requirements <strong>and</strong> design concept<br />
5<br />
3 r<br />
= 6.88<br />
r0<br />
whereζ is the angle with zenith <strong>and</strong>r0 is the Fried parameter defined as:<br />
r0 =<br />
<br />
0.423<br />
2 2π 1<br />
λ cos(ζ)<br />
∞<br />
0<br />
C 2 N(h)dh<br />
− 3<br />
5<br />
.<br />
(2.1)<br />
The Fried parameter r0 is the characteristic spatial scale, which for λ=550nm typically<br />
ranges between 5 <strong>and</strong> 20cm [98]. The Fried parameter corresponds to the aperture diameter<br />
Dt of a telescope for which the varianceσ 2 wf of the wavefront aberrations is roughly 1 rad2 .<br />
This variance can be expressed as [136]:<br />
σ 2 wf<br />
= 1.03<br />
Dt<br />
r0<br />
5<br />
3<br />
(2.2)<br />
Other important statistics are described by the spatial Power Spectral Density (PSD),<br />
whichis a measure for the relative contribution of aberrations with spatial frequency<br />
κ = κ2 x +κ2y +κ2 z to the total wavefront distortion. For the Kolmogorov turbulence<br />
model this is given by [136]:<br />
Φ(κ,h) = 0.033C 2 N(h)κ −11<br />
3 ,<br />
Φ(κ) = 0.023r −5<br />
3<br />
0 κ−11 3 (assuming isotropy). (2.3)<br />
This spatial PSD is often truncated at the outer <strong>and</strong> inner scale of the turbulence in which<br />
the Kolmogorov model is valid. This is mostly done using the Von Karmann model:<br />
Φ(κ) = 0.023r−5 3<br />
0<br />
(κ2 +κ2 o) 11<br />
2 κ<br />
exp−<br />
6 κi<br />
(2.4)<br />
whereκo = 2π/L0 corresponds to the boundary set by the outer scaleL0 <strong>and</strong>κi = 5.92/l0<br />
corresponding to the lower boundary set by the inner scale l0. The outer boundary is in the<br />
order of tens of meters [34] <strong>and</strong> the inner scale is in the order of tens of millimeters [48, 118].<br />
The outer scale constrains the lower order wavefront distortions. Since these are dominant,<br />
the outer scale also determines the total stroke requirements for the actuators in adaptive<br />
<strong>mirror</strong>s. Knowledge of the outer scale at a certain telescope location for Extremely Large<br />
Telescopes (ELTs) is therefore of great importance. For intensity variations (scintillation)<br />
the inner scale is more relevant. In figure 2.3 the Kolmogorov PSD defined by (2.3) <strong>and</strong> the<br />
Von Karmann PSD defined by (2.4) is shown.<br />
Temporal model of atmospheric turbulence<br />
In analogy with the refractive index structure function a temporal structure functionDφ(δt)<br />
can be defined between two wavefront phase values separated in time byδt:<br />
Dφ(δt) = 〈|φ(r,t)−φ(r,t+δt)| 2 〉,
2.1 Requirements 29<br />
Φ [m −3 ]<br />
10 3<br />
10 1<br />
10 −1<br />
10 −3<br />
10 −5<br />
10 −7<br />
10 −9<br />
10 −11<br />
10 −2<br />
Outer scale Inner scale<br />
10 0<br />
Spatial frequency [m −1 ]<br />
Kolmogorov PSD<br />
Von Karmann PSD<br />
10 2<br />
Figure 2.3: The spatial PSDs of<br />
the wavefront aberrations for the Kolmogorov<br />
<strong>and</strong> Von Karmann turbulence<br />
models.<br />
where〈·〉 denotes the variance of the enclosed expression over space (r) <strong>and</strong> time (t). Under<br />
the assumption that the wavefront aberrations are fixed <strong>and</strong> turbulence layers at altitude h<br />
are moving with a wind speedv(h) over the telescope aperture – the frozen flow assumption<br />
– the temporal structure function can be expressed in the spatial frequencyκ as [33]:<br />
∞ 8 −3 1 κ<br />
Dφ(v,κ) ∝ C<br />
v(h) v(h)<br />
2 N (h)dh.<br />
0<br />
This function integrates the effect of all turbulence layers. When this integration is performed<br />
for a single turbulent layer at altitudehof thicknessδh traveling with a wind speed<br />
vw, the temporal power spectrumP of the phase valueφobserved at a certain point in space<br />
can be expressed in terms of the temporal frequencyf as:<br />
P(f,h) ∝ C 2 N (h)δh<br />
vw<br />
f<br />
vw<br />
−8/3<br />
.<br />
This −8/3 power law is often used in the context of <strong>control</strong>ler design for AO [98, 179],<br />
where integrator structures approximate the−8/3 power law by -2.<br />
In the previous paragraph, the characteristic spatial scale r0 was introduced to quantify<br />
the spatial variance of atmospheric turbulence. A similar value exists that describes the<br />
characteristic timescale for changes in wavefront aberrations [153]: the coherence time τ0.<br />
Various definitions exist [26, 179], but let it here be defined as the time for wind to carry<br />
frozen flow turbulence over an aperture of size r0. Based on the mentioned assumptions,<br />
this would imply that the wind speed is indicative of the coherence time τ0. This is in fact<br />
the case, even though the validity of the frozen flow assumption is questionable: it is e.g.<br />
shown in [40, 153] that the so-called boiling effect plays a major role in the evolution of<br />
phase errors on the timescales of practical interest. Let the coherence time τ0 be expressed<br />
through its inverse, the Greenwood frequencyfG [80]:<br />
fG = 1<br />
τ0<br />
= 2.31λ −6<br />
5<br />
<br />
1<br />
cos(ζ)<br />
∞<br />
0<br />
C 2 N(h)v 5<br />
3<br />
5<br />
3(h)dh .<br />
For a single turbulence layer with constant wind speedvw the Greenwood frequency can be<br />
approximated as:<br />
fG = 0.43 vw<br />
. (2.5)<br />
r0<br />
2
2<br />
30 2 Design requirements <strong>and</strong> design concept<br />
For representative values of the wind velocity vw = 10m/s <strong>and</strong> the Fried parameter r0 =<br />
0.166m the Greenwood frequency is approximately 25Hz. Since the Greenwood frequency<br />
is a measure for the rate of change of the wavefront distortion, it is related to the required<br />
<strong>control</strong> b<strong>and</strong>width of an AO system.<br />
2.2 Error budget<br />
The atmospheric conditions <strong>and</strong> the desired optical quality after correction are the main<br />
design drivers for the AO system. They determine the number of actuators <strong>and</strong> the <strong>control</strong><br />
b<strong>and</strong>width. The optical quality is often expressed by a Strehl-ratio S. This ratio can be<br />
related to the varianceσ 2 of the wavefront measured in radians using the extended Maréchal<br />
approximation [20, 98, 179]:<br />
S ≈ e −σ2<br />
This approximation is valid up to σ = 2rad [98]. For the design of the DM the practical<br />
limit to the diffraction limited level is set at a Strehl ratio of 0.85. This leads to a total error<br />
budget of σ ≈ 2π<br />
16 rad, which forλ=550nm corresponds to 550/16 ≈ 34nm. Assuming that<br />
all error sources are independent, the total variance can be approximated as the sum of the<br />
variances corresponding to the main contributing sources:<br />
σ 2 = σ 2 fit +σ 2 temp +σ 2 meas +σ 2 delay +σ 2 angle +σ 2 cal. (2.6)<br />
<br />
σ 2 ctrl<br />
The fitting errorσfit arises from the limited number Degrees Of Freedom (DOF) of the DM<br />
<strong>and</strong> thus the limited number of spatial frequencies that it can correct. The temporal error<br />
σtemp is due to the limited <strong>control</strong> b<strong>and</strong>width of the AO-system.<br />
If the light source used for wavefront sensing (i.e. the reference star), is not the same as<br />
the object for which the correction is used (the science object), a so called anisoplanatic<br />
error is made. The variance of this error is related to the angle θa by which the two objects<br />
are separated as σ2 angle ∝ θ5/3 a . Further, σ2 meas covers all the measurement errors<br />
(e.g. measurement noise in the wavefront sensor) <strong>and</strong> σ2 delay the errors due to delays in<br />
the wavefront sensor <strong>and</strong> the <strong>control</strong>ler. As will be discussed in section 2.4, a closed-loop<br />
<strong>control</strong>ler influences not only the temporal, but also the measurement <strong>and</strong> delay related errors,<br />
hence in (2.6) the combination of these sources is related to the <strong>control</strong>ler <strong>and</strong> denoted<br />
consists of all calibration errors. Calibration is needed for the correction<br />
σ2 ctrl . Finally,σ2 cal<br />
of static aberrations that are not seen by the wavefront sensor <strong>and</strong> are called Non Common<br />
Path Aberrations (NCPAs) [157]. A good review of the main errors in an AO system can<br />
be found in [98]. Since a large part of the total error budget is consumed byσ2 fit <strong>and</strong>σ2 temp<br />
which both can be influenced by the DM <strong>and</strong> <strong>control</strong>ler design, the other error sources will<br />
further be neglected. The fitting <strong>and</strong> temporal errors will be considered in the next two<br />
paragraphs to derive requirements for inter-actuator stroke <strong>and</strong> <strong>control</strong> b<strong>and</strong>width.
2.2 Error budget 31<br />
2.2.1 The fitting error<br />
The variance of the fitting error can be approximated by [98]:<br />
σ 2 fit = κf<br />
dt<br />
r0<br />
5<br />
3<br />
, (2.7)<br />
where dt is the inter actuator distance projected onto the primary aperture <strong>and</strong> r0 the Fried<br />
parameter. The fitting error coefficientκf depends on the type of <strong>mirror</strong> that is used:<br />
⎧<br />
⎪⎨ 1.26 for segmented <strong>mirror</strong>s with only piston correction,<br />
κf = 0.18 for segmented <strong>mirror</strong>s with tip, tilt <strong>and</strong> piston correction,<br />
⎪⎩<br />
0.28 for membrane <strong>mirror</strong>s.<br />
In [122] it is shown that for piston, continuous face-sheet <strong>and</strong> membrane <strong>mirror</strong>s for an<br />
equal number of actuator the correction quality does not significantly depend on the actuator<br />
geometry as long as the actuator distribution is fairly homogenous. This means that (2.7)<br />
gives an estimate for the fitting error variance to be expected for a specific DM on a telescope<br />
with known diameter at a site with a certainr0. Althoughκf given above is the smallest for<br />
segmented <strong>mirror</strong>s with tip, tilt <strong>and</strong> piston correction, this type of <strong>mirror</strong> has three actuators<br />
per segment whereas the inter actuator distance dt is assumed to be the segment size. For<br />
a more fair comparison, let the fitting error be expressed in terms of the total number of<br />
actuatorsNa, which can be achieved by writing the inter actuator spacingdt as a function of<br />
Na. For piston <strong>and</strong> membrane type <strong>mirror</strong>s, the inter actuator spacing can be approximated<br />
as dt ≈ Dt/2 π/Na, whereas for segmented <strong>mirror</strong>s with piston, tip <strong>and</strong> tilt correction<br />
the number of actuators must be scaled by three, yielding dt ≈ Dt/2 3π/Na. After<br />
substitution into (2.7),σ2 fit can thus alternatively be expressed as:<br />
σ 2 fit = κf,NaDt<br />
π/Na<br />
r0<br />
5/3<br />
where<br />
⎧<br />
⎪⎨ 0.63 for segmented <strong>mirror</strong>s with only piston correction,<br />
κf,Na = 0.23 for segmented <strong>mirror</strong>s with tip, tilt <strong>and</strong> piston correction,<br />
⎪⎩<br />
0.14 for membrane <strong>mirror</strong>s.<br />
This implies that for the same number of actuators, the fitting error is the smallest for a<br />
membrane type <strong>mirror</strong>.<br />
2.2.2 The temporal error<br />
Although in practice the temporal error depends on all components of the AO system as<br />
well as on actual atmospheric conditions, Greenwood [80] showed that the variance of the<br />
temporal error can be related to the Greenwood frequencyfG as:<br />
,<br />
σ 2 5<br />
3<br />
fG<br />
temp = k , (2.8)<br />
fc<br />
2
2<br />
32 2 Design requirements <strong>and</strong> design concept<br />
wherefc is the <strong>control</strong> b<strong>and</strong>width <strong>and</strong>k a scaling constant. For the ideal – though unrealistic<br />
– case that the <strong>control</strong>ler fully suppresses the wavefront disturbance up to the b<strong>and</strong>widthfc<br />
<strong>and</strong> does not affect higher frequencies, the scaling constantk is equal to 0.191. For a more<br />
realistic integrator type <strong>control</strong>ler it is equal to 1, which means that (2.8) gives an estimate<br />
for the temporal error to be expected for a given type of <strong>control</strong>ler <strong>and</strong> a given temporal<br />
behavior of the wavefront disturbance.<br />
However, the derivation of this relation is based on many assumptions. Starting point is a<br />
wavefront disturbance with a Kolmogorov spectrum <strong>and</strong> a frozen flow behavior, which is<br />
corrected by a DM system that is able to track a comm<strong>and</strong> signal up to the b<strong>and</strong>width fc.<br />
The temporal error is then defined as the servo tracking error of the DM with respect to<br />
the assumed type of wavefront disturbance. This means that the estimate of the temporal<br />
error variance in (2.8) does not take into account the ability of a (closed-loop) <strong>control</strong> law to<br />
reduce the detrimental effects of measurement noise or DM <strong>dynamics</strong>. It does not consider<br />
the <strong>dynamics</strong> of the wavefront sensor or delays in the disturbance signal to track. In AO<br />
literature, the latter is considered as a separate effect on the eventual performance <strong>and</strong> is<br />
quantified as the varianceσ2 delay between wavefronts measuredτ seconds apart [64]:<br />
σ 2 delay = 28.44(fGτ) 5<br />
3 .<br />
By considering <strong>control</strong> system delays as a separate source of errors, it is not taken into<br />
account that the <strong>control</strong> system can exploit spatio-temporal correlations of the wavefront<br />
distortion to make accurate short term predictions to compensate delays [100]. However,<br />
as the latter strongly depends on the atmospheric turbulence conditions, (2.8) will further<br />
be used for the estimation of the expected error. Since the wavefront sensor is regarded<br />
as a given part of the AO system <strong>and</strong> delays in a <strong>control</strong>ler affect its already considered<br />
b<strong>and</strong>width, the effect of delays will further be neglected as a separate source of errors.<br />
2.2.3 Error budget division<br />
If the main atmospheric parameters (r0 <strong>and</strong>fG) are known for a specific telescope location<br />
<strong>and</strong> only the fitting <strong>and</strong> temporal errors are considered, the actuator spacing dt <strong>and</strong> <strong>control</strong><br />
b<strong>and</strong>width fc can be related to a desired Strehl ratio (figure 2.4). When also the diameter<br />
Dt of a telescope is known, the number of actuatorsNa can be calculated in approximation<br />
as Na = π<br />
4 (Dt/dt) 2 . For an 8-meter telescope (Dt = 8m), figure 2.5 shows the Strehl<br />
ratios for the number of actuators Na <strong>and</strong> <strong>control</strong> b<strong>and</strong>width fc based on fG = 25Hz <strong>and</strong><br />
r0 = 0.166m (λ=550nm). Observe from figures 2.4 <strong>and</strong> 2.5 that the same Strehl ratio can<br />
be achieved by different combinations of <strong>control</strong> b<strong>and</strong>widthfc <strong>and</strong> number of actuatorsNa.<br />
According to figure 2.5 the effect of increasing the number of actuators is limited when<br />
this it not matched by an increase in <strong>control</strong> b<strong>and</strong>width <strong>and</strong> vice versa. A combination of<br />
actuator count <strong>and</strong> <strong>control</strong> b<strong>and</strong>width should be chosen for which the fitting <strong>and</strong> temporal<br />
errors are approximately equal. For a desired Strehl ratio of0.85 this leads to a combination<br />
of 5000 actuators <strong>and</strong> 200Hz <strong>control</strong> b<strong>and</strong>width, which is marked by a white star in figures<br />
2.4 <strong>and</strong> 2.5. The corresponding Root Mean Square (RMS) fitting <strong>and</strong> temporal errors are<br />
σfit = √ 0.28(Dt/ 4Na/π/r0) 5/6 = 0.34rad <strong>and</strong> σtemp = fG/fc = 0.17rad, which<br />
forλ =550nm corresponds to 30nm <strong>and</strong> 15nm respectively.
2.3 Actuator requirements 33<br />
Figure 2.4: The Strehl ratio as function of the<br />
relative actuator densitydt/r0 <strong>and</strong> <strong>control</strong> b<strong>and</strong>widthfG/fc.<br />
2.3 Actuator requirements<br />
Figure 2.5: The Strehl ratio as function of the<br />
number of actuators Na <strong>and</strong> <strong>control</strong> b<strong>and</strong>width<br />
fc based on Dt = 8m, r0 = 0.166m <strong>and</strong> fG =<br />
25Hz.<br />
Before stating the requirements for the actuators, it should be noted that the Optical Path<br />
Difference (OPD) of the light is the double of the facesheet displacement. This is explained<br />
in figure 2.6 <strong>and</strong> implies that the magnitude of the <strong>mirror</strong> deflection required to correct<br />
a wavefront needs to be half the magnitude of the wavefront unflatness. For the nearly<br />
diffraction limited correction of an 8 meter telescope in the visible part of the light spectrum<br />
the main requirements for the actuators are as follows:<br />
• Mirror diameter <strong>and</strong> actuator spacing<br />
Given a number of actuators Na, the actuator spacing depends on the diameter of<br />
the DM. The location of a DM is not restricted to a single position in the optical<br />
path of a telescope system. Because of the dynamic requirements <strong>and</strong> the ease of<br />
manufacturing, usually a flat surface with a smaller diameter is chosen. The lower<br />
limit is set by the Smith-Lagrange invariant. This optical invariant is explained in<br />
figure 2.7 <strong>and</strong> states that at all cross-sections in the optical path the product DΘ is<br />
constant. Herein is D the illuminated diameter or the envelope of all rays <strong>and</strong> Θ the<br />
angle between the optical axis <strong>and</strong> the chief (outer) ray of the beam. At the primary<br />
Mirror shape<br />
Incoming distorted wavefront<br />
Outgoing corrected wavefront<br />
Figure 2.6: The magnitude of the <strong>mirror</strong> deflection<br />
required to correct a wavefront needs only<br />
be half the magnitude of the wavefront unflatness<br />
because the deflection distance of the <strong>mirror</strong> is<br />
traveled twice. The summed lengths of all pairs<br />
of black <strong>and</strong> grey arrows are equal.<br />
2
2<br />
34 2 Design requirements <strong>and</strong> design concept<br />
Dt<br />
D ′ = Dt/10<br />
focus<br />
Θ Θ ′<br />
ΘDt = Θ ′ D ′ → Θ ′ = ΘDt/D ′ = 10Θ<br />
Figure 2.7: The optical Smith-Lagrange invariant, which states that at all cross-sections in the optical<br />
path the product DΘ is constant. D is herein the illuminated diameter, the envelope of<br />
all rays <strong>and</strong> Θ the angle between the optical axis <strong>and</strong> the chief (outer) ray. When D<br />
increases, Θ must decrease <strong>and</strong> vice versa.<br />
<strong>mirror</strong> of this telescope the invariant is equal to DtΘ, where Θ equals the Field Of<br />
View (FOV) <strong>and</strong> at apertures further along the optical path the angleΘ ′ will become<br />
Θ ′ = ΘDt/D ′ . To keep Θ ′ within realizable values for e.g. the 8m Very Large<br />
Telescope (VLT) with a half degree FOV <strong>and</strong> the 42m European Extremely Large<br />
Telescope (E-ELT) with 10 degree FOV, a realistic lower bound for the DM diameter<br />
lies in the order of 500mm. This will be chosen as a starting point in the design. With<br />
the 5000 actuators this defines a 500·10 −3 / 4·5000/π ≈ 6mm actuator spacing.<br />
• Total actuator stroke<br />
The total actuator stroke can be derived using (2.2) describing the RMS unflatness of<br />
the wavefront. With Dt = 8m <strong>and</strong> r0 = 0.166m (λ=550nm) this gives σ 2 stroke =<br />
σ 2 wf = 1.03(Dt/r0) 5/3 = 657rad 2 . The square root of this variance relates to the<br />
RMS actuator position, whereas in fact the Peak To Valley (PTV) value is sought<br />
that forms the total actuator stroke. For AO applications the RMS <strong>and</strong> PTV values<br />
are often related via a scaling factor 5, yielding a total actuator stroke of 5· √ 657 =<br />
128rad. Considering the reflection doubling the OPD (figure 2.6) forλ = 550nm, this<br />
corresponds to a required actuator stroke ofλ/2π ·128/2 = ±5.6µm. In addition to<br />
this stroke a fewµm are added to be able to deal with misalignment of the DM in the<br />
optical system.<br />
• Inter actuator stroke<br />
The inter actuator stroke can be calculated using the structure function in (2.1) describing<br />
the mean square difference between two wavefront phase values separated<br />
by a distance r. Substitution of r = dt <strong>and</strong> using r0 = 0.166m (λ=550nm) then<br />
yields the required mean square inter-actuator stroke as σ 2 ia = 6.88(dt/r0) 5/3 =<br />
2.8·10 −2 rad 2 . Using the factor 5 between the RMS <strong>and</strong> PTV strokes, the latter becomes<br />
5 · √ 2.8 = 8.3rad. Due to the reflection doubling the OPD (figure 2.6) for<br />
λ = 550nm this corresponds to a required inter-actuator stroke of±0.36µm.<br />
• Actuator resolution<br />
The actuator resolution should be well below the error budget as derived in section
2.4 Control system <strong>and</strong> electronics requirements 35<br />
2.2 of 2π<br />
16 rad (for λ = 550nm this is 34nm RMS). The design value for the actuator<br />
displacement resolution is therefore set significantly smaller at 5nm.<br />
• Power dissipation<br />
To avoid the need of active cooling, all energy dissipated in heat should be convected<br />
from the <strong>mirror</strong> surface by natural convection. The temperature difference between<br />
the <strong>mirror</strong> surface <strong>and</strong> the surrounding air of 1K is usually allowed. A typical value<br />
for the heat transfer coefficienthn is 1 < hn < 40W/m 2 [17]. Using hn = 12W/m 2<br />
<strong>and</strong> 5000 actuators on the ∅500mm DM, this allows for ≈ 0.5mW per actuator.<br />
Assuming that only half of the heat dissipated in the actuators is transferred to the<br />
<strong>mirror</strong> surface, the maximum heat dissipation per actuator is set to≈ 1mW.<br />
Dependence on telescope diameterDt<br />
Observe that according to (2.7) <strong>and</strong> (2.8) an increase of the telescope diameter Dt only<br />
affects the fitting error varianceσ 2 fit <strong>and</strong> not the temporal error varianceσ2 temp. To maintain<br />
the desired Strehl ratio, the actuator spacing dt must therefore remain constant <strong>and</strong> the<br />
number of actuators Na must increase with D 2 t . For the E-ELT this would result in 52 ·<br />
5000 = 625000 actuators.<br />
However, (2.7) <strong>and</strong> (2.8) do not consider any beneficial effects that a scale increase may have<br />
on the achievable <strong>control</strong>ler error σctrl. For instance, a larger number of correlated sensor<br />
inputs may lead to better short term predictions <strong>and</strong> a lower sensitivity to measurement<br />
noise. It is therefore likely that the actual number of actuators required at the E-ELT for the<br />
same Strehl ratio is smaller.<br />
2.4 Control system <strong>and</strong> electronics requirements<br />
The goal of the <strong>control</strong> system is to calculate suitable actuator comm<strong>and</strong>s based on wavefront<br />
sensor measurements. The AO system performance ultimately depends on the accuracy<br />
with which the <strong>control</strong> system can match the <strong>mirror</strong> shape to half that of the actual<br />
wavefront disturbance. In section 2.2.3 a desired <strong>control</strong> b<strong>and</strong>width of 200Hz was specified.<br />
To be able to achieve this using a PID (Proportional Integrator Derivative) type <strong>control</strong>ler,<br />
the delays due to sampling with a Zero Order Hold (ZOH), the sensor <strong>and</strong> the communication<br />
to the DM system must be limited to leave sufficient phase margin. Although this<br />
reasoning does not directly apply to Multi-Input Multi-Output (MIMO) systems in general,<br />
the foreseen DM system can with its first resonance frequency of around 1kHz be well<br />
diagonalized up to the b<strong>and</strong>width – i.e. decoupled into a number of Single-Input Single-<br />
Output (SISO) systems – by the inverse of the DM influence matrix. This is discussed in<br />
more detail in chapter 3.<br />
The phase margin PM is equal to 180 ◦ minus the total phase budget that can be divided<br />
into four parts: the plant, the <strong>control</strong>ler, the sensor <strong>and</strong> other delays in the loop. When the<br />
first resonance frequency of the DM lies above 1kHz, the phase delay of the plant around<br />
the 200Hz b<strong>and</strong>width will insignificant, but let the delay due to digital communication between<br />
the <strong>control</strong>ler <strong>and</strong> the DM system be budgeted to 1/10 th of a sample time. Further,<br />
the sampling with ZOH <strong>and</strong> the wavefront sensor both introduce half a sample delay <strong>and</strong> let<br />
the <strong>control</strong>ler computation delay be budgeted to one sample. The total loop delay is then<br />
2
2<br />
36 2 Design requirements <strong>and</strong> design concept<br />
Figure 2.8: Influence of sampling time Ts <strong>and</strong><br />
exposure time Te. The black line represents<br />
|S(2πjf)| 2 <strong>and</strong> the grey line the disturbance<br />
spectrum P(f) that has a horizontal asymptote<br />
on the measurement noise levelσn.<br />
Magnitude 2 [-]<br />
1<br />
− 8<br />
3<br />
σ 2 n<br />
Ts<br />
Te<br />
fc<br />
P<br />
Frequency [Hz]<br />
2.1 sample times, which means that the phase marginPM can be expressed as:<br />
PM = 180 ◦ −2.1 200Hz<br />
360 ◦ +γc,<br />
where fs is the sampling frequency <strong>and</strong> γc is the phase added by the <strong>control</strong>ler. For a PID<br />
<strong>control</strong>ler with a 2 nd order roll-off filter the latter can be at most 45 ◦ , whereas for a PI<br />
<strong>control</strong>ler with a 1 st order roll-off this reduces to −30 ◦ <strong>and</strong> for a pure integrator to −90 ◦ .<br />
This means that for a 10 ◦ phase margin PM the sampling frequency fs must be above<br />
700Hz for a PID type <strong>control</strong>ler, above 1kHz for a PI type <strong>control</strong>ler <strong>and</strong> above 2kHz for<br />
the pure integrator commonly used for AO.<br />
On the other h<strong>and</strong>, the sampling frequency cannot exceed the frame rate of the CCD<br />
camera of the wavefront sensor. For current devices 1kHz is a realistic rate <strong>and</strong> for state-ofthe-art<br />
devices this can even be slightly higher. However, as the exposure time is reduced the<br />
measurement noise becomes more significant. The variance σ2 n of the measurement noise<br />
of a Shack Hartmann sensor (SHS) consists of several components that are related to the<br />
exposure time Te in various ways. For instance, photon noise is attenuated by increasing<br />
Te, whereas dark current <strong>and</strong> read-out noise are attenuated by decreasing it [105, 106, 135].<br />
When measurement noise becomes significant for Te ≤ Ts <strong>and</strong> its variance is added to the<br />
temporal error variance in (2.8), this sum may not strictly decrease with the sampling frequencyfs.<br />
This can be illustrated using figure 2.8, which sketches the disturbance spectrum<br />
P(f), white measurement noise with varianceσ2 n <strong>and</strong> a disturbance rejection characteristic<br />
or sensitivity function|S(s)| 2 of the <strong>control</strong> system, where s is the Laplace variable. Note<br />
that S(s) is thus the transfer function between disturbance <strong>and</strong> residual error <strong>and</strong> that the<br />
notationS(s) implicitly assumes the AO system to be Linear Time-Invariant (LTI).<br />
The external disturbance acting on the <strong>control</strong> loop does not only include the wavefront<br />
disturbance with temporal spectrum P(f), but also the measurement noise with variance<br />
σ2 n . In contrast to the servo tracking point of view that forms the basis for the temporal<br />
error variance discussed in section 2.2.2, a realistic disturbance suppression characteristic<br />
|S(2πjf)| is here used that includes the effect of loop delays <strong>and</strong> obeys the Bode-sensitivity<br />
integral. This integral implies that the disturbance rejection at low frequencies leads to amplification<br />
at higher frequencies.<br />
When the filter S(s) is applied to a disturbance signal with temporal spectrumP(f)+σ 2 n ,<br />
fs<br />
|S| 2
2.4 Control system <strong>and</strong> electronics requirements 37<br />
the output (i.e. the residual error) spectrum can be expressed as |S(2πjf)| 2 (P(f)+σ 2 n).<br />
Using Parseval’s theorem, the <strong>control</strong> error variance σ2 ctrl introduced in (2.6) on page 30<br />
can then be expressed as:<br />
σ 2 ctrl =<br />
<br />
0<br />
∞<br />
|S(2πjf)| 2 P(f)+σ 2 n df.<br />
Now let this be applied to figure 2.8. Accordingly, a decrease of the sampling time Ts<br />
may lead to an increased b<strong>and</strong>width fc, but also to an increased disturbance amplification<br />
at high frequencies <strong>and</strong> since 0 < Te ≤ Ts also to higher measurement noise σn. As a<br />
result, the error variance σctrl may not diminish by a decrease of Ts. The same error may<br />
be achieved using various choices forTs <strong>and</strong>Te.<br />
As mentioned above, the Te <strong>and</strong> Ts form <strong>control</strong> loop delays, as do communication delays<br />
<strong>and</strong> computation time. In contrast to the exposure time, a reduction of the communication<br />
delays or computation time will always be beneficial to performance. However,<br />
communication speeds have limits <strong>and</strong> more computational power will result in limited<br />
performance gain at significant costs. A detailed specification of Te <strong>and</strong> Ts is complicated<br />
by the fact that the measurement noise σn <strong>and</strong> read-out time are highly device specific,<br />
whereas the WaveFront Sensor (WFS) is not included in the AO system to be designed.<br />
Moreover, model based <strong>control</strong>ler designs are able to predict (to some extent) future<br />
wavefront disturbances <strong>and</strong> so compensate for loop delays. In an optimal <strong>control</strong>ler design<br />
the effects of measurement noise are also minimized with respect to some cost function.<br />
The noise residual then becomes dependent also on the DM <strong>dynamics</strong> <strong>and</strong> the accuracies of<br />
the models used.<br />
Due to these a priori unknowns, the temporal error variance in (2.8) will further be used to<br />
express the worst case error. An indicative sampling time of Ts = 1ms will be assumed,<br />
equal to the exposure timeTe.<br />
As discussed in [65, 66], the best performance is obtained when the number of measurement<br />
positions of the wavefront sensor is proportional to the number of actuators. In the<br />
supplied references the location of the actuators with respect to the wavefront sensors is<br />
not explicitly analyzed, whereas it is known from <strong>control</strong> theory that performance may<br />
degrade when actuators <strong>and</strong> sensors are not collocated. On the other h<strong>and</strong>, the gradient<br />
measurement concept of a SHS versus the deflection based DM actuation already clouds<br />
the notion of collocation.<br />
Nevertheless, for a SHS with two measurements per lenselet, it will be assumed that the<br />
number of measurements Ns is approximately equal to twice the number of actuators<br />
Na. The total processing power of the <strong>control</strong> system must then be sufficient to evaluate<br />
the comm<strong>and</strong> update equations from ∼ 10000 measurements to 5000 comm<strong>and</strong> signals<br />
within the sampling time Ts of 1ms. This also involves the processing of the CCD of the<br />
SHS image to obtain the actual gradient measurements [106, 175]. This will be further<br />
elaborated in chapter 3.<br />
The final part of the <strong>control</strong> system is formed by the electronics hardware. The displacement<br />
of each actuator is changed by a current through the actuator coil, which will be<br />
generated by dedicated electronics that must have sufficient accuracy to meet the specified<br />
5nm actuator position accuracy. Further, the <strong>dynamics</strong> introduced by these driver electron-<br />
2
2<br />
38 2 Design requirements <strong>and</strong> design concept<br />
Deformable facesheet<br />
<strong>mirror</strong> - actuator connection<br />
Actuator grid<br />
Actuator grid - base frame connection<br />
Deformable facesheet<br />
Actuatorgrid with actuator modules<br />
Base frame<br />
Base frame<br />
Figure 2.9: Schematic of the adaptive <strong>deformable</strong> <strong>mirror</strong> design.<br />
ics must not reduce the lowest eigenfrequency or rise time of the system that may lead to a<br />
lower achievable <strong>control</strong> b<strong>and</strong>width.<br />
2.5 The design concept<br />
The design concept for the adaptive <strong>deformable</strong> <strong>mirror</strong> that meets the requirements, as listed<br />
in the previous sections, is schematically given in figure 2.9. The design concept is based<br />
on [81]. In the design a few layers are distinguished, which will be discussed in more detail:<br />
• the <strong>mirror</strong> facesheet,<br />
• the actuator grid,<br />
• the base frame.<br />
The first layer consists of the thin reflective facesheet, which is the <strong>deformable</strong> element. The<br />
facesheet is continuous <strong>and</strong> stretches out over the whole <strong>mirror</strong>. In the underlying layer - the<br />
actuator grid - low voltage electro-magnetic push-pull actuators are located. The actuator<br />
grid consists of a number of identical actuator modules. Each actuator is connected via<br />
a strut to the <strong>mirror</strong> facesheet. The <strong>mirror</strong> facesheet, the <strong>mirror</strong>-actuator connection <strong>and</strong><br />
the actuator modules form a thin structure with low out-of-plane stiffness so a third layer<br />
is added, the base frame, to provide a stable <strong>and</strong> stiff reference plane for the actuators.<br />
This base frame is a mechanically stable <strong>and</strong> thermally decoupled structure. Besides these<br />
distinguished layers a <strong>control</strong> system <strong>and</strong> electronics is present which is also described<br />
briefly.<br />
2.5.1 The <strong>mirror</strong> facesheet<br />
A membrane-type <strong>mirror</strong> is chosen because of its low moving mass <strong>and</strong> low out-of-plane<br />
stiffness. This results in low actuator forces <strong>and</strong> the best wavefront correction for a given
2.5 The design concept 39<br />
number of actuators (section 2.2.1). Because of the low out-of-plane stiffness of the thin<br />
facesheet, the inter-actuator coupling <strong>and</strong> the width of the influence functions can be kept<br />
small. This is desirable for currently available efficient <strong>control</strong> algorithms [53, 72, 188]<br />
<strong>and</strong> facilitates the implementation of a distributed <strong>control</strong> system (chapter 3). Finally, the<br />
limited thickness leads to a short thermal time constant that allows for quick adaptation to<br />
changing environmental temperatures.<br />
The <strong>mirror</strong>-actuator connection<br />
The connection between the actuators <strong>and</strong> the <strong>mirror</strong> facesheet is made by struts. Via these<br />
struts, the actuators impose the out-of-plane displacements on the facesheet. The struts<br />
constrain one DOF since their bending stiffness is significantly lower than the local bending<br />
stiffness of the facesheet. Because the struts leave the φ <strong>and</strong> ϕ rotation free, the bending<br />
stiffness of the facesheet can form a smooth surface through the imposed z-positions, as is<br />
shown schematically in figure 2.11. The piston-effect, shown schematically in figure 2.10 is<br />
thereby avoided. As a result no higher order aberrations are introduced into the wavefront.<br />
The struts neither constrain the x- <strong>and</strong> y-positions of the <strong>mirror</strong> facesheet. Differences<br />
in the thermal expansion coefficient of the facesheet material <strong>and</strong> the actuators <strong>and</strong>/or a<br />
temperature difference between them is therefore possible without unwanted deformation<br />
of the <strong>mirror</strong> surface. The struts are glued with small droplets to both the <strong>mirror</strong> facesheet<br />
as the actuators.<br />
The <strong>mirror</strong> facesheet <strong>and</strong> <strong>mirror</strong>-actuator connection is discussed in detail in [82–85]<br />
<strong>and</strong> in [174].<br />
2.5.2 The actuator modules<br />
Because a large number of actuators is needed it is attractive to produce actuator arrays<br />
with a layer based construction instead of single actuators, where each of them is positioned<br />
with respect to its neighbors. Therefore, a st<strong>and</strong>ard actuator module with 61 low voltage<br />
electromagnetic actuators in hexagonal arrangement is designed. More than 80 of such<br />
actuator modules are needed for a DM with 5000 actuators. figure 2.12 shows the <strong>mirror</strong><br />
facesheet with the <strong>mirror</strong>-actuator connections <strong>and</strong> the working principle of the actuators<br />
Actuator side<br />
Actuator - facesheet connections with high bending stiffness<br />
Figure 2.10: A connection, between the actuators<br />
<strong>and</strong> the facesheet, with high bending stiffness<br />
constrains the localφ- <strong>and</strong>ϕ-rotation in the<br />
<strong>mirror</strong> surface, causing local flattening.<br />
Actuator side<br />
Actuator - facesheet connections with low bending stiffness<br />
Figure 2.11: A connection, between the actuators<br />
<strong>and</strong> the facesheet, with low bending stiffness<br />
results in a smooth surface <strong>and</strong> allows lateral expansion<br />
between the facesheet <strong>and</strong> the actuators.<br />
2
2<br />
40 2 Design requirements <strong>and</strong> design concept<br />
in schematic. The actuators are of the variable reluctance type <strong>and</strong> consist of a closed<br />
magnetic circuit in which a Permanent Magnet (PM) provides static magnetic force on a<br />
ferromagnetic core that is suspended in a membrane. This attraction force is influenced by<br />
a current through a coil, which is situated around the PM to provide movement of the core.<br />
figure 2.12 shows that with the direction of the current the attractive force of the PM is either<br />
increased or decreased, allowing movement in both directions.<br />
The efficient actuators are free from mechanical hysteresis, friction <strong>and</strong> play <strong>and</strong> therefore<br />
have a high positioning resolution with high reproducibility. The stiffness of the actuator<br />
is determined by the membrane suspension <strong>and</strong> the magnetic circuit. There exists a large<br />
design freedom for both. The stiffness of the actuators is chosen such that, if one should<br />
fail, no hard point will form in the <strong>mirror</strong> surface.<br />
The coil wires are soldered to a flex foil. This flex foil is connected to a Printed Circuit<br />
Board (PCB) with dedicated electronics for 61 actuators. Each actuator module is connected<br />
to the base frame via three A-frames . The actuator grid is scalable, so the actuator pitch<br />
can be chosen freely <strong>and</strong> extendable since many modules can form large grids of actuators.<br />
The actuator module design is described in [88–91] <strong>and</strong> in chapter 5 <strong>and</strong> in more detail in<br />
[174].<br />
2.5.3 The <strong>control</strong> system <strong>and</strong> electronics<br />
To make the <strong>control</strong> system <strong>and</strong> the electronics of the AO system extendable without a full<br />
redesign, a <strong>modular</strong> structure is foreseen for these components. Each grid of 61 actuators<br />
is given a dedicated electronics module to supply each actuator with its current. These<br />
modules include 61 Pulse Width Modulation (PWM) drivers implemented using Field Programmable<br />
Gate Arrays (FPGAs) <strong>and</strong> 61 2 nd order analog low-pass filters. For the DM<br />
prototypes, the FPGAs receive their setpoint updates from a PC via a custom designed,<br />
multi-drop Low Voltage Differential Signalling (LVDS) communication link (chapter 6). A<br />
distributed <strong>control</strong> system is to be implemented in the FPGAs of the electronics modules<br />
chapter 3 <strong>and</strong> chapter 4. These modules communicate with a limited number of neighboring<br />
modules instead of the complete set, corresponding to a centralized <strong>control</strong>ler.<br />
This architecture has consequences for the <strong>control</strong>ler design, which is discussed in detail in<br />
permanent magnet<br />
reflective <strong>deformable</strong> facesheet<br />
membrane suspension<br />
strut<br />
N<br />
S<br />
coil<br />
ferromagnetic core<br />
N<br />
S<br />
base plate<br />
PM PM - Coil PM + Coil<br />
Figure 2.12: Three actuators shown in schematic.<br />
N<br />
S<br />
undeformed<br />
deformed
2.5 The design concept 41<br />
chapters 3 <strong>and</strong> 4.<br />
2.5.4 The base frame<br />
To support the 80 actuator modules, a light <strong>and</strong> stiff <strong>and</strong> thermally stable base frame has<br />
been designed. The diameter of this support structure is 500mm, its height 150mm <strong>and</strong><br />
its mass 5kg with a first mechanical resonance frequency of 1kHz. This base frame is a<br />
welded hexagonal box with ribs made of 2mm thick aluminium plates. The cover of the box<br />
is a 25mm thick aluminium honeycomb plate, which supports the actuator modules. Aluminium<br />
is chosen because of its good thermal properties. Since the box is well ventilated,<br />
it will adapt quickly to changes in the ambient temperature <strong>and</strong> hereby exp<strong>and</strong> <strong>and</strong> contract<br />
homogeneously. Finally, the box can contain the electronics for the <strong>mirror</strong> modules.<br />
2
ÔØÖØÖ<br />
ÒØ ÓÒØÖÓÐÓÖÇ<br />
ÓÒÔØ×Ò ÐÐÒ×<br />
As the number of degrees of freedom of the <strong>deformable</strong> <strong>mirror</strong> <strong>and</strong> the wavefront<br />
sensor increase, scaling problems arise for the number of computations<br />
required by the <strong>control</strong> system. Without efficient algorithms, the dem<strong>and</strong>ed<br />
computational power increases quadratically with the number of degrees of<br />
freedom <strong>and</strong> thus to the fourth power in the telescope’s aperture diameter. Further,<br />
practical problems arise for cabling <strong>and</strong> driver electronics for the actuators.<br />
A <strong>modular</strong>, distributed <strong>control</strong> system is proposed to solve these problems,<br />
but this approach raises new challenges that will be discussed <strong>and</strong> possible<br />
solutions will be proposed.<br />
43
3<br />
44 3 Efficient <strong>control</strong> for AO: concepts <strong>and</strong> challenges<br />
3.1 Introduction<br />
In this chapter, the problems that will arise for the <strong>control</strong> system when the number of degrees<br />
of freedom of the <strong>Adaptive</strong> Optics (AO) system increases will be discussed. Therefore,<br />
first the general <strong>control</strong> problem will be formulated <strong>and</strong> several commonly used approaches<br />
for <strong>control</strong>ler design will be discussed.<br />
It will be shown in section 3.3 that the computational dem<strong>and</strong> of these approaches scales<br />
with the fourth power in the diameter of the telescope’s aperture. This is not acceptable<br />
for future large telescopes <strong>and</strong> has lead to lots of research into efficient algorithms (e.g.<br />
[53, 72, 124]) of which a few will be discussed in more detail. However, these algorithms<br />
are tailored for a specific <strong>control</strong> law, whereas recent reports indicate that a significant performance<br />
increase can be achieved using designs that are both temporally <strong>and</strong> spatially<br />
optimal (e.g. using the H2 framework [100, 123]). This is mainly due to the incorporation<br />
of more detailed models of the atmospheric disturbance. Therefore, in this thesis such models<br />
will form an important part of the <strong>control</strong>ler design.<br />
Besides computational dem<strong>and</strong>, other more practical issues arise when the AO system is<br />
extended to more actuators <strong>and</strong> sensors: design of the electronics, wiring, communication,<br />
etc. It will be discussed how these issues are simplified if a distributed structure is chosen<br />
for the <strong>control</strong> system that will be elaborated in section 3.4. However, such a structure has<br />
major consequences for the <strong>control</strong>ler design. These consequences <strong>and</strong> problems will be<br />
discussed <strong>and</strong> several ideas will be presented that may offer (partial) solutions. These ideas<br />
also form the basis for the approaches described in the following chapters of this thesis.<br />
3.2 Existing <strong>control</strong> methods<br />
In this section the general AO <strong>control</strong> problem will be stated <strong>and</strong> followed by several approaches<br />
towards a solution. Subsequently, these methods will be analyzed with respect to<br />
their computational dem<strong>and</strong>s as the diameter of the telescope for which the AO system is<br />
designed increases.<br />
3.2.1 Generic problem statement<br />
The aim of the AO system is to compensate the local variations in the wavefront of<br />
the incoming light that are caused by atmospheric turbulence. The observed wavefront<br />
distortion can be influenced with a wavefront corrector such as the <strong>deformable</strong> <strong>mirror</strong><br />
designed within this project. As mentioned in chapter 2, the performance of an AO system<br />
can be measured using the Strehl ratio. According to the Maréchal approximation, this ratio<br />
is related to the variance of the wavefront distortion. It is assumed that the wavefront is flat<br />
before it enters the earth’s atmosphere such that the general aim for the AO <strong>control</strong> system<br />
is to minimize the wavefront variance. Note here that the image quality is not affected by<br />
the global average of the wavefront (the piston mode). For <strong>control</strong>ler design, the wavefront<br />
entering the earth’s atmosphere is therefore defined to be zero.<br />
As mentioned in chapter 2, the <strong>control</strong> system will have a closed-loop configuration,<br />
which means that the sensors register the effects of the <strong>control</strong> actions. The <strong>control</strong> loop
3.2 Existing <strong>control</strong> methods 45<br />
φ<br />
φ u u y<br />
H<br />
C<br />
n<br />
G<br />
φ e<br />
Figure 3.1: Schematic of a<br />
generic AO system in closedloop<br />
configuration.<br />
can be described in general terms using operators to represent the dynamic systems <strong>and</strong><br />
functions of timetto represent the signals to avoid representation or implementation details.<br />
Let the wavefront phase disturbance of the light entering the telescope at time instant t be<br />
denoted by the signalφ(t). This is influenced by the wavefront corrector that introduces an<br />
additional wavefrontφ u(t) that is related to the actuator comm<strong>and</strong>su(t) as:<br />
φ u (t) = H{u(t)}, (3.1)<br />
where the operator H represents the wavefront corrector. The combined wavefront error<br />
φ e(t) that forms the input to the wavefront sensor then becomes:<br />
φ e(t) = φ(t)+φ u(t). (3.2)<br />
Let the wavefront sensor be represented by the operator G <strong>and</strong> its measurements y(t) be<br />
subject to additive measurement noisen(t) such that it can be expressed as:<br />
y(t) = G{φ e(t)}+n(t). (3.3)<br />
The measurementy(t) is fed back to the <strong>control</strong>lerC with outputu(t). With a slight misuse<br />
of notation, let 〈φ e (t)〉 denote the average or piston term of the wavefront that does not<br />
affect the quality of the resulting image. The <strong>control</strong>ler must then minimize the piston-free<br />
wavefront errorφ e(t)−〈φ t(t)〉 in a minimum variance sense:<br />
C = argmin<br />
C<br />
〈φ e(t)−〈φ e(t)〉2〉, (3.4)<br />
subject to (3.1), (3.2) <strong>and</strong> (3.3). This <strong>control</strong> loop is shown schematically in figure 3.1. This<br />
<strong>control</strong>ler synthesis problem can be approached in various ways. Each one mainly differs<br />
in the choices for the models of the wavefront corrector systemH<strong>and</strong> the wavefront sensor<br />
G <strong>and</strong> by the assumptions on the disturbance signalsφ(t) <strong>and</strong>n(t). In the next subsections,<br />
these choices <strong>and</strong> assumptions will be elaborated on for several important design approaches<br />
described in literature.<br />
3.2.2 Traditional approach<br />
Traditionally, the plant P to be <strong>control</strong>led is taken as the combination of the wavefront<br />
corrector system H <strong>and</strong> the wavefront sensor G [55]. The <strong>control</strong> loop elements are considered<br />
to be discrete time systems, expressed using the dimensionless time variable t <strong>and</strong><br />
z-transform variablez. The plant P is described as a static gain matrixP with an arbitrary<br />
numberd ∈ N of samples delay, s.t. the measurement can be expressed as:<br />
y(t) = Gφ(t)+Pu(t−d)+n(t).<br />
3
3<br />
46 3 Efficient <strong>control</strong> for AO: concepts <strong>and</strong> challenges<br />
Figure 3.2: Schematic of a<br />
generic AO system in openloop<br />
configuration with <strong>control</strong>lerR.<br />
φ<br />
G<br />
The <strong>control</strong>lerC is given a predefined discrete time structure that consists of integrators with<br />
gainα<strong>and</strong> leak factorβ <strong>and</strong> an inverse plant model described by a static matrixP # :<br />
n<br />
s<br />
R<br />
u<br />
H<br />
C(z) : u(t) = −α(I−βz −1 I) −1 P # y(t), (3.5)<br />
whereTs the sampling time. Although this equation suggests that the gainsα <strong>and</strong>β must be<br />
equal for all integrators, this is not necessarily the case. Due to the wavefront sensing principle,<br />
the plant matrix P is usually singular <strong>and</strong> its inverse does not exist. In practice, the<br />
inverse model P # is therefore obtained as a pseudo-inverse to circumvent this. Nevertheless,<br />
theP # term in the <strong>control</strong>ler will decouple the plant <strong>dynamics</strong> up to the unseen modes<br />
of the WaveFront Sensor (WFS) <strong>and</strong> allows separate Single-Input Single-Output (SISO)<br />
<strong>control</strong>lers – in this case integrators – to be designed for each actuator loop. Integrators provide<br />
a high gain at low-frequencies <strong>and</strong> thus a high disturbance suppression. The gain <strong>and</strong><br />
leak factor α <strong>and</strong> β can be tuned for maximum <strong>control</strong> b<strong>and</strong>width – <strong>and</strong> thus disturbance<br />
rejection performance – while satisfying certain stability margins.<br />
This <strong>control</strong>ler structure is commonly used in practice <strong>and</strong> provides respectable performance.<br />
Partly this is because the static plant assumption is often not far from the truth,<br />
but also because an integrator has the right magnitude profile over frequency to suppress<br />
the atmospheric wavefront disturbance. This can be explained in a SISO setting by first<br />
recalling that for stochastic disturbance rejection the residual error of the optimal <strong>control</strong>ler<br />
is white noise. This means that the product of the disturbance model <strong>and</strong> the sensitivity<br />
function of the <strong>control</strong> loop has a magnitude that is constant over frequency. When the<br />
<strong>control</strong> law is an integrator <strong>and</strong> the Deformable Mirror (DM) a static gain, the sensitivity<br />
function has a +1 asymptote for frequencies up to the <strong>control</strong> b<strong>and</strong>width, which depends on<br />
the integrator gain. As elaborated in chapter 2, the atmospheric turbulence is often modeled<br />
by assuming a Kolmogorov [98] spatial spectrum <strong>and</strong> the Taylor hypothesis (frozen flow)<br />
to provide the temporal spectrum. According to these assumptions, the temporal power<br />
spectrumΦ(f) is related to the temporal frequencyf as:<br />
Φ(f) ∝ f −8 3.<br />
This −8/3 exponent is close to −2, which means that for frequencies up to the <strong>control</strong><br />
b<strong>and</strong>width the residual error spectrum described by the product Φ(f)|S(f)| 2 becomes<br />
approximately constant (’white’).<br />
Nevertheless, an integrator only compensates for the temporal behavior of the atmospheric<br />
wavefront disturbance. A spatial model is usually included by designing the inverse<br />
plant modelP # – often referred to as the reconstructor – as the open-loop, minimum vari-<br />
ance reconstructorR:<br />
<br />
ˆR<br />
2<br />
= argmin φ(t)−HRs(t)F , (3.6)<br />
R<br />
φ u<br />
φ e
3.2 Existing <strong>control</strong> methods 47<br />
wheres(t) is the open-loop measurement defined as (figure 3.2):<br />
s(t) = Gφ(t)+n(t) (3.7)<br />
Note that in literature additional regularization <strong>and</strong> weighting matrices are applied for specific<br />
cases that will not be considered here. This optimal open-loop reconstructorR replaces<br />
P # in the closed-loop <strong>control</strong> law in (3.5), which means that the effect of closing the <strong>control</strong><br />
loop on the spatial spectrum of the disturbance is not accounted for. The pseudo open-loop<br />
<strong>control</strong> concept presented in [54, 187] compensates for this using an approach based on the<br />
internal model <strong>control</strong> principle (section 4.4.2).<br />
To obtain ˆ R, it is assumed that the behavior of both the wavefront corrector <strong>and</strong> the wavefront<br />
sensor can be described by the static matrices H <strong>and</strong> G respectively. Then rewrite<br />
(3.6) in terms of the vectorialr = vec(R) while omitting references to timetfor brevity:<br />
φ−(s ˆr = argmin<br />
r<br />
T ⊗H)r 2 <br />
,<br />
F<br />
<br />
= argmin φ<br />
r<br />
T φ−2φ T (s T ⊗H)r+r T (ss T ⊗H T <br />
H)r ,<br />
<br />
= argmin φ<br />
r<br />
T <br />
φ −2vec T H T φs T r+r T ss T ⊗H T H r,<br />
<br />
= vec (H T H) −1 H T φs T <br />
T−1<br />
ss , (3.8)<br />
<br />
ˆR<br />
where⊗denotes the Kronecker product [23] <strong>and</strong> use is made of the identity vec(ADB) =<br />
(BT ⊗ A)vec(D). The last step follows from a completion of squares argument. By regarding<br />
the plant as two separate processes instead of only their product, statistical knowledge<br />
of the wavefront disturbance can thus be exploited. By using the sensor model<br />
s(t) = Gφ(t−d)+n(t), the covariance matrices can be rewritten to:<br />
T<br />
φs <br />
= φ(t−d)φ T (t−d)G T<br />
<br />
= φφ T<br />
G T ,<br />
T<br />
ss = G φ(t−d)φ(t−d) T G T + n(t)n T (t) <br />
= G φφ T<br />
G T + nn T .<br />
The covariance matrix<br />
<br />
φφ T<br />
can be calculated from the Kolmogorov or Von Karman<br />
spatial spectra [95] <strong>and</strong> nnT is often taken as a diagonal matrix whose elements are<br />
determined from e.g. sensor <strong>and</strong>/or guide star properties.<br />
It is important to notice that the optimal static reconstructor in (3.8) consists of two parts:<br />
T<br />
φs ss T−1<br />
ˆR = H T H −1 H T<br />
<br />
part 2<br />
<br />
part 1<br />
. (3.9)<br />
Part one is the reconstruction of the wavefront phase from the WFS measurements <strong>and</strong> part<br />
two is the projection of the wavefront onto the actuator space. These form the solutions to<br />
two sequential least-squares problems formulated as:<br />
ˆφ s(t) = arg min φ(t)−φ s(t)<br />
φ (t) s 2 F, (part 1) (3.10)<br />
û(t) = argmin<br />
u(t) ˆ φ s(t)−Hu(t) 2 F, (part 2)<br />
3
3<br />
48 3 Efficient <strong>control</strong> for AO: concepts <strong>and</strong> challenges<br />
where ˆ φ s(t) is the minimum variance reconstruction of the actuator wavefront phase disturbanceφ(t)<br />
given the open-loop WFS measurement vectors(t).<br />
Using the internal model <strong>control</strong> principle (section 4.4.2), the equivalent closed-loop <strong>control</strong>ler<br />
can be derived from the obtained open-loop <strong>control</strong>ler ˆ R as:<br />
C = C(z) = −(I− ˆ RGHz −1 H) −1ˆ R. (3.11)<br />
When ˆ <br />
R is based on the covariance matrices φφ T<br />
= I <strong>and</strong> nnT = 0 then the<br />
equivalent closed-loop <strong>control</strong>ler reduces to the traditional integrator structure in 3.5 with<br />
α = β = 1 . In practice the optimal open-loop reconstructor from 3.9 is often used to<br />
replace theP # matrix in 3.11<br />
However, as mentioned this method is only optimal under several specific conditions<br />
[123]. Besides the DM being static, the spatial <strong>and</strong> temporal parts of the wavefront disturbance<br />
behavior should be independent. Moreover, the temporal behavior of the latter should<br />
be well described by a r<strong>and</strong>om walk (integrated white noise) process. Since none of these<br />
conditions are in general fully met, recent work has been aimed at finding more accurate<br />
plant <strong>and</strong> disturbance models to be applied within e.g. theH2 optimal <strong>control</strong> framework.<br />
3.2.3 More generic <strong>control</strong>ler designs<br />
The H2 optimal <strong>control</strong> framework allows the direct synthesis of an Linear Time-Invariant<br />
(LTI), dynamic, closed-loop <strong>control</strong>ler C = C(z) based on dynamic models for the components<br />
in the <strong>control</strong> loop. These models comprehend the turbulence, the corrector <strong>and</strong><br />
the sensor <strong>and</strong> together form the so called generalized plant. Results have been reported by<br />
[16, 100, 123, 150] showing a significant improvement in optical quality when compared to<br />
that achieved with the traditional <strong>control</strong>ler design. Moreover, this advantage is shown to<br />
increase for higher Greenwood frequencies [100].<br />
Where [123] only suggests to use a dynamic disturbance model, in [16, 150] a physical<br />
model is actually used that is based on Kolmogorov statistics <strong>and</strong> Zernike polynomials [136]<br />
in combination with Taylor’s frozen flow assumption <strong>and</strong> a wavefront corrector modeled as<br />
a static matrix gain with a delay. In [100] a data driven approach is proposed in which an<br />
LTI disturbance model is estimated from measurement data alone <strong>and</strong> the wavefront corrector<br />
is modeled as a two-tap Finite Impulse Response (FIR) filter.<br />
Despite their differences, all these approaches yield closed-loop <strong>control</strong>lers C = C(z) of<br />
the form:<br />
C(z) :<br />
<br />
x(t+1)<br />
=<br />
u(t)<br />
Ac Kc<br />
Cc Dc<br />
<br />
x(t)<br />
, (3.12)<br />
y(t)<br />
where x(t) ∈ R Nx is the state vector whose length Nx is the sum of the lengths of the<br />
states of all models considered in the <strong>control</strong> loop.<br />
In practice the behavior of the atmospheric disturbance varies slowly over time, which<br />
means that the process that generatesφ(t) is not fully stationary. For best performance, the<br />
disturbance model should therefore also vary over time, which can be accomplished using<br />
adaptive <strong>control</strong> schemes [56, 117]. Such schemes are available to update the covariance<br />
matrices of the traditional <strong>control</strong> law <strong>and</strong> also to update the coefficients of more generic<br />
<strong>control</strong> laws [44].
3.3 Scaling problems 49<br />
3.3 Scaling problems<br />
As discussed in chapter 2, there should be a balance between the spatial b<strong>and</strong>width of the<br />
wavefront corrector that is determined by the number of actuators Na <strong>and</strong> the temporal<br />
b<strong>and</strong>width of the AO system that is also determined by the <strong>control</strong> system. The spatial b<strong>and</strong>width<br />
determines the fitting error whose variance is inversely proportional to Na, whereas<br />
the latter determines the temporal error. This depends on the disturbance behavior <strong>and</strong> is<br />
independent of Na. Thus, to keep the balance for a telescope with a larger aperture area,<br />
Na must be chosen linearly proportional to this area such that the actuator density remains<br />
constant. Similarly, the spatial density of the WFS grid must remain constant when the<br />
aperture area is enlarged. In this section, the consequences hereof will be discussed for both<br />
the hard- <strong>and</strong> software of the <strong>control</strong> system.<br />
3.3.1 Computational dem<strong>and</strong><br />
When the numbers of actuators Na <strong>and</strong> sensors Ns are linearly proportional to the telescope’s<br />
aperture area, the number of in- <strong>and</strong> outputs of the <strong>control</strong> system scale quadratically<br />
with the aperture diameterDt.<br />
At each sampling time, the <strong>control</strong> system has to perform two sequential tasks. Firstly, when<br />
assuming the WFS to be of the Shack-Hartmann type [98] – it has to process the Charge<br />
Coupled Device (CCD) images of the WFS to find the gradients. Subsequently, it has to<br />
update the setpoints of the actuators of the wavefront corrector according to a given <strong>control</strong><br />
law.<br />
Centroiding<br />
Many algorithms have been devised to determine the spatial gradients y(t) from the CCD<br />
image of a Shack-Hartmann sensor [27, 106, 142, 175]. However, the centroiding or center<br />
of mass method [98] is probably the best known. The number of computations of this<br />
algorithm is linearly proportional to the number of spots of the WFS. It can be split into<br />
parallel problems for each sensor spot <strong>and</strong> even be implemented in the CCD hardware.<br />
Therefore, this part of the computational dem<strong>and</strong> will not be further considered.<br />
The traditional <strong>control</strong> law<br />
The second task of the <strong>control</strong> system is to implement the <strong>control</strong> law. The <strong>control</strong>ler receives<br />
the measured gradients as input <strong>and</strong> has to calculate the new setpoints for the wavefront<br />
corrector. For the traditional <strong>and</strong> optimal static <strong>control</strong>ler structures, this consists of<br />
the application of the inverse model P # or the reconstruction matrix R <strong>and</strong> the temporal,<br />
SISO integrators.<br />
Even though the plant matrix P is for most corrector technologies very sparse – i.e. each<br />
actuator influences only the WFS readings within its vicinity – all elements of the (pseudo-)<br />
inverse P # are usually non-zero. The same holds for the optimal reconstruction matrix<br />
R in (3.9). Thus, if implemented as a direct matrix multiplication, the number of computations<br />
Nc required by the matrix-vector product between P # or R <strong>and</strong> the measurement<br />
vectory(t) is equal to the productNaNs, where Ns is the number of measurements of the<br />
3
3<br />
50 3 Efficient <strong>control</strong> for AO: concepts <strong>and</strong> challenges<br />
Figure 3.3: Sparsity of the matrix G T for a<br />
Fried geometry <strong>and</strong> a sensor grid of 15 × 15<br />
lenselets.<br />
Figure 3.4: Sparsity of the matrix G T G for a<br />
Fried geometry <strong>and</strong> a sensor grid of 15 × 15<br />
lenselets.<br />
WFS. Both Na <strong>and</strong> Ns are linearly proportional to the aperture area Aa of the telescope<br />
<strong>and</strong> quadratically in its diameterDt:<br />
Na ∝ Ns ∝ Aa ∝ D 2 t .<br />
The number of computations required for application of the integrator scales only linearly<br />
with Na <strong>and</strong> does not change the order of magnitude of the total number of computations.<br />
The number of computations Nc required for the traditional <strong>control</strong> law is thus related to<br />
the number of actuatorsNa <strong>and</strong> the telescope aperture diameterDt as (figure 3.5):<br />
Nc ∝ N 2 a ∝ D 4 t.<br />
This means that to increase the aperture diameter with a factor two, the number of processors<br />
would have to increase at least a factor 16, which is expensive <strong>and</strong> leads to many<br />
practical problems. Therefore, lots of research effort have been spent at reducing the computational<br />
dem<strong>and</strong> of the application ofP # orRas in (3.9).<br />
This can be done by exploiting structure <strong>and</strong> sparsity of the involved matrices [53, 73, 74,<br />
187–189]. As mentioned, for most corrector technologies, each actuator has only a local effect<br />
on the wavefront induced by the corrector, which renders the influence matrixHsparse.<br />
Further, as illustrated in figures 3.3 <strong>and</strong> 3.4 the geometry matrix G that relates the actual<br />
wavefront error to the WFS measurements is sparse for zonal relations such as Fried or<br />
Hudgin (chapter 1). Finally, when considering a Kolmogorov spatial spectrum, the inverse<br />
of the covariance matrix<br />
<br />
φφ T<br />
can be approximated using discrete approximations of the<br />
Laplacian operator [53, 187]. The approximation lies in the fact that the −11/3 exponent<br />
of the Kolmogorov spatial spectrum is changed into−4. Using sparse-plus-low-rank matrix<br />
techniques, this allows for an efficient implementation of the application of the reconstruction<br />
matrixRthat scales approximately asN 3/2<br />
s [53].
3.3 Scaling problems 51<br />
Computations per second<br />
10 16<br />
10 14<br />
10 12<br />
10 10<br />
10 8<br />
10 6<br />
10 1<br />
10 2<br />
Primary <strong>mirror</strong> diameter [m]<br />
0.4 0.6 1 2 4 6 10 20 40<br />
Classic integrator<br />
Optimal <strong>control</strong><br />
Distributed <strong>control</strong><br />
10 3<br />
Number of actuators<br />
10 4<br />
10 5<br />
10 6<br />
10 4<br />
10 2<br />
10 0<br />
Number of processors<br />
Figure 3.5: The estimated<br />
number of computations or<br />
processors required for the unoptimized<br />
implementation of<br />
various <strong>control</strong> algorithms including<br />
the centroiding step<br />
versus the telescope’s aperture<br />
diameterDt <strong>and</strong> the number<br />
of actuators Na of the<br />
wavefront corrector.<br />
The latter referenced method is aimed at obtaining an efficient algorithm to find the exact<br />
solution to the problem posed in (3.6) using sparse approximations of the matrices involved<br />
in the solution in (3.9). Another class of iterative algorithms is aimed at finding an approximation<br />
of this solution with a known accuracy. These methods are usually based on Krylov<br />
subspace methods [152], in particular on the conjugate gradient method [73, 74, 188, 189].<br />
These methods can a.o. be used to solve linear problems of the formAx = b for the vector<br />
x of unknowns <strong>and</strong> the known matrix A <strong>and</strong> vector b. The reconstructor in (3.9) can be<br />
written into this form. The solution vector x (m) at iteration m is updated each iteration in<br />
the directiond (m) with a step size α (m) as:<br />
x (m+1) = x (m) +α (m)d (m), where α (m) = rT (m) r (m)<br />
d T (m) Ad (m)<br />
(3.13)<br />
<strong>and</strong> r (m) = Ax (m) − b is the residual vector. The method requires the search directions<br />
d (m) to be A-orthogonal or conjugate, i.e. d T (m) Ad (n) = 0 for m = n. This makes sure<br />
that the exact solution is reached after Nm iterations, where Nm is the dimension of the<br />
square matrixA. The search directions are built from the residual vectors as:<br />
d (m+1) = r (m+1) +β (m+1)d (m), where β (m+1) = rT (m+1) r (m+1)<br />
rT (m) r . (3.14)<br />
(m)<br />
These choices for α (m) <strong>and</strong> β (m+1) imply that the conjugacy condition is satisfied.<br />
However, in practice this is prone to numerical round-off errors.<br />
The number of computations of the conjugate gradient method scales as Nalog(Na) times<br />
the number of iterations required for convergence. The latter number depends on the spread<br />
of the eigenvalues of A, but in general the computation time of this method is of the same<br />
order as that of the direct matrix product. It mainly offers a storage reduction as the usually<br />
dense inverse matrixA −1 does not need to be stored.<br />
To improve the convergence speed of the algorithm for the specific problem of<br />
wavefront reconstruction for AO, this method can be extended with a variety of precon-<br />
3
3<br />
52 3 Efficient <strong>control</strong> for AO: concepts <strong>and</strong> challenges<br />
ditioners [73, 188]. Using the preconditioner M, the system of equations is modified to<br />
M −1 Ax = M −1 b. The matrix M −1 should approximate A −1 <strong>and</strong> calculation of the<br />
productM −1 v wherev is an arbitrary vector should require few computations. This allows<br />
shaping of the eigenvalues ofM −1 A for better convergence properties. The preconditioner<br />
can be inserted into the conjugate gradient algorithm in a matrix-free way, i.e. neitherM −1<br />
itself nor the product M −1 A needs to be explicitly evaluated. At each iteration, only the<br />
product ofM −1 with the residual vectorr (m+1) needs to be calculated.<br />
In practice, the multi-grid [73, 189] <strong>and</strong> the Fourier domain preconditioners [188] yield the<br />
most efficient solutions. The first is much related to the hierarchic reconstructor described<br />
in [124]. It is also used for solving discretized Partial Differential Equation (PDE) problems<br />
such as Finite Element Model (FEM) problems <strong>and</strong> steady state diffusion equations <strong>and</strong><br />
exploits the mesh structure of a problem. The residual vector is resampled on different<br />
hierarchic levels of mesh density. A so-called V-cycle is then often performed, starting<br />
from the finest mesh, through all intermediate levels to the coarsest mesh <strong>and</strong> back again.<br />
Transition from one level to another involves interpolation <strong>and</strong> restriction operations<br />
that are alternated with smoothing operations. These smoothing operations can be one or<br />
several iterations of another – computationally cheap – iterative solver such as Gauss-Seidel<br />
or successive over-relaxation (SOR). This way, the wavefront reconstruction problem is<br />
solved on hierarchically alternating levels of spatial resolution. However, the computational<br />
efficiency depends highly on the specific choices for the interpolation, restriction <strong>and</strong><br />
smoothing operations. For only the wavefront reconstruction step, an order-N computational<br />
dem<strong>and</strong> – i.e. Nc ∝ Ns ∝ D 2 t – has been reported [72, 189]. It is further shown that<br />
bilinear <strong>and</strong> even bicubic DM influence functions can be used to to solve the second step of<br />
mapping the wavefront to actuator comm<strong>and</strong>s with the same method [189].<br />
Since the 2D Fourier transforms of the sensor geometry matrixG, the covariance matrix<br />
φφ T<br />
<strong>and</strong> the influence matrix H can often be well approximated by sparse – or even<br />
diagonal – matrices, a Fourier domain preconditioner is a viable alternative to multi-grid<br />
[188, 196]. Here the preconditionerM−1 consists of three operations:<br />
M −1 = F −1 ˜ M −1 F,<br />
whereF represents the 2D Fourier transform <strong>and</strong> ˜ M −1 a sparse Fourier domain operation.<br />
It is also possible to transform the whole problem into the Fourier domain, but this yields<br />
no further reduction in computational dem<strong>and</strong>s [188]. The Nslog(Ns) cost of the Fourier<br />
transforms dominates the computational dem<strong>and</strong> of this preconditioner.<br />
Besides the sparse matrix <strong>and</strong> iterative approaches, several other efficient algorithms<br />
are reported in literature. Examples are the direct Fourier domain approach in [143]<br />
<strong>and</strong> the hierarchic approach in [124]. The computational dem<strong>and</strong>s of Fourier domain<br />
approaches are dominated by the Nc ∝ Nslog(Ns) cost of the Fourier transform. For<br />
the hierarchic approach this cost is claimed to be Nc ∝ Ns, but similar to the direct<br />
P # <br />
approach described earlier no covariance matrix φφ T<br />
is considered <strong>and</strong> also the<br />
corrector influence matrixHis taken as an identity matrix.<br />
It can be concluded that for the traditional <strong>control</strong> law with a static reconstructor R,
3.3 Scaling problems 53<br />
many efficient algorithms are available. However, linear scaling of the computational dem<strong>and</strong><br />
(i.e. Nc ∝ Ns) is only reported when specific assumptions or approximations are<br />
made on G, H <strong>and</strong> φφ T<br />
[72, 124]. This may still be acceptable for large telescopes,<br />
since the involved operations can be well parallelized <strong>and</strong> the computation speed of computation<br />
hardware still increases. The fact that the traditional <strong>control</strong> law in general does not<br />
provide the best achievable performance is then taken for granted.<br />
The H2 optimal <strong>control</strong> law<br />
In its most generic form, the computational dem<strong>and</strong> of the H2 optimal <strong>control</strong>ler in (3.12)<br />
involves the products of all four system matrices with the corresponding vectors. This<br />
dem<strong>and</strong> can be quantified as (figure 3.5):<br />
Nc = N 2 x +NxNs+NxNa +NaNs,<br />
whereNx denotes the number of states. It is possible to reduce this computational dem<strong>and</strong><br />
using a state transformation into for example the output normal [186] or block companion<br />
form [161]. However, this approach does not reduce the order of magnitude of the dem<strong>and</strong><br />
<strong>and</strong> therefore offers no solution to the scaling problem. Moreover, such transformations<br />
usually have a detrimental effect on the numerical conditioning of the system matrices <strong>and</strong><br />
thus on the <strong>control</strong>ler’s performance.<br />
The computational dem<strong>and</strong> depends highly on the number of states Nx, which makes the<br />
relation between Nx <strong>and</strong> the aperture diameter of special importance to be able to relate<br />
the computational dem<strong>and</strong> to the aperture size. However,H2 optimal <strong>control</strong>ler design for<br />
AO is only recent work <strong>and</strong> there is little literature available on possibilities for efficient<br />
implementations; neither for the approaches based on physical modeling of the wavefront<br />
disturbance [16, 123, 150] nor for the data driven ones [100]. An important unknown is the<br />
relation between the number of sensors <strong>and</strong> the required number of states for a atmospheric<br />
disturbance model identified from measurement data to maintain a sufficient level of accuracy.<br />
For the analysis at h<strong>and</strong> it will therefore be assumed that the number of states Nx is<br />
linearly proportional to the number of actuators.<br />
Consequently, the computational dem<strong>and</strong> of implementations of theH2 optimal <strong>control</strong> law<br />
is also quadratically proportional to the number of actuators Na <strong>and</strong> to the fourth power<br />
with the aperture diameterDt. The same scaling problems will thus have to be faced as for<br />
the traditional approach. However, since optimal <strong>control</strong> for AO is more recent work, few<br />
approaches are documented in literature.<br />
The system matrices are usually derived from the solution of an algebraic Riccati equation.<br />
This makes it very difficult to enforce a structure for the resulting system matrices that<br />
can be exploited for an efficient implementation. For the data driven approach described<br />
in [100] the structure of the optimal <strong>control</strong>ler is given more explicitly in terms of the disturbance<br />
model, the corrector model <strong>and</strong> the sensor geometry. This approach allows more<br />
freedom to enforce structure to be exploited for implementation efficiency. It is shown in<br />
[100] that similar to the traditional <strong>control</strong> law the H2 optimal <strong>control</strong> law contains a least<br />
squares inversion of the sensor geometry matrixG. Depending on the model chosen for the<br />
corrector this is also the case for the influence matrixH. Both steps can be efficiently done<br />
using methods described for the traditional <strong>control</strong> laws. Further, the model set to describe<br />
the atmospheric disturbance can be chosen as desired, although this will affect its quality.<br />
3
3<br />
54 3 Efficient <strong>control</strong> for AO: concepts <strong>and</strong> challenges<br />
These model structures will be more extensively discussed throughout this thesis in section<br />
3.6 <strong>and</strong> chapter 4.<br />
3.3.2 Practical problems<br />
Besides computational problems there are many problems of a more practical nature that<br />
arise when the number of actuators is increased [21]. The actuators usually require an analog<br />
<strong>control</strong> signal that is provided by driver electronics. In case of thous<strong>and</strong>s of actuators,<br />
it is desirable to keep the interconnections between drivers <strong>and</strong> actuators short to limit mechanical<br />
defects, cross-talk, electromagnetic interference, etc. This can be realized using<br />
driver modules that drive a single actuator or a small group <strong>and</strong> are positioned close to<br />
them. Fast, digital communication links can then be used between the <strong>control</strong> system <strong>and</strong><br />
the drivers for setpoint updates.<br />
However, digital communication links require additional electronics <strong>and</strong> introduce latency<br />
that reduces the achievable performance of the AO system <strong>and</strong> should if possible be avoided.<br />
This has given rise to the concept of distributed <strong>control</strong> that will be discussed in the next<br />
section.<br />
3.4 Distributed <strong>control</strong><br />
The distributed <strong>control</strong> concept used throughout this thesis is based on the desire for a<br />
<strong>control</strong> system that is <strong>modular</strong>ly extendible both in hardware <strong>and</strong> in software (algorithms).<br />
Extendible in hardware means that the hardware consists of modules that all contain<br />
processing power <strong>and</strong> interface with each other, with the DM actuators <strong>and</strong> with the WFS.<br />
This implies that neither the computational dem<strong>and</strong> of each module, nor the number<br />
of connections to other modules, nor the number of data words to be communicated<br />
by each module may depend on the total number of DM actuators. Each module may<br />
receive only a fixed number of measurements from the WFS, which – as the system is<br />
extended – becomes a smaller <strong>and</strong> smaller fraction of the total size of y(t). It can only<br />
obtain other measurements indirectly via communication with other modules. A <strong>control</strong>ler<br />
implementation architecture is chosen for this concept in which each actuator is <strong>control</strong>led<br />
by its own <strong>control</strong>ler – implemented on a local hardware module – which can communicate<br />
with a the <strong>control</strong>lers of a few neighboring actuators.<br />
When considering the frozen flow characteristic of the atmospheric disturbance, the<br />
measurements that are the most relevant to a local <strong>control</strong>ler – e.g. for prediction – are<br />
those that correspond to a location in the wavefront nearby its actuator. In particular the<br />
nearby upstream measurements provide accurate information for short-term prediction, but<br />
the flow directions of the disturbance may vary <strong>and</strong> are not a priori known. Therefore, no<br />
directionality will be used in the selection of available sensors. It will be assumed that each<br />
<strong>control</strong>ler has access to all measurements taken within a small, fixed radius from its actuator.<br />
The influence of the actuators on the wavefront correction is assumed to be the identity<br />
operation: the vector of actuator comm<strong>and</strong>s equals the wavefront correction. The assumption<br />
of a Fried geometry then leads to the actuator/sensor layout in figure 3.6. In practice<br />
the spatial <strong>dynamics</strong> of a DM is often approximated by a static influence matrix. Although
3.4 Distributed <strong>control</strong> 55<br />
Figure 3.6: Schematic of the<br />
proposed distributed <strong>control</strong><br />
concept. The boxes represent<br />
actuators with their designated<br />
<strong>control</strong>lers <strong>and</strong> the<br />
dots mark the measurement locations<br />
of the WFS.<br />
an arbitrary matrix may significantly affect the performance of a distributed system (e.g.<br />
consider the case that each distributed <strong>control</strong>ler node is wired to a r<strong>and</strong>om actuator), for<br />
most DMs each actuator has a local effect on the wavefront. Distributed <strong>control</strong>lers can<br />
then coordinate their actions, such that performance loss may be limited. However, this is<br />
considered an important subject of further research.<br />
In figure 3.6 each box represents a <strong>control</strong>ler with its corresponding DM actuator <strong>and</strong> each<br />
dot the location of gradients measured by the WFS. The solid arrows indicate undirected<br />
communication that is possible between the central <strong>control</strong>ler node <strong>and</strong> those within a radius<br />
indicated by the solid circle. Similarly, the dashed arrows indicate the sensors whose<br />
measurements are available to the central node. These sensors lie within a radius indicated<br />
by the dashed circle. These radii may be different, as long as they are independent of the<br />
number of degrees of freedom of the AO system. This method of selecting nodes <strong>and</strong> sensors<br />
for direct communication will be used throughout this thesis for reasons of simplicity.<br />
3.4.1 Hardware considerations<br />
The proposed concept does not automatically imply that each <strong>control</strong>ler is implemented<br />
in its own <strong>control</strong> hardware. For the new DM designed within this project, each actuator<br />
module containing 61 actuators is driven by a single driver board. These boards could be<br />
augmented by computation hardware for a similarly clustered implementation of 61 local<br />
<strong>control</strong>lers.<br />
Depending on the communication requirements of the local <strong>control</strong>lers, the boards must<br />
have communication links with surrounding boards. When the DM design would be extended<br />
with another module of 61 actuators, this module would only need to be given its<br />
own electronics board with its communication links to its neighbors.<br />
This allows the hardware of each electronics board to be identical, which reduces production<br />
costs for large systems. The modules can be used for correctors with various numbers<br />
of actuators <strong>and</strong> can be easily replaced in case of failure. They include the required driving<br />
electronics to prevent the need for additional communication <strong>and</strong> be positioned close to the<br />
actuators.<br />
A final element that has so far not been considered is how each <strong>control</strong> module receives its<br />
3
3<br />
56 3 Efficient <strong>control</strong> for AO: concepts <strong>and</strong> challenges<br />
WFS measurements. Within this research project this sensor has been considered a given<br />
part of the system <strong>and</strong> its properties have not been investigated. However, an extendible<br />
scenario could be foreseen in which each CCD or quad-cell chip of the WFS is read out by<br />
an electronics module that calculates the wavefront gradients <strong>and</strong> uses a fast, one-to-many<br />
type of communication link to send the results to all relevant <strong>control</strong>ler modules.<br />
3.4.2 Control considerations<br />
A distributed hardware architecture has major consequences for the synthesis <strong>and</strong> performance<br />
of a <strong>control</strong> law that must be thoroughly evaluated. Since each actuator has a <strong>control</strong>ler<br />
with associated processing power, the available processing power scales linearly with<br />
the number of actuators (figure 3.5). As mentioned, algorithms with linearly scaling computational<br />
cost are available for wavefront reconstruction, but it will be shown in the sequel<br />
of this section that these algorithms are not readily suitable for the distributed architecture.<br />
Moreover, it has been shown by e.g. [16, 100, 150] that anH2 optimal <strong>control</strong> design offers<br />
major performance advantages. They are mainly due to the incorporated disturbance model<br />
that provides accurate predictions based on past measurements. As the number of measurements<br />
increases for large telescopes, this accuracy <strong>and</strong> thus the optimal <strong>control</strong> advantage<br />
are expected to further increase.<br />
Therefore, in this thesis the distributed <strong>control</strong>ler design will be aimed at approaching the<br />
H2 optimal performance. Ideally, the <strong>control</strong> nodes together implement a <strong>control</strong> lawC that<br />
minimizes the criterion in (3.4) subject to the constraint of a distributed structure. This constraint<br />
will inevitably reduce the <strong>control</strong> performance in comparison to an unconstrained,<br />
centralized <strong>control</strong>ler. As will be investigated in the next chapter, this reduction depends on<br />
the WFS measurements the nodes receive, on the number of neighbors that each <strong>control</strong>ler<br />
can communicate with <strong>and</strong> on the information that is communicated. However, the effect<br />
that sampling frequency has on the performance trade-off will not be considered. Sampling<br />
frequency determines the integration time of the WFS <strong>and</strong> thus the Signal to Noise<br />
Ratio (SNR) of the measurements. It is conceivable that in a distributed setting the sampling<br />
frequency that gives optimal performance in terms of correction quality is different<br />
than in the centralized setting.<br />
In the sequel of this section, an initial investigation will be presented of the consequences of<br />
the distributed architecture for the implementation of existing <strong>control</strong>ler designs. The main<br />
challenges will be shown together with possible solutions that will be further elaborated in<br />
the sequel of this thesis.<br />
3.5 Challenges<br />
As mentioned, the proposed distributed architecture restricts the total available processing<br />
power to a multiple of the number of actuators. This implies that the access of local <strong>control</strong>lers<br />
must also be restricted to a sub-set of the measurements in y(t). The significance<br />
of this can be shown for the traditional <strong>control</strong> law <strong>and</strong> will also be elaborated in chapter 4.<br />
As mentioned in section 3.2.2, the traditional <strong>control</strong> law consists of two parts: the<br />
reconstructor <strong>and</strong> the integrator. Since the integrators are independent SISO <strong>control</strong>lers,<br />
this part directly fits the distributed architecture. However, for the reconstructor part this is
3.5 Challenges 57<br />
entirely different. The open-loop reconstructed comm<strong>and</strong> for actuator i is a weighted sum<br />
of all WFS measurements. These weights form the i th row of the reconstruction matrix<br />
R <strong>and</strong> for the solution in (3.9) all elements are non-zero. This implies that each actuator<br />
requires access to all measurements, which is not allowed for the distributed architecture.<br />
In fact, since the available computational power is linearly proportional to the number of<br />
actuators, algorithms with a higher scaling of computational dem<strong>and</strong> are not suitable. They<br />
may be implemented for a certain number of actuators, but require a redesign of the <strong>control</strong><br />
hardware if this number increases.<br />
Although there are only a few wavefront reconstruction algorithms with a claimed linear<br />
scaling of computational dem<strong>and</strong> [72, 124], it is interesting to briefly consider several<br />
principles of these algorithms in more detail. The conjugate gradient methods used<br />
in [72, 73, 188] involve the step distance calculations for α <strong>and</strong> β given in (3.13) <strong>and</strong><br />
(3.14). Both expressions contain in-products over the entire residual vector that cannot<br />
be evaluated on the distributed architecture. Also the multi-grid <strong>and</strong> Fourier domain<br />
preconditioners of these methods involve operations that do not fit its structure. The<br />
interpolation <strong>and</strong> restriction operations of the first require a hierarchic communication<br />
structure, whereas the distributed communication structure has only a single-level. The<br />
recursive, fast implementation of the Fourier transform requires a similar structure.<br />
On the other h<strong>and</strong>, in [124] an approach is presented that does actually fit the distributed<br />
architecture. Local <strong>control</strong>lers are designed that receive only a few WFS measurements<br />
from a small area surrounding the actuator. Global performance is partly recovered by using<br />
the previous phase estimates ˆ φ(t−1) that are locally available, which is possible because<br />
the wavefront distortion is strongly correlated in both space <strong>and</strong> time. However, according<br />
to the authors this comes at the cost of a significantly smaller rejection of disturbances<br />
with low spatial frequency that can be recovered by using again a hierarchic architecture.<br />
Without changing to a hierarchic structure, this concept of using past, local phase estimates<br />
to recover global performance is considered within a more general framework in the next<br />
chapter.<br />
When considering the distributed <strong>control</strong> problem within theH2 optimal <strong>control</strong> framework<br />
then the structural constraints imposed on the sought <strong>control</strong>ler lead to a non-convex<br />
optimization problem. Results have been shown for distributed systems satisfying certain<br />
spatial invariance properties [14, 41, 79, 129], but no generic results are available for distributed<br />
H2 optimal <strong>control</strong>. This suggests that for efficient synthesis of distributed <strong>control</strong>lers,<br />
it is important to design the AO system components appropriately.<br />
However, not all AO system components can be freely designed. Particularly for this<br />
project, the WFS is considered to be given as a Shack-Hartmann sensor. This sensor leads to<br />
the wavefront reconstruction step discussed for the traditional <strong>control</strong> approach, for which<br />
no scalable distributed implementations are currently available.<br />
Further, <strong>control</strong>ler synthesis for an AO system is a disturbance rejection problem. Although<br />
the structure for the disturbance model can be chosen as desired, the disturbance itself can<br />
obviously not be influenced. It is possible to choose structured model sets with amenable<br />
properties for distributed <strong>control</strong>, but this may come at the cost of performance. This forms<br />
one of the possible concepts to overcome the problems for distributed <strong>control</strong> that will be<br />
discussed in the remainder of this chapter.<br />
3
3<br />
58 3 Efficient <strong>control</strong> for AO: concepts <strong>and</strong> challenges<br />
3.6 Possible solutions<br />
Given the mentioned problems that arise from the choice for a distributed hardware architecture<br />
of the <strong>control</strong> system, several (partial) solutions will now be introduced <strong>and</strong> discussed<br />
of which some are already suggested in literature.<br />
3.6.1 Phase reconstruction through analog electronics<br />
One of the first functional AO systems used a shearing interferometer [98] to measure<br />
the wavefront disturbance. Similar to the Shack Hartmann WFS this sensor needs postprocessing<br />
to estimate the phase disturbance. At that time, computing power was almost<br />
non-existent <strong>and</strong> phase reconstruction could not be done numerically.<br />
The sensor geometry of the shearing interferometer corresponds to the Hudgin geometry<br />
[98] <strong>and</strong> yields a very sparse, fixed geometry matrix G. This geometry can be translated<br />
to a simple electronic resistor network, where node voltages correspond to wavefront phase<br />
<strong>and</strong> resistor currents to gradient measurements. Phase reconstruction – i.e. the application<br />
of G # to the vector of gradient measurements – can be performed using this circuit<br />
by applying the measured gradients as currents using current sources <strong>and</strong> measuring the<br />
node voltages [96]. The effect of measurement noise can be included by applying a suitable<br />
amount of noise to the currents <strong>and</strong> adding shunt capacitances between the nodes <strong>and</strong><br />
ground.<br />
The settling speed of the network depends on the design of the current sources <strong>and</strong> the shunt<br />
capacitances <strong>and</strong> was reported to be in the order of ms. Although this settling time depends<br />
on the network size <strong>and</strong> thus the number of measurementsNs, studies have shown that this<br />
principle can be extended to systems with 10000 Degrees Of Freedoms (DOFs) [98] <strong>and</strong><br />
can even incorporated into the CCD chip of the WFS.<br />
For this the patented principle must be adapted to the Fried geometry. However, it does<br />
not allow knowledge of phase statistics to be included, but e.g. for the data driven optimal<br />
<strong>control</strong> approach in [100] this is not required. The <strong>control</strong>ler derived therein involves only<br />
the unweighted pseudo-inverseG # .<br />
3.6.2 A distributed disturbance model<br />
When considering the frozen flow properties that are generally attributed to the disturbance,<br />
it can be concluded that the information most relevant for predicting it at a discrete point is<br />
found in a small area surrounding that point. This suggests that the reduction in prediction<br />
accuracy between an unstructured model <strong>and</strong> a distributed model structure will be limited.<br />
In literature this has been verified by [45]. Here an FIR type predictor is used to predict the<br />
local open-loop sensor readings (i)(t) fori = 1...Ns of the form:<br />
ˆs (i)(t) = <br />
A(i,j)(z)−1 s (j)(t),<br />
j∈Ni<br />
whereA (i,j)(z) are scalar, monic polynomials in the z-transform variablez <strong>and</strong> the setsNi<br />
contain all indices of measurements within a small patch around sensor i includingiitself.<br />
Optimality of this predictor structure implies an underlying auto-regressive model structure
3.6 Possible solutions 59<br />
for the Wide Sense Stationary (WSS) data generating process:<br />
<br />
A (i,j)(z)s (j)(t) = e (i)(t),<br />
j∈Ni<br />
where e (i)(t) are zero-mean, uncorrelated white noise signals. Since in practice the wavefront<br />
disturbance causes all measurements to be somehow correlated, the achievable accuracy<br />
of this model depends on the choice for the setsNi: the larger the areas containing the<br />
locally available sensors are chosen, the higher the prediction accuracy.<br />
In the next sections, the property that the disturbance signals at nearby wavefront positions<br />
are highly correlated will be exploited for the design of distributed wavefront reconstructors<br />
<strong>and</strong> predictors.<br />
3.6.3 Iterative distributed phase reconstruction<br />
Although the iterative conjugate gradient algorithms discussed in section 3.3.1 are not suitable<br />
for a distributed architecture, this is not the case for all iterative solvers. Consider the<br />
partial reconstruction problem from open-loop gradient measurements s(t) to phase ˆ φ u (t)<br />
as posed in (3.10) on page 47. If the covariance matrices in (3.9) are taken as φs T = G T<br />
<strong>and</strong> ss T = G T G <strong>and</strong> a Fried or Hudgin geometry is used to constructG, then the update<br />
equations of the Jacobi <strong>and</strong> Steepest Descent (SD) algorithms [152, 197] fit the distributed<br />
communication structure. This is indicated by the sparsity structures of the matrices G T<br />
<strong>and</strong>G T G shown in figures 3.3 <strong>and</strong> 3.4.<br />
Using iteration indexm<strong>and</strong> step size α, the update equation <strong>and</strong> residual vectorr (m)(t) of<br />
the SD algorithm to solve (3.6) can be written as:<br />
ˆφ (m+1)(t) = ˆ φ (m)(t)−αr (m)(t), (3.15)<br />
r (m)(t) = G T G ˆ φ (m)(t)−G T s(t). (3.16)<br />
Similar to the conjugate gradient algorithm in (3.13), for fastest convergence the step sizeα<br />
should be determined at each iteration as [152]:<br />
α (m) = rT (m) r (m)<br />
rT (m) Ar .<br />
(m)<br />
Since the involved in-products do not fit the distributed framework, α can alternatively be<br />
chosen a priori fixed. When assuming a zonal sensor geometry (i.e. the gradient measurements<br />
are expressed in terms of adjacent phase points),Ghas a sparse, distributed structure.<br />
Consequently, each row i of the productG T G will have only a few non-zero elements per<br />
row, located at columnsj, where phase pointj is located nearbyi. The update equation in<br />
(3.15) thus fits a distributed communication structure <strong>and</strong> the number of computations per<br />
node per iteration is fixed.<br />
However, the number of iterations mc required for convergence is not fixed. Using (3.15),<br />
the propagation of the residual vector in (3.16) can be expressed as:<br />
r (m+1) = Wr (m), where W = I−αG T G.<br />
The iterations are usually stopped when the Frobenius norm of the residual drops below a<br />
certain fraction ǫ of that of the initial residual r (0)(t) = G T G ˆ φ (0)(t) − G T s(t), where<br />
3
3<br />
60 3 Efficient <strong>control</strong> for AO: concepts <strong>and</strong> challenges<br />
ˆφ (0)(t) is the initial guess. The Frobenius norm of the residual after a single iteration must<br />
satisfy r (m+1)(t)F ≤ W·r (m)(t)F , where W denotes the spectral norm of W.<br />
Accordingly, the norm of the residual afterm iterations becomes:<br />
r (m)(t)F ≤ W m r (0)(t)F.<br />
SinceW is symmetric, its spectral normW is the square of its spectral radiusρ(W) <strong>and</strong><br />
thus iterations are stopped whenρ(W) 2m < ǫ.<br />
For fastest convergence, the fixedα should thus be chosen that minimizesρ(W). The eigenvalues<br />
λ(W) can be expressed in terms of λ(G T G) as λ(W) = 1 −αλ(G T G). Let the<br />
eigenvalues of the symmetric, positive semi-definite matrixG T G be spread on the positive<br />
real axis between 0 <strong>and</strong> λ. The zero eigenvalues ofG T G (i.e. its kernel) correspond to the<br />
unseen modes that cannot be reconstructed without additional (e.g. statistical) information.<br />
They are mapped toλ(W) = 1 irrespective ofα, which can also be directly observed from<br />
the update law. The residual in (3.16) is orthogonal to the unseen modes such that by (3.15)<br />
the energy of these modes in ˆ φ (m+1)(t) is unaffected. If the smallest positive eigenvalue of<br />
G T G is denotedλ, then:<br />
1−αλ ≤ λ(W) ≤ 1−αλ.<br />
The α that minimizes max ρ(W) must thus satisfy |1 − αλ| = |1 − αλ| <strong>and</strong> is found as<br />
αo = 2/(λ+λ). The minimal spectral radius can now be expressed as:<br />
ρ(W) = λ/λ−1<br />
λ/λ+1 ∝ γNφ +δ −1<br />
, (3.17)<br />
γNφ +δ +1<br />
where the second step uses the fact that in case of a Fried geometry [98], the matrixGis a<br />
discrete Laplacian matrix. This yields an eigenvalue spread of the productG T G such that<br />
in approximation the ratioλ/λ is linearly proportional to the number of phase pointsNφ.<br />
When assuming thatr (0)F is invariant to the number of phase pointsNφ, after substitution<br />
of (3.17) intoρ(W) 2m < ǫ <strong>and</strong> solving form, the required number of iterationsmc is found<br />
as:<br />
mc ∝<br />
logǫ<br />
2log(γNφ +δ −1)−2log(γNφ +δ +1) ,<br />
∝ − 1<br />
4 Nφlogǫ forNφ ≫ 1.<br />
In approximation, the required number of iterations <strong>and</strong> thus computations increases<br />
linearly inNφ, which is not allowed for the distributed architecture.<br />
However, the required number of iterations mc is not only determined by the convergence<br />
speed of the algorithm. If the initial guess ˆ φ (0)(t) is already close to the solution,<br />
then only a few iterations may be required for convergence. Since the disturbance can be<br />
well approximated as r<strong>and</strong>om walk process, it makes sense to use the previous solution as<br />
the initial guess for the current problem, i.e. ˆ φ (0)(t) = ˆ φ (mc)(t−1).<br />
In figure 3.7 the number of steepest descent iterations required for convergence (ǫ = 10 −4 )<br />
is shown that was evaluated using numerical simulations on an artificially created data<br />
set. First, a static phase screen with Kolmogorov spatial statistics was generated using a<br />
midpoint displacement algorithm [95]. Then this was interpolated over a square aperture
3.6 Possible solutions 61<br />
Number of iterations m c<br />
10 2<br />
10 1<br />
10 1<br />
ˆφ (0)(t) = 0<br />
ˆφ (0)(t) = ˆ φ(mc)(t − 1)<br />
10 2<br />
Number of phase points Nφ<br />
Figure 3.7: The average number of SD iterations<br />
required for phase reconstruction withǫ = 10 −4<br />
evaluated on synthetic disturbance data.<br />
10 3<br />
Normalized variance accounted for [%]<br />
100<br />
90<br />
80<br />
70<br />
60<br />
50<br />
40<br />
10 SD iterations, no delay<br />
10 SD iterations, 1 sample delay<br />
30<br />
0 200 400 600 800 1000<br />
Number of phase points [−]<br />
Figure 3.8: Performance in terms of (3.18) using<br />
10 iterations of SD with ˆ φ (0)(t) = ˆ φ (10)(t−1)<br />
for measurement delays d of 0 <strong>and</strong> 1 samples.<br />
window that was translated over the phase screen, thus simulating a frozen flow <strong>and</strong><br />
yielding φ(t) for t = 0...5000. The open-loop measurement data set s(t) was then<br />
obtained through the sensor model in (3.7) where the variance of the white measurement<br />
noise signaln(t) was chosen according to a SNR of 20dB.<br />
As can be seen from the figure, the number of iterations can be reduced by using the<br />
previous reconstruction result as the initial guess for the new sample. However, the number<br />
of iterations still increases with the number of phase points Nφ, which means that this<br />
extension to the SD algorithm provides insufficient efficiency gain to fit the distributed<br />
architecture.<br />
A similar scheme is shown in [192], where the authors propose the use of single iterations<br />
of Jacobi <strong>and</strong> Richardson [152] per sample for closed-loop wavefront reconstruction<br />
with an ideal DM. They use the traditional integrator <strong>control</strong>ler <strong>and</strong> show that for small<br />
integrator gains, the approach leads to a smaller noise propagation compared to using<br />
R = G # . However, they do not analyze the performance effects of increasing the number<br />
of DOF of the AO system.<br />
Alternatively this performance trend can be analyzed by keeping the number of SD<br />
iterations constant, while increasing the number of DOF. This has yielded the results in<br />
figure 3.8. This shows the performance obtained using 10 SD iterations <strong>and</strong> ˆ φ (0) (t) =<br />
ˆφ (10)(t − 1) as evaluated for the same data sets as previously while considering zero <strong>and</strong><br />
one sample measurement delay. The performance is measured using the following Variance<br />
Accounted For (VAF) criterion:<br />
<br />
VAF =<br />
1−<br />
5000 t=0 s(t)−Gˆ φ (10)(t)2 F<br />
5000 t=0 s(t)2 <br />
·100%. (3.18)<br />
F<br />
The VAF values in the figure are normalized to those of a centralized solution to show<br />
the trend in the performance lost due to the chosen communication structure. For the case<br />
without measurement delay this centralized solution is taken as ˆ φ(t) = G # s(t) <strong>and</strong> for a<br />
3
3<br />
62 3 Efficient <strong>control</strong> for AO: concepts <strong>and</strong> challenges<br />
one sample delay this is taken as an FIR prediction:<br />
ˆφ(t) = B1s(t−1)+...+Bns(t−n) (3.19)<br />
with n lags. The coefficient matrices Bi are estimated from the data set as the minimizers<br />
of the cost function:<br />
<br />
s(t)−G n<br />
<br />
<br />
<br />
JB = Bks(t−k) .<br />
k=1<br />
The number of FIR lags has been chosen n = 10, yielding the highest VAF values for a<br />
separate validation data set.<br />
Observe from figure 3.8 that for a small number of DOF <strong>and</strong> no delay the 10 SD iterations<br />
yield almost the same performance as the centralized solution which deteriorates as<br />
the number of DOF increases. However, in all practical situations the measurement delay<br />
will be at least one sample, in which case the performance will significantly drop (figure<br />
3.8). Although the depth of this drop depends on the disturbance behavior, the trend for an<br />
increasing number of DOF is the same as for the zero delay case. The loss of performance<br />
due to the choice for a distributed architecture increases with the number of DOF. A concept<br />
will now be introduced to improved this trend.<br />
3.6.4 Recursive adaptive distributed reconstruction <strong>and</strong> prediction<br />
The discussed SD <strong>and</strong> Jacobi algorithms can be further extended to also include knowledge<br />
of the spatio-temporal <strong>dynamics</strong> of the wavefront disturbance. As suggested both by the<br />
Kolmogorov structure function in (2.1) on page 28 as well as by the frozen flow assumption<br />
discussed in section 2.1.2, this disturbance is highly correlated for nearby points in the<br />
wavefront. Now consider a network of processing nodes – one associated to each phase<br />
point – that performs both wavefront prediction <strong>and</strong> reconstruction. Each nodei updates its<br />
output ˆ φ (i)(t) using the following c<strong>and</strong>idate update law:<br />
ˆφ (i) (t) = <br />
j∈C (i)<br />
n−1 <br />
a ˆ<br />
(i,j) φ(j) (t−1)+<br />
<br />
k=0 j∈M (i)<br />
2<br />
F<br />
b (i,j,k)s (j)(t−k). (3.20)<br />
Here the sets C (i) <strong>and</strong> M (i) contain the indices of the phase points <strong>and</strong> sensor spots whose<br />
phase values <strong>and</strong> measurements are available for the prediction of φ (i)(t) respectively.<br />
Further, s (j)(t) is the two-element gradient measurement vector from sensor spot j <strong>and</strong><br />
n is the number of temporal lags of the Moving Average (MA) structure. Let the Auto-<br />
Regressive (AR) coefficientsa (i,j) of the filter structure be given a priori in accordance with<br />
a single iteration of the SD or the Jacobi algorithm. For the SD case described in (3.15) <strong>and</strong><br />
(3.16), the coefficientsa (i,j) form the (i,j) elements of the matrix A = I−αG T G. This<br />
implies that for α = αo the filter in (3.20) is marginally stable with two poles on the unit<br />
circle corresponding to the unseen modes of the Shack Hartmann sensor (SHS). Moreover,<br />
due to the sparsity structure of G T G (figure 3.4) the sets C (i) can be restricted to consist<br />
of i <strong>and</strong> the indices of the four phase points diagonally adjacent to i. This set becomes<br />
larger when a larger numberm > 1 of SD iterations is performed within one sample time,
3.6 Possible solutions 63<br />
C (i)<br />
S (j)<br />
j<br />
i<br />
W (i)<br />
M (i)<br />
Figure 3.9: Summary of the introduced<br />
sets of indices.<br />
W(i):gradients that are dependent on<br />
φ (i) <strong>and</strong> define the local cost function.<br />
M(i):gradients available for the output<br />
update ˆ φ (i) in (3.20).<br />
C(i): phase points used for the output update<br />
ˆ φ (i) in (3.20) as a result of the<br />
SD iteration(s).<br />
S(j):phase points defining the measurements(j)<br />
according to the Fried geometry.<br />
corresponding to A = (I−αG T G) m .<br />
Now let the MA termsb (i,j,k) be estimated as the minimizers of the cost function:<br />
s(t)−G <br />
J = φ(t−d) ˆ <br />
,<br />
j=1<br />
i∈S (j)<br />
2<br />
F<br />
<br />
Ns s(j)(t)− =<br />
<br />
G ˆ <br />
[j,i] φ(i)(t−d)<br />
2<br />
F<br />
<br />
, (3.21)<br />
where G [j,i] ∈ R2×1 denotes the (j,i) th block-element of the matrix G <strong>and</strong> the set S (j)<br />
contains the indices of the four adjacent phase points that define the two-element gradient<br />
vector s (j)(t). A schematic summary of the used sets is shown in figure 3.9. Further,<br />
d represents the number of samples delay due to the CCD based measurements <strong>and</strong> any<br />
processing time to obtain ˆ φ(t). Estimation of the coefficients b (i,j,k) can be performed<br />
by partitioning the cost function <strong>and</strong> using Least Mean Squares (LMS) or Recursive Least<br />
Squares (RLS) algorithms for each subsystem. The adaptiveness of these algorithms also<br />
allows to compensate for the slowly time-varying <strong>dynamics</strong> of the wavefront disturbance.<br />
Substitution of the update law in (3.20) into the cost function in (3.21) yields:<br />
<br />
Ns <br />
J =<br />
j=1<br />
<br />
<br />
˜s (j)(t)− <br />
<br />
2<br />
<br />
<br />
G [j,m]p (m)(t) <br />
,<br />
m∈S <br />
(j) F<br />
where ˜s (j)(t) = s (j)(t)− <br />
a ˆ<br />
(m,l) φ(l)(t−d−1),<br />
<strong>and</strong> p (m)(t) = <br />
l∈M (m) k=0<br />
m∈S (j)<br />
G [j,m]<br />
l∈C (m)<br />
n−1 <br />
b (m,l,k)s (l)(t−d−k).<br />
This cost function can be partitioned into overlapping, local pieces for each node i ∈<br />
{1...Nn}:<br />
J (i) = <br />
<br />
˜s(j)(t)− <br />
<br />
<br />
<br />
G [j,m]p (m)(t) . (3.22)<br />
j∈W (i)<br />
m∈S (j)<br />
2<br />
F<br />
3
3<br />
64 3 Efficient <strong>control</strong> for AO: concepts <strong>and</strong> challenges<br />
Each node minimizes its cost function w.r.t. the variables b (i,j,k) for j ∈ M (i) <strong>and</strong> k =<br />
0...n − 1. Let the sets W (i) be defined as the union of all sets S (l) for l = 1...Ns that<br />
containi. This assures that only those errors are weighted that can be directly influenced by<br />
the local variables.<br />
By partitioning the cost function <strong>and</strong> letting each node optimize its own piece, the global<br />
optimum is traded for a Nash equilibrium: optimization stops when no node can reduce<br />
its cost function by its own decision variables alone. Such a game theoretic objective is<br />
explicitly used in [147] to find a distributed <strong>control</strong>ler via an iterative process based on<br />
price mechanisms.<br />
The signals p (m)(t) form = i involve coefficientsb (m,l,k) that do not form variables local<br />
to node i. For node i to evaluate its cost function J (i), neighbors must communicate the<br />
value of these signals at each sample time. This also requires communication of the phase<br />
estimates ˆ φ (l)(t−d−1) from neighboring nodesl ∈ <br />
j∈M (i)<br />
<br />
m∈S (j) C (m).<br />
To arrive at a procedure to estimate the variables b (i,j,k) associated to each node i, let the<br />
local cost functionJ (i) from (3.22) be rewritten to:<br />
<br />
˜s(i)(t)−G J (i) = [W(i),L (i)]p (Li)(t) −G [W(i),i]s<br />
<br />
˜e (i)(t)<br />
T (i) (t)b <br />
<br />
(i) 2<br />
<br />
, (3.23)<br />
F<br />
where ˜s (i)(t) stacks all ˜s (j)(t) for j ∈ W (i) in lexicographical order <strong>and</strong> s (i)(t) stacks all<br />
s (l)(t−d−k) forl ∈ M (i) <strong>and</strong>k = 0...n−1. The vectorb (m) stacks the corresponding<br />
coefficients b (i,l,k) <strong>and</strong> the set L (i) is defined as L (i) = ( <br />
j∈M S (j)) \ {i}. Using a<br />
(i)<br />
completion of squares argument, the vector ˆ b (i) that minimizes J (i) can then be solved<br />
from: <br />
s (i)(t)γ (i)s T (i) (t)<br />
<br />
ˆb(i) = s (i)(t)G T [Wi,i] ˜e <br />
(i)(t) ,<br />
where the scalar γ (i) = G T [W (i),i]G [W(i),i]. Using e.g. the LMS algorithm [3, 119] this<br />
estimation can be performed recursively on-line by updating the parameter vector ˆ b (i) at<br />
each time instantt:<br />
<br />
ˆb (i)(t) = I−α (i)γ (i)s (i)(t)s T (i) (t)<br />
<br />
ˆb(i)(t−1)+α<br />
(i)s (i)(t)G T [W (i),i] ˜e (i)(t),<br />
where α (i) is the step size for node i. For all nodes these can either be tuned manually or<br />
optimal values can be derived from known statistical properties of the disturbance signals<br />
[3].<br />
An important issue with the proposed method is that neither the local nor the global cost<br />
functions weight the unseen piston mode (section 1.3.2). As a result, the energy corresponding<br />
to this mode will show r<strong>and</strong>om walk behavior that can be suppressed by adding a<br />
regularization termZ (i) of the form<br />
Z (i) =<br />
<br />
<br />
j∈W (i)<br />
<br />
<br />
<br />
ρ ˆ<br />
(i,j) φ(j)(t−d) <br />
<br />
<br />
to the local cost functions in (3.23). For suitable coefficients ρ (i,j), the regularization affects<br />
piston <strong>and</strong> waffle components with as little effect as possible on the remainder. The<br />
2<br />
F
3.6 Possible solutions 65<br />
Normalized variance accounted for [%]<br />
98<br />
97<br />
96<br />
95<br />
94<br />
93<br />
SD−LMS with 4 neighbors<br />
SD−LMS with 8 neighbors<br />
92<br />
0 200 400 600 800 1000<br />
Number of phase points [−]<br />
Figure 3.10: Results of the<br />
distributed SD-LMS reconstruction<br />
procedure applied<br />
to the artificial data sets of<br />
section 3.6.3.<br />
waffle mode is specific to the Fried geometry matrix (section 1.3.2) <strong>and</strong> can also be evaded<br />
by choosing e.g. the Hudgin geometry matrix. The unseen piston mode is fundamental to<br />
the Shack-Hartmann WFS, but incorrect, large piston values do not affect the AO correction<br />
quality <strong>and</strong> only require an overly large stroke of the wavefront corrector. Nevertheless,<br />
these are trade-offs that require further research.<br />
Figure 3.10 shows the results of performing this procedure for the artificial data set described<br />
in section 3.6.3 using d = 1, n = 3 <strong>and</strong> with sets M (i) that contain the indices of<br />
four <strong>and</strong> eight nearest sensor spots. For simplicity, the step sizes α (i) were chosen equal.<br />
Similar to figure 3.8 the results are expressed in terms of a VAF-value that has been normalized<br />
against that for the centralized predictor in (3.19). The VAF-value itself is the normalized<br />
average prediction/reconstruction error that determines the cost function in (3.21):<br />
VAF =<br />
<br />
1−<br />
t1<br />
t=t0 s(t)−Gˆ φ(t) 2 F<br />
t1<br />
t=t0 s(t)2 F<br />
<br />
·100%.<br />
The data set was evaluated over 5000 samples <strong>and</strong> the VAF values were computed using<br />
t0 = 3000 <strong>and</strong> t1 = 5000 to skip the settling time of the LMS algorithm. For these<br />
data sets, the VAF values corresponding to the traditional, static reconstructor R = G #<br />
that does not take the one sample delay into account is approximately 20%, stressing the<br />
importance of a predictor. Observe that for the four neighbor case, the normalized VAF<br />
value remains virtually constant, whereas for eight neighbors it increases with the number<br />
of DOF. This means that for the used synthetic data sets the difference with the centralized<br />
solution actually decreases as the system becomes larger. However, note that these results<br />
will depend highly on the spatio-temporal behavior of the wavefront disturbance. Nevertheless,<br />
the notion that the decrease of the locally available fraction of SHS measurements can<br />
be compensated by exploiting spatio-temporal correlations of the wavefront disturbance is<br />
promising. It will be addressed in more detail <strong>and</strong> using more general parameterizations of<br />
the local subsystems in the next chapter.<br />
3
3<br />
66 3 Efficient <strong>control</strong> for AO: concepts <strong>and</strong> challenges<br />
3.6.5 Local, identical influence functions<br />
Both the described traditional <strong>control</strong> law <strong>and</strong> the H2 optimal <strong>control</strong> law described in<br />
[100] involve an inverse of the influence matrix H of the wavefront corrector to fit actuator<br />
comm<strong>and</strong>s to a desired <strong>mirror</strong> shape. As mentioned, there exists extensive literature<br />
on performing this operation in a computationally efficient manner [53, 188]. However, to<br />
achieve this efficiency, these methods use approximations of the influence matrixHwhose<br />
accuracy could be improved by suitable design of the DM.<br />
For instance, in [53] the influence matrix is approximated as a sparse matrix. It is assumed<br />
that each actuator only affects the wavefront in a small area surrounding it <strong>and</strong> influence<br />
below a certain threshold is neglected. This approximation thus benefits from narrow influence<br />
functions that makeHmore sparse <strong>and</strong> the algorithm implementation more efficient.<br />
Other approaches [188] assume the DM influence functions to be spatially invariant. A circulant<br />
structure is attributed such that the influence matrix becomes a spatial filtering kernel<br />
with a sparse Fourier domain representation. This can be efficiently applied using the previously<br />
described Fourier domain techniques. Further, spatial invariance is important for the<br />
more generic distributed <strong>control</strong>ler synthesis methods described in [14, 41, 79, 129, 168].<br />
The spatial invariance approximation becomes more accurate when the variation between<br />
the shapes of different influence functions is reduced. These differences may e.g. be the<br />
result of manufacturing tolerances or variations in material properties (chapters 5 <strong>and</strong> 6) or<br />
edge effects of the DM (chapter 7).<br />
Further, the above efficient algorithms all assume the DM to be well modeled as a static<br />
gain, which means that their applicability or performance will be compromised if the DM<br />
has significant temporal <strong>dynamics</strong>. Moreover, the performance of a distributed <strong>control</strong>ler<br />
will deteriorate when the spatial propagation speed of the plant <strong>dynamics</strong> exceeds that of the<br />
inter-node communications. For quasi-static DMs with local influence functions, this means<br />
that after each <strong>control</strong> signal update the <strong>control</strong>ler nodes should be able to communicate to<br />
all nodes that can influence their measurements. A requirement for the inter-node communication<br />
distance can thus be reduced by a DM design with narrow influence functions.<br />
3.7 Conclusions<br />
In this chapter the scaling problems that arise when the number of DOF of the AO system<br />
increases were investigated. This has been based on both traditional approaches towards<br />
<strong>control</strong> design for AO <strong>and</strong> for more general, optimal approaches. For all methods the computational<br />
complexity is proportional to the number of DOF with an exponent larger than<br />
one, which indicates that for these cases the computational power of the <strong>control</strong> system<br />
must grow faster than the number of DOF. This leads to huge parallel processor systems for<br />
the new large telescopes such as the European Extremely Large Telescope (E-ELT) <strong>and</strong> the<br />
Thirty Meter Telescope (TMT). Since these systems must be custom-designed, they will be<br />
very costly. Without a <strong>modular</strong> architecture, such systems do not befit from the <strong>modular</strong> AO<br />
system to be realized within this research project.<br />
To arrive at a fully <strong>modular</strong> system, a <strong>control</strong> system with a distributed architecture is required.<br />
This was proposed, consisting of computational nodes that each drive a single actuator,<br />
receive a limited set of wavefront sensor measurements <strong>and</strong> can communicate to a few<br />
neighboring nodes. This structure was chosen for practical reasons <strong>and</strong> will significantly
3.7 Conclusions 67<br />
affect the achievable performance. This performance loss was investigated based on literature,<br />
but is difficult to quantify because it depends entirely on the behavior of the DM, the<br />
sensor <strong>and</strong> the atmospheric disturbances. Therefore, several approaches towards distributed<br />
<strong>control</strong> are proposed, some of which are evaluated in numerical simulations <strong>and</strong> compared<br />
to centralized solutions.<br />
A first idea is to use an electrical resistor network, similar to the one used for wavefront<br />
reconstruction in the first operational AO system. Such a network can perform basic wavefront<br />
reconstruction very fast, but is sensitive to variation in the properties of its analog<br />
components <strong>and</strong> does not allow to exploit any wavefront statistics. The latter is possible<br />
with the subsequently proposed distributed disturbance model whose performance has been<br />
verified in literature.<br />
An alternative for the analog resistor-network reconstructor can be an iterative, distributed<br />
wavefront reconstruction algorithm based on the SD algorithm. All computations of this algorithm<br />
are fully distributed, but the number of computations required per computation node<br />
grows with the spatial dimension of the system. This approach does not allow to exploit any<br />
wavefront statistics, but this becomes possible by extending it with LMS based wavefront<br />
prediction. Simulation results of this distributed, adaptive wavefront predictor <strong>and</strong> reconstructor<br />
showed that its performance actually increases with the number of DOF, but further<br />
research is required into dealing with the unseen waffle-mode of the Shack-Hartmann WFS.<br />
Finally, a number of properties was discussed in the context of a DM system that are often<br />
assumed of the plants in distributed <strong>control</strong>ler design approaches found in literature. A spatial<br />
invariance assumption may for instance be fulfilled by a suitable mechatronic design.<br />
However, the main challenge for <strong>modular</strong>ly distributed <strong>control</strong> for AO lies with wavefront<br />
reconstruction.<br />
3
ÔØÖÓÙÖ<br />
ØÖÚÒ×ØÖÙØ ÓÒØÖÓÐ<br />
To deal with the increase in the computational dem<strong>and</strong> for <strong>Adaptive</strong> Optics<br />
(AO) <strong>control</strong> systems of future large telescopes, a <strong>modular</strong>, distributed <strong>control</strong>ler<br />
structure is proposed consisting of local Auto-Regressive Moving<br />
Average (ARMA) <strong>control</strong>lers. A data based <strong>control</strong>ler design approach is presented<br />
that is elaborated for the case of an ideal wavefront corrector. The internal<br />
model <strong>control</strong> principle is used to transform the closed-loop problem<br />
into an open-loop problem while preserving the distributed structure. A twostage<br />
procedure is proposed to identify the unknown <strong>control</strong>ler coefficients<br />
from open-loop measurement data based on a minimum variance, output error<br />
criterion. Constrained optimization is used to guarantee stability of the<br />
identified open-loop <strong>control</strong>ler using Gershgorin’s circle theorem. Application<br />
results are presented on both experimental <strong>and</strong> synthetic data.<br />
69
4<br />
70 4 Data driven distributed <strong>control</strong><br />
4.1 Introduction<br />
In this chapter a synthesis approach will be proposed that yields a <strong>control</strong>ler that is<br />
suitable for the distributed system architecture proposed in chapters 2 <strong>and</strong> 3. The hardware<br />
architecture of the <strong>control</strong> system is assumed to be a network of small <strong>control</strong>ler modules<br />
whose hardware is identical. Here it will be assumed that each module <strong>control</strong>s a single<br />
actuator, but for an actual implementation this can also be a small group. The modules can<br />
communicate with a small number of neighbors <strong>and</strong> have only access to a small set of local<br />
WaveFront Sensor (WFS) measurements.<br />
As mentioned, such a distributed hardware architecture has major consequences for the<br />
synthesis <strong>and</strong> performance of a <strong>control</strong>ler that must be thoroughly evaluated. Firstly,<br />
because the available computational power will only increase linearly with the number of<br />
actuators. But also because the constraints that must be applied to the optimization problem<br />
in (3.4) to restrict the solution to distributed <strong>control</strong>lers render this problem non-convex.<br />
In the previous chapter the bottlenecks arising for traditional <strong>control</strong> laws have been<br />
investigated together with several approaches towards solutions. The main challenge for<br />
distributed <strong>control</strong> has been identified as the wavefront reconstruction step that in its pure<br />
form requires all sensor measurements for the reconstruction of each phase point. Further,<br />
as the optimality of the traditional <strong>control</strong> law has been shown to be subject to very specific<br />
conditions [56, 100, 123], in this chapter a more general, but distributed <strong>control</strong> law will be<br />
used. However, the distributed structure is considered to be the design driver <strong>and</strong> as will be<br />
shown this leads to a loss of performance when compared to centralized solutions.<br />
A distributed, minimum variance <strong>control</strong>ler will be proposed that is data-driven in the sense<br />
that its coefficients will be directly estimated from open-loop disturbance measurement<br />
data. The available measurements <strong>and</strong> data from neighboring nodes will be chosen such<br />
that they are unchanged by the transformation between the closed <strong>and</strong> open-loop <strong>control</strong>lers<br />
using the internal model principle. The particular <strong>control</strong>ler structure leads to the implicit<br />
assumption that the gradient measurements are generated by a network of interconnected,<br />
low order, Wide Sense Stationary (WSS) stochastic processes. Further, a Shack-Hartmann<br />
wavefront sensor is considered with a one sample delay, a Fried geometry <strong>and</strong> white<br />
measurement noise. Except for the delay, this sensor model corresponds to the one used in<br />
[53, 124]. Finally, for the sake of simplicity an ideal wavefront corrector will be assumed<br />
whose transfer matrix is the identity matrix, such that the actuator comm<strong>and</strong> vector is<br />
equal to the introduced wavefront correction. This means that both the temporal <strong>dynamics</strong><br />
<strong>and</strong> the influence the actuators have on each other via the reflective facesheet are ignored,<br />
hence this assumption is too restrictive for the Deformable Mirror (DM) system outlined<br />
in chapter 2. Nevertheless, it is adequate to investigate the feasibility of a distributed<br />
<strong>control</strong>ler structure for an <strong>Adaptive</strong> Optics (AO) system based on a Shack-Hartmann WFS<br />
that forms a major challenge for a distributed <strong>control</strong> architecture (chapter 3).<br />
This chapter is outlined as follows: in the next section several preliminaries are introduced,<br />
after which a detailed problem definition is given in section 4.3. In section 4.4 the<br />
approach towards the design of a distributed <strong>control</strong>ler is described. The closed-loop problem<br />
is first translated into the equivalent open-loop problem <strong>and</strong> then a system identification<br />
problem is formulated to identify the unknown coefficients from open-loop measurement<br />
data. In section 4.7 an algorithm will be proposed to do so. Finally, application results to
4.2 Preliminaries 71<br />
both breadboard measurement data as well as synthetic data will be presented in section 4.8.<br />
4.2 Preliminaries<br />
Throughout this chapter, individual, distributed, discrete time sub-systems will be denoted<br />
through a subscript index, e.g. C (i)(z). The global, interconnected systemC(z) is obtained<br />
by lifting the sub-systems in this index space <strong>and</strong> applying the interconnection constraints.<br />
The fact that this system is structured is reflected in the bold italic typeface used forC. The<br />
input <strong>and</strong> output vectors of the global system will be described with structured vectors that<br />
stack the vectors or scalars corresponding to local sub-systems. Such structured vectors are<br />
also denoted in a bold italic typeface as x instead of the regular bold typeface x used for<br />
unstructured vectors.<br />
Throughout this chapter the following symbols will be often used: the total number of nodes<br />
Nn, the number of lenselets Ns in the wavefront sensor array <strong>and</strong> the order n of the individual<br />
distributed <strong>control</strong>lers.<br />
For more general definitions of notation the reader is referred to section 1.1 <strong>and</strong> the nomenclature<br />
from page xvii onwards.<br />
4.3 Problem description<br />
Consider a linear, discrete time, WSS stochastic processT(z) that generates the wavefront<br />
phase disturbanceφ(t) = [φ (1)(t),...,φ (Nn)] T at phase grid points1...Nn <strong>and</strong> is driven<br />
by an unknown noise signal v(t) ∈ N Nn (0,Cv). The residual wavefront disturbances<br />
ǫ (i)(t) that affect the science image is influenced by the wavefront corrector (e.g. a DM).<br />
As discussed above, it is assumed that the transfer function matrix of the corrector is the<br />
identity matrix. Therefore, there will be no further reference to its <strong>dynamics</strong> <strong>and</strong> both its<br />
input <strong>and</strong> output signals will be called u (i)(t) for i = 1...Nn. Let the residual wavefront<br />
ǫ (i)(t) be defined as:<br />
ǫ (i)(t) = φ (i)(t)−u (i)(t).<br />
The residuals ǫ (i)(t) are observed through a Shack-Hartmann wavefront sensor that is<br />
modeled using the Fried geometry [179], a one sample delay <strong>and</strong> measurement noise<br />
w (j)(t) ∈ N2(0,C (w,j)) forj = 1...Ns. The Fried geometry has been treated in detail in<br />
section 1.3.2 on page 6 of which several important properties will here be briefly repeated.<br />
Firstly, the Fried geometry defines the closed-loop gradient measurementse (j)(t) ∈ R 2 in<br />
terms of the four surrounding phase values as:<br />
e (j)(t) = GFǫS (j) (t−1)+w (i)(t), (4.1)<br />
where the setsS (j) contain the indices of the four phase points that define the two gradients<br />
e (j)(t) as in (1.1) on page 8. Further, when e (j)(t) for j = 1...Ns <strong>and</strong> ǫ (i)(t) for<br />
i = 1...Nφ are stacked in the vectorse(t) <strong>and</strong>ǫ(t) respectively, the measurements can be<br />
expressed as e(t) = Gǫ(t−1)+w(t). The matrix G is rank deficient, which causes the<br />
piston <strong>and</strong> waffle modes to yield a zero measurement. Implications of this for <strong>control</strong> will<br />
be addressed at a later point in this chapter.<br />
4
4<br />
72 4 Data driven distributed <strong>control</strong><br />
Figure 4.1: Schematic of the<br />
design problem of the con-<br />
v<br />
w<br />
φ<br />
+ -<br />
+<br />
e<br />
trollerC(z). T(z) Gz -1<br />
u I<br />
u<br />
C(z)<br />
In this chapter a spatially interconnected network of discrete time <strong>control</strong>lers C (i)(z)<br />
with outputu (i)(t) is considered that together form the global <strong>control</strong>lerC(z) <strong>and</strong> minimize<br />
the output error variance (figure 4.1):<br />
Ĉ(z) = arg min<br />
C(z)<br />
2<br />
e(t)F .<br />
This optimization criterion does not consider a priori knowledge of statistics of the wavefront<br />
phase disturbance or the measurement noise, which is similar to the approach discussed<br />
in chapter 3 whereR(z) is restricted to a static matrixR such that:<br />
ˆR = argmin<br />
R<br />
s(t)−GRs(t) 2 F . <br />
When assuming s(t) ∈ N2Ns(0,I), this has the solution R = G # , where G # is the<br />
unweighted pseudo-inverse of G. The latter does not exist uniquely <strong>and</strong> in practice the<br />
Singular Value Decomposition (SVD) inverse is used that constrains the unseen modes in<br />
the reconstructed phase vector to zero. When evaluating the performance of the distributed<br />
<strong>control</strong>ler in section 4.8, thisR = G # will be used as a reference.<br />
Further, the local <strong>control</strong>lers will further be called <strong>control</strong> nodes. The communication<br />
between these nodes will be limited to their output signals u (i)(t) instead of their full state<br />
information. This limits the inter-node communication load <strong>and</strong> makes this independent<br />
of the order (state dimension) of the nodes. However, by denying the nodes full access to<br />
their neighbors’ states an additional loss of generality is introduced <strong>and</strong> a possible loss of<br />
performance.<br />
Let the inputs of node i consist of the measurements e (j)(t) for j ∈ M (i) <strong>and</strong> the outputs<br />
u (k)(t) of their neighbors k ∈ C (i). In the sequel of this chapter a specific choice will be<br />
made for the setsM (i) <strong>and</strong>C (i).<br />
4.4 Design approach<br />
The approach taken in this chapter to obtain the distributed <strong>control</strong>lerC(z) can be sketched<br />
as follows. First, a distributed parametrization is proposed for the <strong>control</strong>ler <strong>and</strong> a choice is<br />
made for the sets M (i) that specify which measurements are available to each node. Then,<br />
using the internal model principle the closed-loop synthesis problem is translated into an<br />
open-loop problem. The sets C (i) that specify which output estimates are available to each<br />
node are defined such that this transformation does not compromise the goal that the output<br />
of each local <strong>control</strong>ler node is determined entirely on information from spatially nearby<br />
sources.<br />
ǫ
4.4 Design approach 73<br />
rc<br />
4.4.1 Parametrization of the distributed <strong>control</strong>ler<br />
1<br />
1<br />
Figure 4.2: Distributed <strong>control</strong><br />
setting. The boxes represent<br />
the <strong>control</strong>ler nodes <strong>and</strong><br />
the dots the WFS gradient<br />
measurements. Arrows indicate<br />
the information flow for<br />
the solid black node with its<br />
greyed neighbors <strong>and</strong> sensors<br />
located within a communication<br />
radius rc.<br />
In AO literature, several distributed <strong>control</strong>ler structures have been described. In [45] <strong>and</strong><br />
[52], local Finite Impulse Response (FIR) filters are described that predict open-loop measurement<br />
as a weighted estimate of past, local measurements. As shown for on-sky measurements<br />
[45] <strong>and</strong> experimental breadboard measurements [52], such local predictors are<br />
still capable of achieving small prediction errors. This is due to the strong local wavefront<br />
correlations both in space – indicated by Kolmogorov or Von Karman statistics – as well<br />
as in time – indicated by the frozen flow assumption. For the first, the phase covariance<br />
between two points in the wavefront decays with the distance between them [95] due to the<br />
definition of the structure function in (2.1). For the second, two adjacent points may even<br />
observe time-shifted versions of the exact same temporal phase fluctuations.<br />
However, in both [45] <strong>and</strong> [52] the wavefront reconstruction step is considered as a separate<br />
part of the <strong>control</strong>ler. Here this step is considered a critical part of the <strong>control</strong>ler, which<br />
significantly complicates its design. In [124] wavefront reconstruction is also implicitly<br />
considered <strong>and</strong> the following distributed <strong>control</strong>ler structure is proposed:<br />
C (i)(z) : u (i)(t) = <br />
a (i,j)u (j)(t−1)+ <br />
b (i,j)e (j)(t).<br />
j∈C (i)<br />
j∈M (i)<br />
This Auto-Regressive Moving Average (ARMA) <strong>control</strong>ler is similar to the one proposed<br />
in section 3.6.4, except for the presence of only a single Moving Average (MA) term.<br />
The coefficients a (i,j) <strong>and</strong> b (i,j) correspond to the application at each sampling time<br />
of one or a few iterations of a distributed linear solver, starting from the last estimate<br />
of the previous sample. Therefore, these coefficients are determined from steady state<br />
convergence criteria. An important difference is that the inputs of <strong>control</strong>ler node i<br />
now also consist of the previous outputs of itself <strong>and</strong> neighboring <strong>control</strong>ler nodes:<br />
i.e. of u (j)(t − 1) for j ∈ C (i). At each time step, this leads to propagation of wavefront<br />
phase information over the network of nodes. This allows partial recovery of the<br />
performance loss arising from gradient measurements being unavailable to local <strong>control</strong>lers.<br />
In this chapter an approach will be proposed to improve the recovery of the performance<br />
loss, take into account the WFS delay <strong>and</strong> exploit local knowledge of the behavior of the atmospheric<br />
disturbance. To achieve this, the following class of <strong>control</strong>lers with a distributed<br />
4
4<br />
74 4 Data driven distributed <strong>control</strong><br />
input-output parametrization is proposed:<br />
C (i)(z) : <br />
a (i,j)(z)u (j)(t) = <br />
j∈C (i)<br />
j∈M (i)<br />
k (i,j)(z)e (j)(t), (4.2)<br />
where the polynomialsa (i,j)(z) <strong>and</strong> k (i,j)(z) are of fixed ordernfor all nodes <strong>and</strong> defined<br />
as:<br />
n<br />
a (i,j)(z) = δij − a (i,j,l)z −l ,<br />
n−1 <br />
k (i,j)(z) =<br />
l=0<br />
l=1<br />
k (i,j,l)z −l .<br />
Since δij = 1 for i = j, the polynomial a (i,i)(z) is monic, whereas the first coefficients<br />
of a (i,j)(z) for i = j are zero. This is assumed for well-posedness of the interconnected<br />
system. Further, the coefficients k (i,j,l) are of size 1×2since each lenselet of the Shack-<br />
Hartmann WFS yields two gradient measurements. This parametrization is also proposed<br />
in [61], where a structured, multi-dimensional Transfer Function (TF) model is used for<br />
system identification. The input-output form circumvents the estimation of a local state<br />
sequence for each node, which is a yet unsolved problem.<br />
Similar to [45, 52], the sets M (i) will be defined through the radius rc ≥ 1 as the indices<br />
j of all measurements e (j)(t) taken at a distance from node i smaller than rc (figure 4.2).<br />
When the communication radius rc is increased until for all nodes i the set M (i) contains<br />
all j = 1...Ns, then the problem reduces to the centralized setting. Otherwise, the sets<br />
M (i) can be smaller for nodes near the border of the grid than for nodes in the center, as<br />
there are fewer sensors present within the communication radius. The same will hold for<br />
the setsC (i) that will be defined in the next section.<br />
4.4.2 Internal model <strong>control</strong><br />
The internal model <strong>control</strong> principle can be used to transform a closed-loop <strong>control</strong>ler into<br />
an equivalent open-loop <strong>control</strong>ler. Since the measurement inputs of an open-loop system<br />
are not affected by its outputs <strong>and</strong> in this case the forward plant modelGz −1 is a distributed,<br />
static gain, the coefficients of its update equation will be easier to identify.<br />
The closed-loop schematic depicted in figure 4.1 can be equivalently represented by the<br />
schematics in figures 4.3 <strong>and</strong> 4.4, where the latter more clearly shows the open-loop configuration.<br />
The lower grey area in figure 4.3 now corresponds to the <strong>control</strong>ler C(z), but consists<br />
internally of the open-loop <strong>control</strong>lerR(z) <strong>and</strong> the forward plant model. Expressed on a local<br />
scale, the inputss (j)(t) of the open-loop <strong>control</strong>lerR (i)(z) are equal to the closed-loop<br />
measurementse (j)(t) minus the effects of all <strong>control</strong>ler actions:<br />
s (j)(t) = e (j)(t)+GFu (S(j))(t−1). (4.3)<br />
Substitution of (4.3) fore (j) into the <strong>control</strong>ler parametrization of (4.2) yields the open-loop<br />
<strong>control</strong>ler:<br />
R (i)(z) : <br />
a (i,j)(z)u (j)(t) = <br />
k (i,j)(z)s (j)(t)− <br />
k (i,j)(z)GFu (S(j))(t−1).<br />
j∈C (i)<br />
j∈M (i)<br />
j∈M (i)
4.5 Performance 75<br />
T(z) Gz-1 + -<br />
v<br />
w<br />
φ<br />
+<br />
e<br />
u<br />
C(z)<br />
Gz -1<br />
R(z)<br />
s<br />
+<br />
Figure 4.3: A schematic equivalent to the<br />
closed-loop in figure 4.1.<br />
ǫ<br />
T(z) Gz -1<br />
R(z)<br />
v φ<br />
+<br />
s<br />
w<br />
Figure 4.4: An open-loop schematic equivalent<br />
to figures 4.1 <strong>and</strong> 4.3.<br />
Observe that this <strong>control</strong>ler has an additional term compared to the parametrization of<br />
C (i)(z). Each u (S(j))(t) in the additional term consists of the outputs of nodes l ∈ S (j).<br />
Due to the definition of the setsS (j) all these nodes are spatially located near nodei. To prevent<br />
any differences between the sets of available inputs of the nodesC (i)(z) <strong>and</strong>R (i)(z),<br />
the indices of these nodes should be included in C (i). Therefore, let the sets C (i) be defined<br />
accordingly as:<br />
C (i) = <br />
S (j).<br />
j∈M (i)<br />
The open-loop <strong>control</strong>lerR (i)(z) can then be parameterized more compactly as:<br />
R (i)(z) : <br />
j∈C (i)<br />
where the polynomialsã (i,j)(z) are equal to:<br />
ã (i,j)(z)u (j)(t) = <br />
<br />
−1<br />
ã (i,j)(z) = a (i,j)(z)+z<br />
c∈M (i)<br />
j∈M (i)<br />
k (i,j)(z)s (j)(t), (4.4)<br />
k (i,c)(z)G [c,j]<br />
<strong>and</strong>G [c,j] ∈ R 2×1 denotes the(c,j) block element of the geometry matrixG.<br />
4.5 Performance<br />
+ -<br />
u<br />
ǫ<br />
(4.5)<br />
For H2 optimal <strong>control</strong>ler synthesis, the <strong>control</strong>ler structure follows from the plant model<br />
structure, but in this case the desired distributed <strong>control</strong>ler structure is the design driver.<br />
The consequences of this choice for the <strong>control</strong>ler’s performance will now be investigated<br />
by showing the assumptions that this choice implies for the structure of the plant.<br />
Assuming that the network of open-loop <strong>control</strong>lersR (i)(z) is optimal in the sense that<br />
it minimizes the output error variance e (j)(t)2 F for all j, then the output errors must be<br />
white noise signals, i.e. e (j)(t) ∈ N(0,Cej). After application of the sensor model, the<br />
<strong>control</strong> signals u (i)(t) thus form the optimal one step ahead prediction for the open-loop<br />
measurement signals s (j)(t). Consequently, the network of open-loop <strong>control</strong>lers of (4.4)<br />
4
4<br />
76 4 Data driven distributed <strong>control</strong><br />
can be combined with the output equation in (4.3) to form an alternative process that generates<br />
the open-loop measurementss (j)(t) <strong>and</strong> is driven by the white noise inputse (j)(t). By<br />
considering this innovation form of the predictor, insight into the implicitly assumed structure<br />
of the data generating system is obtained. First recall those two equations <strong>and</strong> consider<br />
them part of a single system:<br />
⎧ <br />
⎨ ã (i,j)(z)u (j)(t) =<br />
j∈C (i)<br />
⎩<br />
<br />
k (i,j)(z)s (j)(t)<br />
j∈M (i)<br />
(4.6)<br />
s (j)(t) = GFu (S(j))(t−1)+e (j)(t)<br />
This process resembles a Kalman filter, but in absence of a well defined state this term is<br />
not applicable. When substituting the output equation of (4.6) for s (j)(t) into its ARMA<br />
equation <strong>and</strong> then using (4.5), this leads to:<br />
⎧ <br />
⎨ a (i,j)(z)u (j)(t) = <br />
k (i,j)(z)e (j)(t),<br />
⎩<br />
j∈C (i)<br />
j∈M (i)<br />
s (j)(t) = GFu (S(j))(t−1)+e (j)<br />
This equivalent representation of the data generating system thus has the same distributed<br />
structure as the <strong>control</strong>ler C(z) in (4.2) <strong>and</strong> is driven by white noise injected at the sensor<br />
locations. However, this form is not realistic as a physical model since the white noise signal<br />
that drives the process generating the phase distortionφ(t) is here coupled with the noise of<br />
the WFS. Nevertheless, if the proposed distributed structure is enforced on the <strong>control</strong>ler,<br />
this can only be the minimizer of the output error variance if the data generating system has<br />
a realization with the same spatially distributed structure.<br />
Conversely, if the <strong>control</strong>ler that minimizes e(t)2 <br />
F for the disturbance process T(z)<br />
with sensor model Gz−1 does not have a realization that fits the structural constraints set<br />
for R(z), the signal e(t) will not be white noise. This can be illustrated by examining the<br />
Markov parameters ofR(z).<br />
Substitution of the update equation in (4.4) into itself allowsu (i)(t) to be written as:<br />
u (i)(t) = <br />
j∈C (i)<br />
+ <br />
a (i,j)(z)<br />
j∈M (i)<br />
⎛<br />
⎝ <br />
m∈C (j)<br />
k (i,j)(z)s (j)(t),<br />
a (j,m)(z)u (m)(t)+ <br />
m∈M (j)<br />
⎞<br />
k (j,m)(z)s (m)(t) ⎠<br />
wherea (i,j)(z) = δij −ã (i,j)(z). Repeated application allowsu (i)(t) to be expressed as:<br />
Nt−1 <br />
u (i)(t) = ξ (i)(t)+<br />
l=0<br />
<br />
j∈F (i,l)<br />
f (i,j,l)s (j)(t−l), (4.7)<br />
where the sets F (i,l) <strong>and</strong> the Markov parametersf (i,j,l) can be expressed recursively as:<br />
⎧<br />
⎨M<br />
(i) for l = 0,<br />
F (i,l) = <br />
⎩ F (j,l−1) for l ≥ 1, (4.8)<br />
j∈C (i)
4.6 Stability 77<br />
⎧<br />
0 for l < 0,<br />
⎪⎨ k (i,j,l) +<br />
f (i,j,l) =<br />
⎪⎩<br />
n−1 <br />
ã (i,p,m)f (p,j,l−m−1) for 0 ≤ l < n,<br />
m=0p∈C<br />
(i)<br />
n−1 <br />
ã (i,p,m)f (p,j,l−m−1) for l ≥ n.<br />
m=0p∈C<br />
(i)<br />
Further,ξ (i)(t) expresses the contribution of past outputsu(tp) fortp ≤ t−Nt. Assuming<br />
thatR(z) is stable, this term goes to zero when the used number of tapsNt increases. This<br />
implies that R(z) can be approximated using (4.7) with ξ (i)(t) = 0 <strong>and</strong> limited Nt, which<br />
forms an FIR description.<br />
Observe from (4.8) that the set F (i,l) of measurements that influenceu (i)(t) grows with the<br />
temporal lagl. Each communication step allows the sensor information to further propagate<br />
over the network of nodes. For the communication structure illustrated in figure 4.2 the<br />
number of <strong>control</strong>ler nodes whose outputs are affected by a measurement initially available<br />
to the central node grows quadratically in time from 9 on the second time step to 16 in the<br />
third <strong>and</strong> so on. In this manner all measurements s (j)(t) can eventually affect all outputs<br />
u (i)(t+d), but with a delaydthat is proportional to the distance between sensorj <strong>and</strong> node<br />
i <strong>and</strong> scaled by the communication radiusrc.<br />
Conversely this means that if the transfer function matrix of the data generating system<br />
in innovation form between the excitation noise e (j)(t) <strong>and</strong> the wavefront phase ˆ φ (i)(t)<br />
has a delay smaller than d, these fast <strong>dynamics</strong> cannot be compensated by the distributed<br />
<strong>control</strong>ler. If the excitation signals e (j)(t) corresponding to the centralized case would be<br />
known, the covariance between e (j)(t) <strong>and</strong> the measurements (m)(t+l) for m ∈ S (i) <strong>and</strong><br />
l = 0...∞ would be indicative of the allowed delay <strong>and</strong> thus the required communication<br />
rangerc.<br />
4.6 Stability<br />
Although it can be safely assumed that the atmospheric disturbance is a stable process, the<br />
distributed open-loop <strong>control</strong>lerR(z) is estimated from measurement data <strong>and</strong> is therefore<br />
not automatically stable even if the independently identified <strong>control</strong>ler nodes are stable.<br />
A constraint is therefore sought that can be enforced on local <strong>control</strong>ler nodes while leading<br />
to stability of the interconnected system of <strong>control</strong>lers. This means that the feedback interconnection<br />
of one <strong>control</strong>ler node with more <strong>and</strong> more other nodes in the network is stable<br />
regardless of the number of interconnected nodes, which implies that the enforced condition<br />
must be retained in the interconnection. This is e.g. not the case when constraining each<br />
<strong>control</strong>ler node to satisfy the small gain theorem (or bounded real lemma [199]), i.e. having<br />
an infinity norm smaller than one. It is well known that the feedback interconnection of two<br />
stable systems H (1) <strong>and</strong> H (2) is stable when H (1)∞ · H (2)∞ < 1. But although the<br />
interconnection of two small gain systems is stable, the infinity norm of the interconnected<br />
system is unknown such that the small gain property is not retained in the interconnection<br />
<strong>and</strong> cannot be used to guarantee stability of a network of <strong>control</strong>lers.<br />
To overcome this, the notion of passivity could be used [8]. Strict passivity implies asymptotic<br />
stability <strong>and</strong> is retained in the interconnection of two strictly passive systems. However,<br />
as shown in appendix B passivity is very restrictive for discrete time systems having a zero<br />
4
4<br />
78 4 Data driven distributed <strong>control</strong><br />
direct feed-through term. Such terms would lead to ill-posedness of the interconnected system.<br />
An approach that is restrictive, but does not suffer these drawbacks is described in the next<br />
subsection.<br />
4.6.1 Gershgorin’s circle theorem<br />
Stability of a discrete time system essentially requires the zeros of its characteristic polynomial<br />
to be located within the unit disc. For state-space systems this characteristic polynomial<br />
is obtained directly from an eigenvalue analysis of its state transition matrix. Computation<br />
of the poles of the interconnected system in (4.4) will here be done by first lifting<br />
its autonomous <strong>dynamics</strong> in the spatial coordinate <strong>and</strong> then considering the block-<strong>control</strong>ler<br />
canonical state-space form [161]. The autonomous part in lifted form can be expressed as:<br />
u(t) = Ã (1)u(t−1)+...+ Ã (n)u(t−n).<br />
where the <strong>control</strong>ler coefficients ã (i,j,l) form the (i,j) elements of the matrices à (l). This<br />
autonomous system can be rewritten into the following block-<strong>control</strong>ler canonical statespace<br />
form:<br />
⎡ ⎤<br />
à (1) ... ... à (n)<br />
⎢ I ... 0 ⎥<br />
x(t+1) = Ax(t) where A = ⎢<br />
⎣<br />
. ⎥<br />
.. ⎦ ,<br />
0 I 0<br />
where the state x(t) = [u(t),...,u(t−n+1)].<br />
Gershgorin’s circle theorem [71] states that each eigenvalue of a matrix A with elements<br />
A [i,j] lies within at least one of the closed discs centered at A [i,i] with radius<br />
r (i) = <br />
i=j |A [i,j]|. Consequently, all eigenvalues of the matrix will lie inside the unit<br />
disc <strong>and</strong> the system is (marginally) stable if the sum of the absolute values in each row of<br />
the matrix A is smaller than or equal to one. For the second <strong>and</strong> higher block-rows this is<br />
automatically guaranteed by the identity matrices, but for the first block-row this leads to<br />
the constraint that for eachi = 1...Nn:<br />
<br />
j∈C (i) k=1<br />
n<br />
|a (i,j,k)| ≤ 1. (4.9)<br />
How restrictive this constraint is in terms of performance depends on the data generating<br />
system, but it further compresses the class of data generating systems derived in section 4.5.<br />
Equation (4.9) gives rise to an independent set of constraints for each nodei that involves the<br />
sum of absolute values. A constraint of this form|a|+|b| < 1 with two parameters can be<br />
written into 4 simultaneous linear constraints: a+b < 1∩−a+b < 1∩a−b < 1∩−a−b < 1.<br />
One constraint is required for each binomial possibility of parameter signs such that the<br />
total number of constraints can be expressed as N (c,i) = 2 M (a,i), where the numberM (a,i)<br />
of involved parametersa (i,j,k) is equal to n times the cardinality of C (i). As will be shown<br />
at the end of the next section, this constraint concept leads to the identification of the local<br />
<strong>control</strong>ler parameters via local quadratic programming problems.
4.7 Identification procedure 79<br />
4.7 Identification procedure<br />
4.7.1 Optimization criterion <strong>and</strong> approach<br />
With the <strong>control</strong>ler structure defined, the values of the unknown coefficients need to be<br />
determined. Let the unknown coefficients of the distributed <strong>control</strong>ler R(z) be estimated<br />
from open-loop measurement data s(t) fort = 0...N −1 as the minimizer of:<br />
Ns e<br />
J = (j)(t) 2 F<br />
j=1<br />
N−1<br />
t0<br />
. (4.10)<br />
This cost function is used both for identification <strong>and</strong> validation. Recall that e (j)(t) is<br />
defined in (4.3) <strong>and</strong> u (S(j)) in (4.4). In contrast to the traditional cost function described in<br />
section 3.6 this cost function weights the error in terms of WFS gradients <strong>and</strong> not in terms<br />
of wavefront phase. The latter would require an explicit model of the wavefront phase<br />
disturbance, which is outside the scope of the data driven approach presented here. For<br />
the purpose of identification the time t0 will be chosen as small as possible as determined<br />
by the identification algorithm. For validation it will be chosen large enough to prevent<br />
weighting of initial transient behavior.<br />
When taking into account the update law in (4.4), the definition of e (i)(t) contains a<br />
product between the unknown signal u (S(j))(t) <strong>and</strong> the unknown coefficients ã (i,j,l). This<br />
renders to optimization problem nonlinear in its unknowns <strong>and</strong> the coefficients that minimize<br />
(4.10) cannot be expressed explicitly. However, this is a common problem <strong>and</strong> methods<br />
have been developed to deal with it. Such methods are usually iterative <strong>and</strong> based on<br />
gradient search concepts [119]. However, the large number of unknown coefficients makes<br />
calculation of the Jacobian <strong>and</strong> Hessian terms computationally costly operations. This is<br />
also the case for simulations that must be performed for the intermediate <strong>control</strong>lers on the<br />
measurement data to evaluate the cost function. Since these <strong>control</strong>lers may be unstable,<br />
computationally expensive stability checks must be performed <strong>and</strong> – if necessary – the <strong>control</strong>ler<br />
must be stabilized within its structural constraints. Combined with the experience of<br />
the author that convergence is very slow <strong>and</strong> highly sensitive to initial estimates, an alternative<br />
approach is proposed. Similar to the gradient search algorithm, this algorithm is to<br />
be performed off-line on a global data set <strong>and</strong> is itself not distributed. Only the coefficients<br />
estimated by the algorithm form a distributed <strong>control</strong>ler.<br />
4.7.2 A two-stage approach<br />
An off-line procedure will now be proposed in which the outputsu (i)(t) <strong>and</strong> the coefficients<br />
ã (i,j,l) <strong>and</strong>k (i,u,l) are estimated sequentially. An estimate ofu (i)(t) can be obtained by first<br />
estimating the coefficients of the <strong>control</strong>ler’s FIR approximation described in section 4.5 <strong>and</strong><br />
subsequently applying this FIR filter to the measurements. This signal cannot be reproduced<br />
by the distributed <strong>control</strong>ler using the FIR coefficients, because – as was shown in section<br />
4.5 – these do not satisfy the distributed structure. The second identification step therefore<br />
consists of the identification of the coefficients ã (i,j,l) <strong>and</strong> k (i,u,l) that parameterize the<br />
distributed open-loop <strong>control</strong>ler R (i)(z) defined in (4.4) through linear regression on the<br />
locally available measurements <strong>and</strong> the output estimates of the first step.<br />
4
4<br />
80 4 Data driven distributed <strong>control</strong><br />
Stage 1: estimatingu (i)(t)<br />
By substituting (4.3) into the cost function J in (4.10) <strong>and</strong> replacing u (S(j))(t) with the<br />
output of the FIR approximation ofR(z) in (4.7), the following cost function is obtained:<br />
Ns <br />
J1 =<br />
where ũ (i)(t) =<br />
j=1<br />
Nt−1 <br />
l=0<br />
s(j)(t)−GFũ <br />
<br />
(S(j))(t−1)<br />
<br />
p∈F (i,l)<br />
f (i,p,l)s (p)(t−l).<br />
2<br />
N−1 F<br />
Nt<br />
, (4.11)<br />
This can be rewritten into the spatially lifted form for the interconnected system as:<br />
<br />
<br />
J1 = s(t)−G ⎡ ⎤<br />
s(t−1) <br />
2<br />
N−1 ⎢ ⎥<br />
F (0) ... F (Nt−1) ⎣<br />
<br />
. ⎦<br />
,<br />
<br />
Ψ s(t−Nt) F Nt<br />
<br />
s(t−1)<br />
<br />
<br />
<br />
= s(Nt) ... s(N −1)<br />
<br />
<br />
−GΨ<br />
<br />
S<br />
s(Nt −1) ... s(N −2) <br />
<br />
2<br />
<br />
. (4.12)<br />
<br />
F<br />
S<br />
where the only non-zero elements (i,j) of the matrices F (l) for i = 1...Nn <strong>and</strong> j ∈<br />
F (i,l) are equal to the FIR coefficients f (i,j,l). Using the vectorization operator ”vec” <strong>and</strong><br />
Kronecker identities from [23], the cost function in (4.12) can be rewritten to:<br />
J1 = S−GΨS 2 , (4.13a)<br />
F<br />
<br />
<br />
= vec(S)− S T <br />
<br />
⊗G vec(Ψ) 2<br />
, (4.13b)<br />
F<br />
= S 2 2<br />
F − vec<br />
Ne<br />
T (G T SS T )vec(Ψ)+ 1<br />
vec<br />
Ne<br />
T <br />
(Ψ) SS T ⊗G T <br />
G vec(Ψ),<br />
<br />
Q<br />
whereNe = 2Ns(N −Nt). For the case when the communication radiusrc is large <strong>and</strong> a<br />
centralized <strong>control</strong>ler is sought, all elements ofΨ<strong>and</strong> thus vec(Ψ) are free to be estimated.<br />
The Ψ that minimizes J1 can then be expressed in the following form that does no longer<br />
contain a Kronecker product:<br />
ˆΨ = (G T G) −1 G T <br />
T<br />
SS SS T−1 .<br />
Although it can be assumed that then loop measurement signal s(t) is persistently exciting<br />
<strong>and</strong>SS T has full rank, the inverse(G T G) −1 does not exist due to the singularity ofG.<br />
Without restricting the analysis to the centralized case, the solution can become ill defined<br />
due to the pre-multiplication by G in (4.13a). Consequently, any columnΨ·i of Ψ that lies<br />
in the kernel of G does not influence the cost functionJ1. As discussed in section 4.3, the<br />
kernel ofGconsists of the unseen piston <strong>and</strong> waffle modes denoted by the vectorsmp <strong>and</strong><br />
mw respectively. These vectors can be linearly transformed into two orthogonal vectors
4.7 Identification procedure 81<br />
m (1) <strong>and</strong> m (2) consisting of only ones <strong>and</strong> zeros that are complementary <strong>and</strong> divide the<br />
phase grid into two separate sub-grids (figure 1.6, [143]). Sub-grid i consist of the points<br />
corresponding to the non-zero values in m (i). Assuming a rectangular Fried geometry,<br />
these are such that all diagonally adjacent points belong to the same sub-grid, as indicated<br />
by the black <strong>and</strong> white backgrounds in figure 1.6. The solution ˆ Ψ thus becomes ill defined<br />
if its sparsity structure allows a termαm (1) orαm (2) forα = 0 to be added to any column<br />
Ψ·i. Since the sub-grids corresponding to m (1) <strong>and</strong> m (2) are spread over the entire spatial<br />
domain, this occurs only if the temporal lag l is such that the effect of a measurement<br />
s (j)(t−l) has spatially propagated to all outputsu (i)(t) for i = 1...Nn. This means that<br />
Markov matricesF (l) of which all elements are allowed to be non-zero cannot be uniquely<br />
identified regardless of persistence of excitation of the data generating system.<br />
Hard constraints can be applied for the elements of Ψ that must be zero by removing<br />
rows <strong>and</strong> columns from the vectors <strong>and</strong> matrices in (4.13b), but this leads to large matrices<br />
with non-trivial structures. The latter becomes even more stringent when the solution is illdefined<br />
<strong>and</strong> e.g. a pseudo-inverse technique is needed to find one. Instead, soft constraints<br />
can be used, in which case the elements of Ψ that must be zero are explicitly weighted.<br />
These terms are then allowed to be small instead of zero, with their magnitude depending<br />
on the weights used in combination with the correlations present in the measurement signal<br />
s(t). This leads to a new cost function:<br />
˜J1 = J1 + 1<br />
= S 2 F<br />
Ne<br />
vec T (Ψ)Wdvec(Ψ), (4.14)<br />
2<br />
− vec<br />
Ne<br />
T (G T SS T )vec(Ψ)+ 1<br />
vec<br />
Ne<br />
T (Ψ)(Q+Wd) vec(Ψ),<br />
whereWd = diag(vec(W)) is a diagonal weighting matrix, whereW has the same size as<br />
Ψ <strong>and</strong> its elements are only nonzero, positive when the corresponding element inΨis zero.<br />
The solution vec( ˆ Ψ) can then be solved from:<br />
<br />
(Q+W) vec(Ψ) = vec G T <br />
SS , (4.15)<br />
after which the soft-constrained elements that must be zero are forced to zero. By using a<br />
Krylov subspace conjugate gradients solver [77], a pseudo-inverse solution to (4.15) can be<br />
obtained for which the ill defined terms ofΨare constrained to zero.<br />
In this algorithm,Q <strong>and</strong> Wd only occur in product with a residual vectorrcg = vec(Rcg).<br />
When using the definitions of ofWd <strong>and</strong>Q(the latter in (4.13)), these products reduce to:<br />
Qrcg = (SS T ⊗G T G)rcg,<br />
= vec(G T GRcgSS T )<br />
<strong>and</strong><br />
Wdrcg = diag(vec(W))rcg<br />
,<br />
= vec(W◦Rcg)<br />
where ◦ denotes the Hadamard (element-wise) product. The second reduction steps are<br />
based on Kronecker <strong>and</strong> Hadamard identities from [23] <strong>and</strong> [115] respectively. Since these<br />
expressions no longer contain the Kronecker product, this means a large reduction in memory<br />
requirement. Further, as it is assumed thats(t) is generated by an wide sense stationary<br />
stochastic process,SS T should have a block-Toeplitz structure that can be exploited for further<br />
storage reduction. However, even with these optimizations the number of computations<br />
4
4<br />
82 4 Data driven distributed <strong>control</strong><br />
required to solve ˆ Ψ from (4.15) increases to the third power in the number of nodes Nn.<br />
Further reductions will require future research, but this procedure is performed off-line <strong>and</strong><br />
is not time critical.<br />
The so obtained set of estimates for the Markov coefficients differs from the sought set<br />
that minimizes the cost function J1 for two reasons. Firstly, they were derived based on<br />
the modified, soft constraint cost function ˜ J1. Depending on the correlations in the input<br />
data series <strong>and</strong> the chosen weighting matrix Wd, this leads to a different solution that approximates<br />
that for the original cost functionJ1 as Wd increases. Secondly, the conjugate<br />
gradient algorithm leads to an estimation error. Although in theory the algorithm converges<br />
on the correct solution afterNq iterations, whereNq is the number of unknowns, numerical<br />
round-off generally leads to significant error <strong>and</strong> a stopping criterion terminates the iterative<br />
process beforeNq cycles are reached. When iterations are stopped when the Frobenius<br />
norm of the residual vector has been reduced by a certain factor1/ǫ (see section 3.6.3), the<br />
second error can be limited by choosingǫ sufficiently small. However, the total estimation<br />
error <strong>and</strong> its effect on the eventual performance of the model in terms of prediction error<br />
are not further investigated in this report due to the high computational cost involved in<br />
obtaining the exact, hard-constrained solution for systems of relevant dimensions.<br />
Now the estimates for the Markov parametersf (i,j,l) fori = 1...Nn,j ∈ F (i,l) <strong>and</strong>l =<br />
0...Nt−1 from ˆ Ψ can be substituted into the FIR approximation in (4.7). When neglecting<br />
the termsξ (i)(t), estimatesû (i)(t) for the output trajectoriesu (i)(t) can be obtained as:<br />
û (i)(t) =<br />
Nt−1 <br />
l=0<br />
<br />
j∈F (i,l)<br />
Stage 2: estimating the <strong>control</strong>ler coefficients<br />
ˆ f(i,j,l)s (j)(t−l). (4.16)<br />
In the second stage, the unknown coefficientsã (i,j,l) <strong>and</strong>k (i,u,l) of the distributed <strong>control</strong>ler<br />
as in (4.4) can be estimated. The estimates will be obtained by using the trajectory estimates<br />
û (i)(t) <strong>and</strong> solving independent linear regression problems fori = 1...Nn.<br />
Letr (i)(t) be the equation error of (4.4) where the optimal output trajectoryu (i)(t) has been<br />
replaced by the estimate û(t):<br />
r (i)(t) = <br />
ã (i,j)(z)u (j)(t)− <br />
k (i,j)(z)s (j)(t).<br />
j∈C (i)<br />
This can be rewritten using matrix notation as:<br />
j∈M (i)<br />
r (i)(t) = û (i)(t)−θ T (i) υ (i)(t), (4.17)<br />
where the column vector υ (i)(t) stacks all û (j)(t − l) <strong>and</strong> s (m)(t − l + 1) for j ∈ C (i),<br />
m ∈ M (i) <strong>and</strong> l = 1...n. The column vector θ (i) stacks the corresponding unknown<br />
<strong>control</strong>ler coefficientsã (i,j,l) <strong>and</strong>k (i,m,l).<br />
Let the estimate ˆ θ (i) of the coefficient vectorθ (i) now be defined as the argument that minimizes<br />
the cost functionJ 2,(i):<br />
J 2,(i) = r (i) 2 F<br />
N−1<br />
t0<br />
, (4.18)
4.7 Identification procedure 83<br />
<br />
<br />
<br />
= û(i)(t0) ... û<br />
<br />
(i)(N −1) <br />
<br />
Û (i)<br />
−θ T (i)<br />
<br />
υ(i)(t0 −1) ... υ (i)(N −2) <br />
<br />
<br />
<br />
<br />
<br />
Υ (i)<br />
2<br />
F<br />
, (4.19)<br />
where the second step follows by substitution of (4.17). The solution ˆ θ (i) can now be expressed<br />
in the well-known form:<br />
ˆθ (i) =<br />
<br />
Υ (i)Υ T −1Û(i)Υ (i) (i). (4.20)<br />
The coefficients a (i,j,l) of the closed-loop <strong>control</strong>ler C(z) can be calculated from ˆã (i,j,l)<br />
<strong>and</strong> ˆ k (i,u,l) by reverse application of (4.5) on page 75.<br />
Finally, note that Υ (i) stacks only those output <strong>and</strong> measurement signals to which node i<br />
has direct access at each sampling time. Therefore, each solution ˆ θ (i) can be obtained using<br />
only the information available to node i, making the second identification stage a set of<br />
independent, distributed operations. The number of computations required to solve each<br />
local problem depends only on the <strong>control</strong>ler ordern<strong>and</strong> the communication radiusrc <strong>and</strong><br />
not on the number of nodesNn. Although this suggests possibilities for on-line estimation,<br />
this has no practical relevance due to the centralized nature of the first estimation step.<br />
Enforcing stability<br />
As discussed in section 4.6, stability of the network of open-loop <strong>control</strong>lers – which<br />
guarantees stability of the closed-loop provided that the plant model is correct – is not trivial<br />
even if all individual <strong>control</strong>ler nodes are stable. Therefore, the above described two-stage<br />
identification algorithm will be extended to comprehend two different approaches to enforce<br />
stability. Besides the stability constraint based on Gershgorin’s circle theorem discussed in<br />
section 4.6.1, a second method will be employed to enforce stability. By adding explicit<br />
weights to the elements of the coefficient matrices à (k) in the second identification stage,<br />
the pole locations of the identified system can be influenced. This is done by augmenting<br />
the cost-function J 2,(i) in (4.19) with the regularization term θ T (i) W′ (i) θ (i), where W ′ (i) is<br />
a diagonal weighting matrix of which only the diagonal elements that correspond to the<br />
unknown coefficients ã (i,j,l) are non-zero. When all non-zero weights are chosen equal, a<br />
line search method can be used to find the weight for which the poles are contained within<br />
the unit circle. This means that in contrast to the Gershgorin method, global stability is<br />
not guaranteed after local optimization, which renders this approach unfit for the future<br />
development of an on-line adaptive distributed <strong>control</strong> law. Moreover, both the second<br />
identification stage <strong>and</strong> the stability check become computationally costly for large systems.<br />
The principle based on Gershgorin’s circle theorem as proposed in 4.6.1 is applied as<br />
follows. For eachi = 1...Nn the linear constraints in (4.9) together with the cost function<br />
J 2,(i) in (4.19) lead to a Quadratic Programming (QP) problem that can be expressed as:<br />
ˆθ (i) =argmin<br />
θ (i)<br />
subject to<br />
θ T (i) Υ (i)Υ T (i)θ (i) −2 Û (i)Υ T (i)θ (i)<br />
Ξ(i) 0 θ (i) < 1−λ,<br />
(4.21)<br />
4
4<br />
84 4 Data driven distributed <strong>control</strong><br />
where Ξ (i) =∈ R2M (a,i)<br />
×M(a,i) is a binomial matrix consisting of the sign-values ±1. Its<br />
rows form all possible signs for the parametersa (i,j,k) inθ (i) that can be obtained by taking<br />
M (a,i) values from the set {−1,1} with replacement.<br />
The QP problems are solved using Matlab’sÕÙÔÖÓ[130] function <strong>and</strong> – in principle<br />
– must be performed only once. However, for λ = 0 the stability constraint allows poles<br />
arbitrarily close to the unit circle, which in practice may lead to marginally stable systems.<br />
The value of λ that yields the best performance in terms of (4.10) varies per data set <strong>and</strong> is<br />
again found using a line search.<br />
4.7.3 Algorithm summary<br />
The derivation of the closed-loop distributed <strong>control</strong>ler C(z) can be summarized into the<br />
following steps:<br />
1. Start with a set of open-loop WFS measurement data of the wavefront disturbance.<br />
2. Estimate the Markov parametersf (i,p,l) of the distributed <strong>control</strong>lerR(z) as the minimizers<br />
of the cost function (4.14). A conjugate gradient solver is proposed to solve<br />
the resulting system of equations in (4.15) for vec(Ψ). Since the cost-function is<br />
based on soft constraints, Markov parameters that are required to be zero will only<br />
be arbitrarily small <strong>and</strong> must be fixed to zero after conjugate gradient iteration has<br />
converged.<br />
3. Substitute the identified Markov parameters into (4.16) to obtain an estimate for the<br />
output trajectoriesu (i)(t) of the open-loop <strong>control</strong>ler nodesR (i)(z).<br />
4. Estimate the coefficients ã (i,j,l) <strong>and</strong> k (i,j,l) of R(z) as the minimizers of the cost<br />
function (4.18). Two approaches are proposed to enforce stability ofR(z):<br />
• Add regularization terms of the formθ T (i) W′ (i) θ (i) to the cost function hence to<br />
the inverted expression in the explicit solution in (4.20).<br />
• Apply linear constraints based on Gershgorin’s circle theorem (section 4.6.1)<br />
<strong>and</strong> solve the resulting QP problem in (4.21).<br />
5. Obtain the remaining coefficientsa (i,j) of the closed-loop distributed <strong>control</strong>lerC(z)<br />
from (4.2) using (4.5).<br />
4.7.4 Discussion<br />
Unseen modes <strong>and</strong> bias errors<br />
As mentioned in section 4.7, the rank deficiency of G is the reason that only a limited<br />
number of Markov parameters F (k) can be uniquely estimated in the first identification<br />
step. Only these coefficients can yield a non-zero contribution to the unseen modes in û(t)<br />
as for the others this is constrained to zero. For the other coefficients, this contribution is<br />
undefined <strong>and</strong> fixed to zero when solving the estimation problem.<br />
In the centralized case when even the first Markov parameter is not uniquely defined, this<br />
means that the time series in û(t) are not independent, rendering the second identification<br />
step an ill posed problem. A solution can still be obtained using a pseudo-inverse or similar<br />
to [100] an orthogonal projection onû(t) can be applied to remove the dependent signals.<br />
In the distributed case, the ill defined parts of F (l) imply an unknown error on û(t) <strong>and</strong>
4.8 Simulation <strong>and</strong> breadboard results 85<br />
therefore an unknown bias on the estimates ˆã (i,j,l) <strong>and</strong> ˆ k (i,u,l). On the other h<strong>and</strong>, for<br />
a fixed communication radius rc, the number of Markov parameters that contain no illdefined<br />
columns increases with the number of actuators Na. Assuming that the <strong>dynamics</strong><br />
underlying the unseen modes are stable <strong>and</strong> have a finite impulse response, the bias will<br />
therefore decrease for increasing Na. Alternatively, physical models of the atmospheric<br />
wavefront disturbance can be used in the first identification step. For instance, the temporal<br />
spectra of the unseen modes have been modeled based on Kolmogorov statistics <strong>and</strong> the<br />
frozen flow assumption in [33] <strong>and</strong> can be used as additional weighting terms in the cost<br />
functionJ1.<br />
Further, the estimation results will be biased because only a limited numberNt of Markov<br />
parameters is estimated <strong>and</strong> the cost functions in (4.11) <strong>and</strong> (4.19) represent only finite<br />
sample approximations of the actual variances. However, these effects can be limited by<br />
choosing the number of coefficient matrices Nt <strong>and</strong> the number of samples N sufficiently<br />
large.<br />
Order selection<br />
Since neither the true order of the data generating system T(z) is known nor its spatiotemporal<br />
behavior, the choice for Nt <strong>and</strong> n is not straightforward <strong>and</strong> may depend on the<br />
chosen communication radiusrc.<br />
The number of tapsNt must be chosen such that the estimateû(t) is consistent. In practice,<br />
it can be obtained by gradually increasing it until this yields only a marginal decrease of the<br />
cost function J1. The number of states per node n can be chosen using the same approach<br />
on the cost functionsJ 2,(i).<br />
4.8 Simulation <strong>and</strong> breadboard results<br />
In this section, the results of application of the described identification algorithm to several<br />
sets of measurements will be presented. A first set of measurements has been obtained from<br />
an AO breadboard at TNO Science <strong>and</strong> Industry, which is depicted in figure 4.5. Although<br />
the results on this data demonstrate the validity of the described approach, the number of<br />
spots of the wavefront sensor is too limited to show the effect of scaling up the system<br />
dimensions. Therefore, a second data set with a larger number of sensor spots has been<br />
artificially created.<br />
4.8.1 Performance measures<br />
The performance of the <strong>control</strong>lers resulting from both identification stages will be measured<br />
as a relative prediction error in terms of the Variance Accounted For (VAF). For the<br />
first stage this will be defined as:<br />
<br />
VAF1 =<br />
J1<br />
1−<br />
〈sTs〉 N−1<br />
<br />
Nt<br />
·100%, (4.22)<br />
where J1 is the cost function defined in (4.11), but evaluated on a separate validation data<br />
set. For the second stage, the signal ê(t) will be obtained after application of the identified<br />
4
4<br />
86 4 Data driven distributed <strong>control</strong><br />
ARMA <strong>control</strong>ler on the validation data set <strong>and</strong> used to evaluate:<br />
⎛ <br />
T N−1<br />
ê ê t0<br />
VAF2 = ⎝1−<br />
〈sTs〉 N−1<br />
⎞<br />
⎠·100%,<br />
t0<br />
wheret0 is chosen large enough to skip the transient errors.<br />
Further, the results will be compared to that of the baseline strategy, which is defined in<br />
section 4.7.1 as the static <strong>control</strong>ler R(z) = G # . Substitution into (4.3) gives the corresponding<br />
prediction error signalê0(t) as:<br />
ê0(t) = s(t)−GG # s(t−1).<br />
Let the corresponding VAF0 value then be defined as:<br />
4.8.2 Breadboard data<br />
<br />
T N−1<br />
ê0ê0 1<br />
VAF0 = 1−<br />
〈sTs〉 N−1<br />
<br />
·100%.<br />
1<br />
The breadboard (figure 4.5) contains a 630nm laser source whose light is fed through a pinhole<br />
to simulate a point source. A h<strong>and</strong>-polished, glass disc with a 110mm radius is used as<br />
a turbulence simulator. Its spatial characteristics resemble Kolmogorov turbulence with an<br />
intensity characteristicDt/r0 = 5, whereDt denotes the diameter of the telescope aperture<br />
<strong>and</strong>r0 the Fried constant [98]. In order to simulate the temporal behavior of the wavefront<br />
disturbance, the disc rotates through the laser beam, whose diameter at this point is 10mm.<br />
After several reflections, the beam reaches a wavefront sensor consisting of an OKO-Tech,<br />
hexagonal Hartmann array with 127 lenselets <strong>and</strong> an SVS-Vistek CCD camera. As discussed<br />
in chapter 2, a compression factor 16 is foreseen from the telescope aperture onto<br />
a DM surface, such that a 6mm actuator spacing of a DM becomes approximately 96mm<br />
when projected on the telescope’s aperture. When assuming the number of WFS spots to be<br />
equal to the number of actuators – in this case 127 – this means that the telescope diameter<br />
Dt corresponding to the setup is approximately 11·96mm≈ 1m, where 11 is the number<br />
of sensor spots over the diameter.<br />
The disturbance translates over the aperture as a pure frozen flow <strong>and</strong> the rotation speed of<br />
the turbulence simulator can be varied to simulate different wind speeds. These wind speeds<br />
can be related to Greenwood frequencies using the approximation in (2.5). However, since<br />
rotation speed <strong>and</strong> sampling time have the same effect (except for measurement noise <strong>and</strong><br />
motion blur effects), let the temporal behavior of the disturbance be characterized using the<br />
ratio between Greenwood frequencyfG <strong>and</strong> sampling frequencyfs. For the obtained measurements,<br />
this ratio lies in the range 0.01 < fG/fs < 0.41. For an AO system for an 8m<br />
telescope with a sampling frequency of 1kHz, this corresponds to wind speeds between ca.<br />
5 <strong>and</strong> 200m/s.<br />
Identification of the FIR coefficients is done on a set of gradient measurements containing<br />
10000 samples, whereas the performance is evaluated on a separate validation set of 1000<br />
samples. Figure 4.6 shows VAF1 against Nt for various communication radii rc. It can<br />
be observed that even for the minimal communication radius, which allows communication
4.8 Simulation <strong>and</strong> breadboard results 87<br />
DM<br />
Laser<br />
TT<br />
SC<br />
WFS<br />
Figure 4.5: The AO breadboard setup at TNO Science <strong>and</strong> Industry. SC is the science camera, TS<br />
the turbulence simulator <strong>and</strong> TT a tip/tilt steering <strong>mirror</strong>. The picture is courtesy of K.<br />
Hinnen.<br />
Variance accounted for [%]<br />
100<br />
95<br />
90<br />
85<br />
80<br />
75<br />
70<br />
65<br />
r c = 1<br />
r c = 1.5<br />
r c = 2<br />
r c = 2.5<br />
60<br />
1 3 5 7 9 11 13<br />
Baseline<br />
15<br />
Number of taps N [−]<br />
t<br />
TS<br />
Figure 4.6: Performance of<br />
the structured FIR approximation:<br />
VAF1 from (4.22) as a<br />
function of the number of taps<br />
Nt on a breadboard data set<br />
with a Greenwood to sampling<br />
frequency ratiofG/fs ≈ 0.3.<br />
with only the four directly adjacent nodes, a better phase prediction is obtained than for the<br />
centralized baseline strategy. On the other h<strong>and</strong>, with only slightly more than 4 rings or in<br />
total 69 illuminated sensor spots, the distributed <strong>control</strong>ler becomes centralized forrc = 11.<br />
As restriction of the radius rc causes the distributed <strong>control</strong>ler to trade spatial correlations<br />
for temporal ones, a decrease inrc can be observed to yield an increase in the relevant number<br />
of tapsNt to achieve a high VAF1 value. However, this restriction has little effect on the<br />
maximum VAF value achieved for a high numberNt of taps.<br />
Figure 4.7 shows the performance results of the distributed <strong>control</strong>ler obtained after the<br />
second identification step for rc ∈ {1,1.5,2} <strong>and</strong> n = 1...5 obtained by evaluating the<br />
measure VAF2 on the validation data set. Stability was enforced using both the regularization<br />
method described in section 4.7.2 <strong>and</strong> using the Gershgorin approach of section 4.6.1.<br />
4
4<br />
88 4 Data driven distributed <strong>control</strong><br />
Figure 4.7: Performance of<br />
the distributed <strong>control</strong>ler:<br />
VAF2 as a function of the<br />
local <strong>control</strong>ler order n on a<br />
breadboard data set with a<br />
Greenwood to sampling frequency<br />
ratio of fG/fs ≈ 0.3.<br />
The VAF-value of the Gershgorin<br />
approach for rc = 1<br />
varies between 60 <strong>and</strong> 70%.<br />
Variance accounted for [%]<br />
100<br />
95<br />
90<br />
r c = 1.0 (Regularization)<br />
r c = 1.0 (Gershgorin)<br />
r c = 1.5 (Regularization)<br />
r c = 1.5 (Gershgorin)<br />
r c = 2.0 (Regularization)<br />
r c = 2.0 (Gershgorin)<br />
85<br />
1 2 3<br />
Filter order n [−]<br />
Baseline<br />
4 5<br />
The performance obtained for the first approach is slightly better than for the Gershgorin<br />
approach, but this difference is small. For both approaches the performance increases with<br />
the communication radiusrc <strong>and</strong> the <strong>control</strong>ler ordern, but for the latter only up to n ≈ 5.<br />
The performance is worse than that obtained using the structured FIR coefficients, which<br />
shows that the proposed combination of the ARMA <strong>control</strong>ler structure <strong>and</strong> the two-step<br />
identification method does not yield a <strong>control</strong>ler that is able to fully exploit the available<br />
spatio-temporal correlation. Nevertheless, even for rc = 1 <strong>and</strong> n = 1 the performance<br />
of the identified distributed <strong>control</strong>lers exceeds that of the baseline method. By increasing<br />
rc, the performance gradually approaches that of the centralized case, which corresponds<br />
to the distributed system for rc ≥ 11. Therefore, the radius rc serves as a means to make<br />
a trade-off between the costs of computation <strong>and</strong> communication hardware <strong>and</strong> achievable<br />
performance.<br />
4.8.3 Artificial data set<br />
To show the effect of increasing the number of Shack-Hartmann spots on the performance<br />
of the distributed <strong>control</strong>ler, a data set is required with a higher number of sensor spots.<br />
This data set has been artificially created. First, a static phase screen with Kolmogorov<br />
spatial statistics was generated using a midpoint displacement algorithm [95]. Then this<br />
was interpolated over a square aperture window that was translated over the phase screen<br />
with a speed corresponding to fG/fs ≈ 0.24, thus simulating a frozen flow <strong>and</strong> yielding<br />
φ(t). The measurement data set s(t) was then obtained through the sensor model in (4.1)<br />
where the variance of the measurement noise signalw(t) was chosen according to a signal<br />
to noise ratio of 20dB.<br />
The performance of the identified <strong>control</strong>ler is again evaluated using VAF1 <strong>and</strong> VAF2 <strong>and</strong><br />
performed on a different data set. For all data sets, a communication radius rc = 2 was<br />
used <strong>and</strong> the number of coefficient matrices was chosenn = 2. Results are plotted in figure<br />
4.8 for grid sizes between 3×3 <strong>and</strong>19×19 sensor spots. Note here that in contrast to the
4.9 Conclusions <strong>and</strong> future work 89<br />
Variance accounted for [%]<br />
100<br />
90<br />
80<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
0<br />
FIR approximation<br />
Distributed ARMA (Regularization)<br />
Distributed ARMA (Gershgorin)<br />
Baseline<br />
50 100 150 200 250 300 350 400<br />
Number of sensor spots [−]<br />
Figure 4.8: The measures<br />
VAF1 <strong>and</strong> VAF2 for both identification<br />
steps as a function of<br />
Ns, the number of points in<br />
the sensor grid, using rc = 2<br />
<strong>and</strong> n = 2.<br />
previous results, the performance obtained from the Gershgorin approach is slightly better<br />
than that of the regularization approach. For the centralized case the VAF-value is expected<br />
to increase, since the relative contribution of the badly predictable part of the disturbance<br />
entering the aperture at one edge decreases. This is still the case for the FIR approximation<br />
of the distributed <strong>control</strong>ler, but the performance of the distributed ARMA <strong>control</strong>lers shows<br />
a slight decrease when the number of sensor spotsNs increases. This decrease becomes less<br />
steep asNs increases.<br />
4.9 Conclusions <strong>and</strong> future work<br />
Due to the working principle of the Shack-Hartmann wavefront sensor, wavefront reconstruction<br />
forms a critical part of the AO <strong>control</strong> system. As shown in the previous chapter<br />
this poses a serious challenge for the design of a distributed <strong>control</strong>ler, because local<br />
<strong>control</strong>lers have access to only a limited number of measurements. In this chapter some<br />
ideas of the previous chapter were extended towards a more general distributed <strong>control</strong>ler<br />
parametrization in the context of a closed-loop <strong>control</strong> setting in which the WFS observes<br />
the effect of the <strong>control</strong> actions. Further, an idealized DM is considered that is modeled as<br />
an identity transfer matrix <strong>and</strong> the WFS is of the Shack-Hartmann type with one sample<br />
delay <strong>and</strong> white measurement noise. The approach towards a distributed <strong>control</strong>ler is driven<br />
by the desired structure <strong>and</strong> aimed at determining the performance achievable given this<br />
structure to determine the feasibility of distributed <strong>control</strong> for AO.<br />
This structure was defined as a network of output interconnected ARMA <strong>control</strong>lers,<br />
where the connections are chosen based on a communication radius. It was shown that by<br />
assuming this parametrization it is implicitly assumed that the structure of the disturbance<br />
generating system in innovations form is also a network of output interconnected ARMA<br />
systems.<br />
A two-stage approach has been proposed to identify the unknown <strong>control</strong>ler parameters in<br />
which the estimation problem is split into two linear regression problems. In the second<br />
step, stability is enforced using two alternative approaches: regularization <strong>and</strong> constrained<br />
4
4<br />
90 4 Data driven distributed <strong>control</strong><br />
optimization based on Gershgorin’s circle theorem. Both identification steps need to be<br />
solved off-line <strong>and</strong> the first is a centralized instead of a distributed operation.<br />
Results were presented both on measurement data obtained from an optical breadboard<br />
<strong>and</strong> on synthetic data. Little performance difference is found between the Gershgorin<br />
<strong>and</strong> regularization methods to enforce stability. The performance of the distributed<br />
<strong>control</strong>ler is shown to depend on the chosen communication radius, but even for very<br />
small communication radii it is found to exceed that of the baseline strategy. The latter<br />
is defined as centralized reconstruction of the latest vector of measurements using the<br />
pseudo-inverse of the Fried geometry matrix G. Further, the performance decays slightly<br />
with the number of sensor <strong>and</strong> actuator channels, which is not the case for a structured<br />
FIR approximation, which indicates that this is not a fundamental limitation of the distributed<br />
architecture. Future research is required to analyze such scaling properties in more<br />
detail, preferably using wavefront disturbance measurements from an actual large telescope.<br />
In this chapter the DM system has been assumed to be an ideal corrector by neglecting<br />
both its temporal <strong>and</strong> spatial <strong>dynamics</strong> <strong>and</strong> letting its influence on the wavefront phase<br />
be equal to the actuator comm<strong>and</strong>s. This is not the case for the DM system developed<br />
within this project, which means that the presented algorithm cannot be directly applied.<br />
However, in future work it can be investigated whether the principle of exploiting the local<br />
spatio-temporal <strong>dynamics</strong> of the disturbance signal to perform a global operation such as<br />
wavefront reconstruction can also be applied to the servo problem of calculating the optimal<br />
DM comm<strong>and</strong> vector for the predicted wavefront shape. In anticipation of this, the components<br />
of the DM introduced in chapter 2 are described <strong>and</strong> modeled in the next chapters.<br />
This leads to a complete spatio-temporal model of the DM system that is suitable both for<br />
verification purposes as well as for the design of a (distributed) <strong>control</strong>ler.
ÔØÖÚ<br />
ÌÚÖÐÖÐÙØÒØÙØÓÖ<br />
The design of the electromagnetic reluctance type actuator will be introduced.<br />
The actuator converts a current through a coil into a mechanical force – <strong>and</strong><br />
thus a deflection – by means of a magnetic flux that varies with the deflection.<br />
A mathematical model will be derived that describes this relation based on<br />
equations from the magnetic, mechanical <strong>and</strong> electronic domains. A measurement<br />
setup to test a prototype actuator is presented <strong>and</strong> measurement results<br />
are shown. The derived model structure is validated <strong>and</strong> model parameters are<br />
estimated from these results <strong>and</strong> compared to the model. Differences are analyzed<br />
based on a sensitivity analysis, leading to minor design changes before<br />
manufacturing grid modules of 61 actuators.<br />
The design of these modules is then presented <strong>and</strong> a second measurement setup<br />
is described <strong>and</strong> measurement results are shown. From these results the actuator<br />
properties are again identified <strong>and</strong> their statistical spread is analyzed. The<br />
actuators in seven modules are found to behave in accordance with the derived<br />
model <strong>and</strong> within the specifications set in chapter 2. This proves that the manufacturing<br />
<strong>and</strong> assembly process is robust <strong>and</strong> allows the production of reliable<br />
actuator modules.<br />
Joint work with Roger Hamelinck<br />
91
5<br />
92 5 The variable reluctance actuator<br />
5.1 Introduction<br />
Figure 5.1 shows the schematic of the actuator that is used in the adaptive <strong>deformable</strong> <strong>mirror</strong><br />
<strong>and</strong> whose design process is discussed in detail in [174]. It consists of a PM that provides<br />
a pretension on a suspension membrane that is connected to the reflective <strong>mirror</strong> facesheet<br />
via a thin rod. An electric current through the coil at the bottom of the PM modulates the<br />
magnetic force of the PM on a small ferromagnetic core suspended by the membrane <strong>and</strong><br />
thus affects the deflection of the <strong>mirror</strong> facesheet. Figure 5.2 shows the CAD drawing<br />
of the actuator <strong>and</strong> figure 5.5 shows a photo of a single actuator. An insert which contains<br />
the PM, the coil <strong>and</strong> forms part of the magnetic circuit is shown on the left of the photo.<br />
The actuator is a so-called variable reluctance type actuator. This is an electromagnetic<br />
actuator in which an electromagnetic force is influenced by the variation of a reluctance in<br />
the path of the magnetic flux. The actuator is chosen as a variable reluctance actuator for<br />
several reasons. Firstly because the variable reluctance design has no moving coil. This<br />
means that no moving wires are required that form parasitic stiffnesses <strong>and</strong> are sensitive to<br />
break. Moreover, it allows for a low moving mass <strong>and</strong> a high resonance frequency. Finally,<br />
an indirect advantage is that air gaps – <strong>and</strong> thus reluctances – can be very small, leading to<br />
a high efficiency.<br />
The magnetic flux from the PM passes through the axial air gap, the ferromagnetic moving<br />
core <strong>and</strong> a radial air gap to one of the three pole shoes located in the baseplate <strong>and</strong> finally<br />
back to the PM. Since the mechanical spring is not a part of the magnetic flux path, no<br />
trade-off is required for the selection of the materials for the spring <strong>and</strong> the moving core.<br />
They can be chosen for mechanical strength <strong>and</strong> magnetic permeability respectively. In<br />
this specific actuator, the reluctance of the flux path through the axial air gap varies with<br />
the width of this gap, which depends on the position of the moving ferromagnetic core<br />
<strong>and</strong> thus on the deflection of the membrane. The PM pulls at the moving core, which is<br />
counteracted by the suspension membrane that forms a mechanical spring. This spring<br />
provides a positive stiffness that is matched by the negative stiffness of the PM. Since<br />
both stiffnesses vary with the actual deflection, multiple equilibrium points exists of which<br />
one is stable <strong>and</strong> yields a residual, positive actuator stiffness. A suitable value for this<br />
stiffness is desired that is low enough to prevent the need for high actuator forces <strong>and</strong> thus<br />
power dissipation, but high enough to keep the resonance frequency of the Deformable<br />
Mirror (DM) at approximately 1kHz regardless of its size. This minimum resonance<br />
Figure 5.1: Schematic of the variable<br />
reluctance actuator used in the adaptive<br />
<strong>deformable</strong> <strong>mirror</strong>.The magnetic flux<br />
from the PM passes through the axial air<br />
gap, the ferromagnetic moving core <strong>and</strong><br />
a radial air gap to one of the three pole<br />
shoes located in the baseplate <strong>and</strong> finally<br />
back to the PM.<br />
connection strut<br />
z<br />
axial airgap ferromagnetic moving core<br />
radial airgap membrane suspension<br />
pole shoe<br />
baseplate<br />
Φ<br />
n s<br />
PM<br />
<strong>mirror</strong> facesheet<br />
coil
5.1 Introduction 93<br />
Figure 5.2: CAD drawing of the variable reluctance actuator. The membrane suspension, with three<br />
leafsprings <strong>and</strong> the ferromagnetic core is shown. The magnetic flux from the PM passes<br />
through the axial air gap, through the ferromagnetic moving core <strong>and</strong> through a radial<br />
air gap the three pole shoes located in the baseplate, back to the PM.<br />
frequency is required to achieve the 200Hz <strong>control</strong> b<strong>and</strong>width as discussed in chapter 2. It<br />
also means that the DM system has a static gain up the the b<strong>and</strong>width, such that it can be<br />
adequately diagonalized for <strong>control</strong> using a static decoupling matrix. The corresponding<br />
required on actuator stiffness is discussed in detail in [174], where the optimum value is<br />
derived as 500N/m. This stiffness is low compared to the out-of-plane stiffness of the DM’s<br />
reflective facesheet, such that a malfunctioning actuator does not cause a hard point – i.e. a<br />
point with a fixed deflection – in the reflective surface <strong>and</strong> thus has a limited effect on the<br />
optical quality of the DM.<br />
The equilibrium position of the moving core can be influenced by an electric current<br />
flowing through the actuator coil that affects the magnetic force acting on it. This current is<br />
provided by driver electronics discussed in chapter 6.<br />
In this chapter, first the relevant actuator parts will be discussed <strong>and</strong> mathematical equations<br />
for their behavior will be derived. This leads to nonlinear models describing both the<br />
static <strong>and</strong> dynamic behavior of the actuator. A series of measurements is then performed<br />
on a single actuator prototype whose results are used to validate the derived model. The<br />
sensitivity of the actuator behavior is then analyzed w.r.t. geometric, magnetic <strong>and</strong> elec-<br />
5
5<br />
94 5 The variable reluctance actuator<br />
Figure 5.3: Dimensions in mm of the tested membrane designs.<br />
tric properties, which is used to improve the actuator design before manufacturing them in<br />
modules. Finally, measurement results will be shown for the actuator module prototypes<br />
<strong>and</strong> conclusions an recommendations will be formulated. Further information about the<br />
actuator design can be found in [174].<br />
5.2 The single actuator<br />
5.2.1 The actuator membrane suspension<br />
The stiffness of the membrane, in which the moving core is suspended, largely determines<br />
the actuator’s resonance frequencies <strong>and</strong> thereby the resonance frequencies of the adaptive<br />
Force [mN]<br />
200<br />
150<br />
100<br />
50<br />
0<br />
−50<br />
−100<br />
−150<br />
−200<br />
Measurement<br />
FEM data<br />
−150 −100 −50 0 50 100 150<br />
Displacement [µm]<br />
200<br />
150<br />
100<br />
50<br />
0<br />
−50<br />
−100<br />
−150<br />
−200<br />
Measurement<br />
FEM data<br />
−150 −100 −50 0 50 100 150<br />
Displacement [µm]<br />
Figure 5.4: Comparison of the nonlinear spring characteristic of the suspension designs as shown in<br />
figure 5.3, as calculated with FEM <strong>and</strong> as measured. The measurements were performed<br />
using the test setup described in [174].
5.2 The single actuator 95<br />
Figure 5.5: Photo of a single actuator. On the<br />
left, the insert is shown that contains the PM <strong>and</strong><br />
the coil.<br />
<strong>deformable</strong> <strong>mirror</strong> . The magnetic force generated by the PM <strong>and</strong> the coil acts on the moving<br />
core <strong>and</strong> pre-tensions the membrane suspension. According to [177], the nonlinear relation<br />
between spring forceFs <strong>and</strong> membrane deflectionzs can be approximated by:<br />
Fs(zs) = −C1<br />
Emt3 m<br />
r2 Emtm<br />
zs −C2<br />
m r2 z<br />
m<br />
3 s<br />
(5.1)<br />
where tm <strong>and</strong> rm are the suspension membrane thickness <strong>and</strong> radius <strong>and</strong> Em is the<br />
membrane material’s Young’s modulus. The coefficients C1 <strong>and</strong> C2 depend on the design<br />
<strong>and</strong> boundary conditions <strong>and</strong> will be estimated from FEM results <strong>and</strong> measurements.<br />
The membrane deflection zs follows the sign definition as indicated in figure 5.7 <strong>and</strong> is<br />
always negative due to the PM pretension. Since the resulting deflection is larger than the<br />
membrane thickness, a linear approximation of the stiffness is no longer valid: not only the<br />
bending stiffness Emt3 m/r2 <br />
m , but also the nonlinear stiffness due to in-plane stretching<br />
becomes relevant.<br />
Figure 5.3 shows on the left a design with springs placed radially towards a central<br />
disc. Results from both FEM analysis <strong>and</strong> measurements depicted in figure 5.4 show that<br />
the radially placed springs cause the suspension to stiffen quickly. This nonlinear effect is<br />
undesirable as this will complicate the design of a <strong>control</strong> system for the DM <strong>and</strong> reduce<br />
the achievable optical correction quality. One way to reduce the nonlinearity is to use a<br />
relatively thick membrane such that the linear term in (5.1) remains dominant for larger<br />
deflections zs. Another way to reduce the nonlinearity is to allow rotation of the central<br />
part to reduce the tensional forces. This leads to the design shown on the right in figure 5.3.<br />
Here the springs are placed tangentially. Due to the out-of-plane displacement, bending in<br />
the leaf springs occurs <strong>and</strong> the central part will rotate with typically 2 ◦ per Newton [2].<br />
Since the glued connection strut has a low rotational stiffness (figure 2.12 on page 40), this<br />
rotation will not lead to an actuator malfunction after assembly of the DM.<br />
Stiffness measurements<br />
To verify the FEM analysis a test set-up was designed to measure the nonlinear stiffness<br />
for different membrane suspensions. This measurement setup is described in [125] <strong>and</strong><br />
5
5<br />
96 5 The variable reluctance actuator<br />
in [174]. In the setup the displacement is measured optically with a Philtec D21 sensor<br />
with sub-µm resolution. The force required to enforce the displacement is measured with<br />
a Kistler 9203 piezo sensor, with mN resolution. The membrane suspensions are placed in<br />
containers (∅25x8mm) to be able to h<strong>and</strong>le them <strong>and</strong> place them in the measurement setup.<br />
In the containers the membranes are clamped at the outer edge. In the setup the suspensions<br />
are subjected to an out-of-plane displacement of ±100-150µm. For the suspensions shown<br />
in figure 5.3 the results of the nonlinear FEM analyses are compared with the measurements<br />
<strong>and</strong> shown in figure 5.4. A few remarks on the measurements are given below:<br />
• FEM prediction of the stress in the membrane is needed prior to the measurement to<br />
avoid plastic deformation during measurement.<br />
• Around the central position a negative stiffness is observed. This can be explained,<br />
partly from the stress present in the material due to the production of the foil, which is<br />
rolled, <strong>and</strong> partly from the clamping forces in the container. However, the membrane<br />
suspension will not be used in this position as the PM pre-tensions the membrane<br />
suspension. To deal with the negative stiffness, the membrane is placed a little higher<br />
above the magnet than originally designed. The spring force model in (5.1) is well<br />
able to describe a negative stiffness aroundzs = 0 <strong>and</strong> only the coefficientC1 needs to<br />
be adapted. By using a single rolled sheet for all actuators, the variation in membrane<br />
material stresses between the actuators is minimized.<br />
• The measurements are made using slow back <strong>and</strong> forward motion. A small difference<br />
between the two directions is observed, which is attributed partly to hysteresis in the<br />
clamp <strong>and</strong> partly to charge leakage in the piezo based sensor.<br />
The suspension membranes of the first actuators (e.g. figure 5.5) were made with titanium<br />
rolled sheets. The titanium sheets had limited yield strength (250N/mm 2 ) <strong>and</strong> yielded<br />
vulnerable actuators. In later designs, sheets of Havar , a non-magnetic, cobalt based,<br />
high-strength alloy, were used. Its yield strength is 1860N/mm 2 <strong>and</strong> its Young’s modulus<br />
200GPa [1]. The available choice in sheet thickness is limited. The design on the right<br />
in figure 5.4 was made with a 25µm Havar rolled sheet. The constants C1 <strong>and</strong> C2 were<br />
estimated using a least squares fit as:<br />
5.2.2 The electromagnetic force<br />
C1 = -0.12, C2 = 0.02.<br />
The axial magnetic force acting on the ferromagnetic core will be modeled as a function of<br />
the actuator deflection za (figure 5.7) <strong>and</strong> the actuator current Ia. First the magnetic flux<br />
density in the axial air gap is determined, followed by a force derivation based on magnetic<br />
coenergy.<br />
The magnetic circuit of the actuator from figure 5.2 is shown schematically in figure 5.6.<br />
The model includes leakage flux paths: one that short-circuits the coil <strong>and</strong> one that shortcircuits<br />
the PM. As will be shown later, the first one mainly affects the actuator coil inductance,<br />
whereas the latter affects many properties such as motor constant <strong>and</strong> actuator
5.2 The single actuator 97<br />
φ1<br />
φ3<br />
Va<br />
N<br />
S<br />
za<br />
φ2<br />
ℜb<br />
ℜgr<br />
φ1<br />
ℜc<br />
1<br />
ℜga (za)<br />
NIa<br />
ℜm<br />
-Hcm lm<br />
Figure 5.6: Left: a schematic representation of the variable reluctance actuator from figure 5.2.<br />
Right: the electrical equivalent circuit including two leakage flux paths.<br />
stiffness. The indicated flux paths contain two sources – the PM <strong>and</strong> the coil – <strong>and</strong> eight<br />
reluctances: the reluctance ℜm of the PM itself, ℜga(za) of the axial air gap, ℜc of the<br />
ferromagnetic core, ℜgr of the radial air gap, ℜb of the baseplate, ℜbc of the part of the<br />
baseplate that forms the core of the coil <strong>and</strong> ℜflc <strong>and</strong> ℜflm of the leakage flux paths that<br />
short-circuit the coil <strong>and</strong> PM respectively. Based on first principles, the reluctances of the<br />
PM, the axial <strong>and</strong> radial air gaps <strong>and</strong> the coil core are expressed as:<br />
lm<br />
ℜm =<br />
µ0µrmAm<br />
,ℜga(za) = z0 +za<br />
,ℜgr = lgr<br />
µ0Aga<br />
φ4<br />
ℜbc<br />
φ2<br />
lc<br />
,ℜbc =<br />
µ0Agr<br />
µ0µrbAm 2<br />
3<br />
φ5<br />
φ3<br />
ℜflm<br />
ℜflc<br />
, (5.2)<br />
where µ0 is the permeability of vacuum <strong>and</strong> µrm <strong>and</strong> µrb are the relative permeabilities of<br />
the PM <strong>and</strong> the baseplate material respectively.<br />
Since the thickness of the pole shoe is larger than the thickness of the ferromagnetic core<br />
plus the displacement range, the reluctance ℜgr of the radial air gap is considered to be<br />
independent of the displacement za. Am <strong>and</strong> Agr are the cross sectional areas of the flux<br />
paths through the PM <strong>and</strong> the radial air gap respectively <strong>and</strong> (z0 + za), lgr <strong>and</strong> lc are the<br />
axial air gap height <strong>and</strong> the lengths of the flux paths through the radial air gap <strong>and</strong> coil core<br />
respectively. A schematic of the actuator with the definitions of za <strong>and</strong> z0 is depicted in<br />
figure 5.7. The effective lengths <strong>and</strong> areas of the flux paths through the base plate <strong>and</strong> the<br />
ferromagnetic core are estimated from the actuator geometry, leading to the reluctancesℜb<br />
<strong>and</strong>ℜc as listed in table 5.1. Their values lie two orders of magnitude below the reluctances<br />
of the air gaps <strong>and</strong> the PM <strong>and</strong> will be combined into a single reluctance ℜr = ℜc + ℜb<br />
with a characteristic path lengthlr.<br />
Figure 5.6 indicates five different magnetic fluxesφ1...φ5 with positive directions. Since<br />
the sum of the fluxes towards each node must be equal to zero, these fluxes are related as:<br />
⎧<br />
⎪⎨ φ1 −φ2 +φ3 = 0,<br />
φ2 −φ3 −φ4 +φ5 = 0,<br />
(5.3)<br />
⎪⎩<br />
φ4 −φ5 −φ1 = 0.<br />
The PM is represented as a source with an internal reluctance <strong>and</strong> the coil as a source of<br />
magnetomotive force. According to Ampre’s law, the magnetomotive forces F1...F3<br />
5
5<br />
98 5 The variable reluctance actuator<br />
can be derived for the three different flux paths indicated in the figure as:<br />
⎧<br />
F1 =<br />
⎪⎨<br />
⎪⎩<br />
<br />
Hdl = NIa = Hmlm +Hga(za +z0)+Hrlr +Hgrlgr +Hbclbc,<br />
1<br />
F2 = <br />
Hdl = 0 = Hmlm +Hflmlflm,<br />
2<br />
F3 = <br />
Hdl = NIa = Hflclflc +Hbclbc,<br />
3<br />
where Hm, Hga, Hgr, Hr <strong>and</strong> Hbc are the magnetic field intensity in the PM, the axial air<br />
gap, the radial air gap, the combined baseplate <strong>and</strong> moving core <strong>and</strong> the coil core respectively.<br />
Assuming that all flux conductors represent linear magnetic materials, their flux densitiesB<br />
are related to their magnetic field intensityH via the material’s magnetic permeabilityµ as<br />
B = µH. For the PM, this relation includes an offset:<br />
Bm = µ0µrm(Hm −Hcm), (5.4)<br />
where Hcm is the coercivity of the PM. According to Gauss’s law, flux φ is the integral of<br />
the flux densityB over an areaA:<br />
<br />
φ = B ·dA = BA, (5.5)<br />
A<br />
where the latter equality assumes that the flux density is constant over the cross-sectional<br />
areaA.<br />
When substituted into (5.4), this allows the magnetic field intensity of the PM to be<br />
expressed as:<br />
Hm = Bm<br />
µ0µrm<br />
+Hcm =<br />
φ4<br />
µ0µrmAm<br />
+Hcm. (5.6)<br />
Substitution of (5.6), Gauss’s law from (5.5), the definitions of the reluctances from (5.2)<br />
<strong>and</strong> the linear relations between flux density <strong>and</strong> magnetic field intensity into the expressions<br />
forF1...F3 then yields:<br />
⎧<br />
⎪⎨ F1 = NIa = ℜmφ4 +Hcmlm +(ℜga(za)+ℜr +ℜgr)φ1 +ℜbcφ2,<br />
⎪⎩<br />
F2 = 0 = ℜmφ4 +Hcmlm +ℜflmφ5,<br />
F3 = NIa = ℜflcφ3 +ℜbcφ2.<br />
Figure 5.7: Definition of the axial air gap height<br />
through the initial gap z0 <strong>and</strong> the displacement<br />
za. The height h is the axial air gap when<br />
the suspension membrane is not deflected, i.e.<br />
zs = 0. The membrane deflectionzs is related to<br />
the actuator displacementza aszs = za+z0−h,<br />
<strong>and</strong> is negative for downward membrane deflection.<br />
-zs<br />
h z0<br />
za
5.2 The single actuator 99<br />
From these three relations <strong>and</strong> the three flux relations in (5.3), five uncoupled expressions<br />
can be solved for the fluxesφ1...φ5 as:<br />
where<br />
φ1(Ia,za) = NIaℜflcℜ2 −Hcmlmℜflmℜ3<br />
˜ℜ(za)<br />
φ2(Ia,za) = NIaℜ(za)−Hcmlmℜflcℜflm<br />
˜ℜ(za)<br />
φ3(Ia,za) = NIa(ℜflmℜm +ℜ1(za)ℜ2)+Hcmlmℜbcℜflm<br />
˜ℜ(za)<br />
φ4(Ia,za) = NIaℜflcℜflm −Hcmlmℜ(za)<br />
˜ℜ(za)<br />
φ5(Ia,za) = NIaℜflcℜm −Hcmlm(ℜbcℜflc +ℜ3ℜ1(za))<br />
˜ℜ(za)<br />
ℜ1(za) = ℜga(za)+ℜr +ℜgr,<br />
ℜ2 = ℜm +ℜflm,<br />
ℜ3 = ℜbc +ℜflc,<br />
ℜ(za) = (ℜflc +ℜ1(za))ℜ2 +ℜflmℜm,<br />
ℜ(za) = (ℜflm +ℜ1(za))ℜ3 +ℜbcℜflc,<br />
˜ℜ(za) = (ℜflmℜm +ℜ2ℜ1(za))ℜ3 +ℜbcℜflcℜ2.<br />
(5.7a)<br />
(5.7b)<br />
(5.7c)<br />
(5.7d)<br />
(5.7e)<br />
Note that the flux φ4(Ia,za) through the PM will be zero when the winding current Ia is<br />
equal to Iacc = −(Hcmlmℜ(za)/(Nℜflmℜflc) or in absence of leakage flux to Iacc =<br />
−Hcmlm/N. For this current, the coil’s magnetic field fully cancels that of the PM. For<br />
the values in table 5.1 <strong>and</strong> in the absence of leakage flux, this corresponds to a current<br />
Ia =324mA. For the coil’s ∅50µm copper wire this corresponds to an unrealistic current<br />
density of 165A/mm 2 .<br />
From the derived expressions for the fluxes, the operating point of the PM on its B-H curve<br />
is obtained. This operating point indicates how efficiently the volume of the PM is used<br />
to generate a desired flux density. Substitution of the flux φ4(Ia,za) from (5.7d) into the<br />
expression for the magnetic field intensity of the PM in (5.6) provides the magnetic field<br />
intensityHm of the PM as:<br />
Hm = Hcm − NIaℜflmℜflc +Hcmlmℜ(za)<br />
µ0µrmAm ˜ . (5.8)<br />
ℜ(za)<br />
In the unactuated state (i.e. Ia = 0, za = 0), with the use of the values from table 5.1,<br />
this yields Hm =-313kA/m. Subsequent substitution into (5.4) then leads to Bm =0.33T.<br />
The product of flux densityBm <strong>and</strong> magnetic field intensityHm indicates the available PM<br />
energy per unit volume. Using (5.4), the maximum of this product is found as:<br />
<br />
<br />
|BmHm| max = µ0µrm(Hm −Hcm)Hm<br />
= µ0µrmH<br />
max 2 cm /4 = 106kJ/m3 .<br />
5
5<br />
100 5 The variable reluctance actuator<br />
The value derived for the actuator is |HmBm| =104kJ/m 3 , which is very close to the<br />
optimum. In fact, this optimum corresponds to the situation whenHm = −Hcm/2 <strong>and</strong> thus<br />
– when considering (5.8) forIa = 0 – to:<br />
ℜ(za) 1<br />
= .<br />
˜ℜ(za) 2ℜm<br />
In absence of leakage flux (i.e. ℜflc,ℜflm → ∞) this reduces to ℜm = ℜbc + ℜ1 =<br />
ℜbc +ℜgr +ℜr +ℜga(za), which implies that the highest volume efficiency of the PM is<br />
obtained when the internal reluctance ℜm of the PM is exactly equal to the external reluctance<br />
felt by the PM.<br />
The magnetic force on the ferromagnetic core is calculated via flux linkage <strong>and</strong> magnetic<br />
coenergy [59]. In this procedure the PM is modeled as a fictitious winding with the equivalent<br />
magnetomotive force. This means thatHcmlm is replaced by−NfIf , whereNf is the<br />
number of turns of the fictitious winding <strong>and</strong> If the fictitious current through it. The flux<br />
linkagesλ of the coil <strong>and</strong>λf of the fictitious winding are given by [59]:<br />
λ(Ia,za) = Nφ2(Ia,za) = L11(za)Ia +L12(za)If, (5.9)<br />
λf(Ia,za) = Nfφ4(Ia,za) = L21(za)Ia +L22(za)If,<br />
whereL11(za) <strong>and</strong>L22(za) are the self inductances of the coil <strong>and</strong> the PM <strong>and</strong>L12(za) <strong>and</strong><br />
L21(za) the corresponding mutual inductances:<br />
L11(za) = N2 ℜ(za)<br />
˜ℜ(za)<br />
L21(za) = NNfℜflcℜflm<br />
˜ℜ(za)<br />
, L12(za) = NNfℜflcℜflm<br />
,<br />
˜ℜ(za)<br />
, L22(za) = N2 fℜ(za) .<br />
˜ℜ(za)<br />
(5.10)<br />
As expected, the two mutual inductances L12(za) <strong>and</strong> L21(za) are equal. The magnetic<br />
coenergy can be expressed in terms of these (mutual) inductances as [59]:<br />
W(Ia,za) = 1<br />
2 L11(za)I 2 a +L12(za)IaIf + 1<br />
2 L22(za)I 2 f, (5.11)<br />
= 1<br />
2 L11(za)I 2 a<br />
+ L12(za)<br />
Nf<br />
IaHcmlm + L22(za)<br />
2N2 H<br />
f<br />
2 cml2 m ,<br />
1<br />
=<br />
2˜ 2 2<br />
N Iaℜ(za)−2NIaHcmlmℜflmℜflc +H<br />
ℜ(za)<br />
2 cml2 mℜ(za) .<br />
Note that in the second step the magnetomotive forceNfIf of the fictitious winding is again<br />
replaced by the −Hcmlm of the PM. After substitution of the inductances from (5.10) in<br />
the final step, the expression forW(Ia,za) becomes independent ofNf <strong>and</strong>If .<br />
The electromagnetic forceFm(Ia,za) is equal to the partial derivative of the coenergy with<br />
respect to the displacementza <strong>and</strong> can be expressed as:<br />
Fm(Ia,za) = ∂W(Ia,za)<br />
∂za<br />
= −1<br />
2Agaµ0<br />
2 NIaℜflcℜ2 −Hcmlmℜflmℜ3<br />
˜ℜ(za)<br />
(5.12)
5.2 The single actuator 101<br />
5.2.3 A static actuator model<br />
Together, the derived equations for the electromagnetic force Fm(Ia,za) in (5.12) <strong>and</strong> the<br />
mechanical spring force in (5.1) provide static relations for the behavior of the actuator.<br />
This provides insight in the nonlinear actuator stiffness, the required actuator current <strong>and</strong><br />
voltage, its motor constant <strong>and</strong> power dissipation. However, the derived relations do not<br />
provide insight into dynamic properties such as resonance frequency, damping, inductance,<br />
etc. that affect the achievable <strong>control</strong>ler performance <strong>and</strong> thus the correction quality of the<br />
<strong>Adaptive</strong> Optics (AO) system. Therefore, in section 5.2.4 these equations are extended to<br />
also include dynamic behavior.<br />
The static force equilibrium can be expressed as:<br />
Fs(zs)+Fm(Ia,za) = 0. (5.13)<br />
Let the nominal operating point of the actuator be defined as the unactuated equilibrium<br />
point, where the air gap z0 is such that the electromagnetic force Fm(Ia = 0,za = 0)<br />
equals the membrane suspension spring force Fs(zs). The deflection zs of the suspension<br />
membrane can be expressed in terms of theza, z0 <strong>and</strong>has defined in figure 5.7 by:<br />
zs = za +z0 −h.<br />
Using this definition withza = 0, the initial gapz0 can be solved from the force equilibrium<br />
in (5.13) through substitution of (5.12) <strong>and</strong> (5.1):<br />
− 1<br />
2µ0Aga<br />
2 Hcmlmℜflmℜ3<br />
˜ℜ(za)<br />
Emt<br />
−C1<br />
3 m<br />
r2 Emtm<br />
(z0 −h)−C2<br />
r2 (z0 −h)<br />
m<br />
3 = 0, (5.14)<br />
This leads to a fifth order equation in z0 with five solutions. The solution that corresponds<br />
to practice is real-valued. Moreover, it forms a stable equilibrium – i.e. the derivative of<br />
the sum of forces with respect to z0 at the solution for z0 is negative – <strong>and</strong> finally has a<br />
value within the range 0 < z0 < h. This solution is found numerically after substitution<br />
of the parameters from table 5.1. Note that the values for some parameters in this table are<br />
determined from measurements as described in section 5.2.5. The initial gap z0 found is<br />
z0 =109µm. The spring <strong>and</strong> electromagnetic forces of (5.14) are plotted in figure 5.8 as a<br />
function ofz0.<br />
5
5<br />
102 5 The variable reluctance actuator<br />
Force [N]<br />
Table 5.1: Electromagnetic <strong>and</strong> mechanical parameters of the variable reluctance actuator.<br />
0.2<br />
0.15<br />
0.1<br />
0.05<br />
0<br />
Parameter Value Unit Parameter Value Unit<br />
Hcm -540 a kA/m C1 -0.12 c -<br />
lm 0.30 mm C2 0.02 c -<br />
Aga 0.79 mm 2 Em 200 GPa<br />
Agr 0.48 mm 2 tm 25 µm<br />
Am 0.79 mm 2 rm 2.5 mm<br />
ℜga(0) 111 1/µH mac 3.6 mg<br />
ℜm 262 1/µH Ra 39.0 Ω<br />
ℜr 512 1/µH h 230 d µm<br />
ℜgr 248 1/µH z0 109 e µm<br />
ℜbc 1 1/µH ba 0.4 f mNs/m<br />
ℜflc 100 b 1/µH ca 583 e N/m<br />
ℜflm 600 b 1/µH<br />
a Estimated from PM measurements [174]<br />
b Value estimated from measurements on actuator prototypes<br />
c Fitted on nonlinear FEM model with typical value for negative stiffness included<br />
d Design parameter<br />
e Derived value<br />
f Estimated from actuator measurements<br />
Spring force F s<br />
Magnet force F m<br />
Residual force<br />
Equilibrium<br />
−0.05<br />
0 50 100 150 200<br />
Air gap z [µm]<br />
0<br />
Figure 5.8: Forces exerted by the membrane suspension<br />
<strong>and</strong> the PM respectively as a function of<br />
the axial air gap.<br />
Stiffness [N/m]<br />
800<br />
700<br />
600<br />
500<br />
400<br />
300<br />
200<br />
100<br />
Spring<br />
Magnet<br />
Residual<br />
0<br />
−10 −5 0 5 10<br />
Displacement z [µm]<br />
Figure 5.9: The stiffness of the suspension membrane<br />
<strong>and</strong> PM as a function of the actuated displacement<br />
za around the equilibrium z0.<br />
The relation between currentIa <strong>and</strong> displacementza is found by solving the force equilibrium<br />
equation in (5.13) for the actuator currentIa, yielding:<br />
<br />
Hcmlmℜflmℜ3 ±<br />
Ia(za) =<br />
−2Agaµ0 ˜ ℜ 2 (za)<br />
Nℜflcℜ2<br />
C1Emt 3 s<br />
r 2 m<br />
zs + C2Emtm<br />
r2 z<br />
m<br />
3 <br />
s<br />
(5.15)<br />
where zs is used for brevity of the expression. Since (5.12) <strong>and</strong> thus (5.13) is quadratic in
5.2 The single actuator 103<br />
the current, there are two solutionsIa(za) – as indicated by the ± sign – of which only the<br />
solution with a plus sign is valid. Since a current yields a displacement za around the z0<br />
for which Ia = 0, both positive <strong>and</strong> negative currents lead to realistic displacements. The<br />
solution forIa(za) with a minus sign could never lead to a positive current because the first<br />
term(Hcmlmℜflmℜ3) is always negative (Hcm is negative) <strong>and</strong> for realistic displacements<br />
the square root term is always positive. The viable relation forIa(za) is plotted in figure 5.10<br />
for the physical properties in table 5.1. Note that despite the nonlinearity of the equations,<br />
the relation between current <strong>and</strong> deflection is highly linear within the intended operating<br />
range of−10µm< za < 10µm.<br />
The mechanical stiffnessca of the actuator can be derived by taking the partial derivative of<br />
the sum of forces in (5.13) with respect to the deflectionza. Although this remains a function<br />
of both Ia <strong>and</strong> za, these quantities are statically coupled through (5.15). The dependence<br />
onIa can therefore be replaced by an implicit constantIaz = Ia(za) that denotes the static<br />
current required to reach the displacement za = z ′ a. The mechanical stiffness ca(za) can<br />
thus be derived as:<br />
ca(Iaz,za) = − ∂<br />
(Fs(za)+Fm(Ia = Iaz,za))<br />
∂za<br />
= − ∂<br />
<br />
−C1<br />
∂za<br />
Emt3 m<br />
r2 zs −C2<br />
m<br />
Emtm<br />
r 2 m<br />
Emt<br />
= C1<br />
3 m<br />
r2 Emtm<br />
+3C2<br />
m r2 (za +z0 −h) 2<br />
− ℜ2ℜ3(NIazℜflcℜ2 −Hcmlmℜflmℜ3) 2<br />
µ 2 0A2 ˜ℜ ga<br />
3 .<br />
(za)<br />
z 3 s − (NIazℜflcℜ2 −Hcmlmℜflmℜ3) 2<br />
˜ℜ 2µ0Aga<br />
2 (za)<br />
By substituting the relation between current Iaz <strong>and</strong> displacement za from (5.15), an expression<br />
for the actuator stiffness in terms of onlyza is found:<br />
ca(za) = − C2Emtm<br />
r 2 mµ0Aga<br />
˜ℜ(za) z3 s + 3C2Emtm<br />
r2 m<br />
z 2 s −<br />
C1Emt3 m<br />
r2 ˜ℜ(za) mµ0Aga<br />
zs+ C1Emt3 m<br />
r2 m<br />
, (5.16)<br />
where zs was resubstituted for brevity of the expression. Note that this expression is a<br />
function of the initial air gap z0, such that it does not fully reflect the effect of parameters<br />
affectingz0. Moreover, the stiffness is a third order polynomial in the deflectionza, whereas<br />
without the PM this was second order. Based on the numerical solution for z0 <strong>and</strong> the<br />
other properties shown in table 5.1, the actuator stiffness has been plotted in figure 5.9 for<br />
displacements za within the intended operating range. The stiffness decreases for positive<br />
za <strong>and</strong> increases for negative za <strong>and</strong> varies approximately 16.5% over the full intended<br />
operating range. At the equilibriumza = 0 <strong>and</strong>zs = z0 −h, the actuator stiffness is found<br />
asca(0) =583N/m.<br />
Figure 5.10 also shows the dissipated powerPa(za) calculated via Ohm’s law:<br />
Pa(za) = I 2 a (za)Ra,<br />
where Ra is the resistance of the actuator coil that can be expressed in terms of geometry<br />
<br />
5
5<br />
104 5 The variable reluctance actuator<br />
Current I a [mA], Power P a [mW]<br />
40<br />
30<br />
20<br />
10<br />
0<br />
−10<br />
−20<br />
−30<br />
Current<br />
Power<br />
−10 −5 0 5 10<br />
Displacement z [µm]<br />
Figure 5.10: The actuator current Ia <strong>and</strong> corresponding<br />
power dissipation required for a displacement<br />
za.<br />
<strong>and</strong> material properties as:<br />
Actuator force F a [mN]<br />
8<br />
6<br />
4<br />
2<br />
0<br />
−2<br />
−4<br />
−6<br />
Ra = 2πNrcaρe<br />
,<br />
Aw<br />
Force at z = −10µm<br />
Force at z = 0µm<br />
Force at z = 10µm<br />
−8<br />
−30 −20 −10 0 10 20 30<br />
Current ∆I [mA]<br />
Figure 5.11: The force that can be exerted by the<br />
actuator on the DM facesheet as a function of the<br />
current Ia at three deflections za.<br />
whererca is the average coil radius,ρe the specific resistance of the coil’s material <strong>and</strong>Aw<br />
the cross-sectional area of the coil’s wire. For the∅50µm copper wire used this leads to the<br />
valueRa ≈39.0Ω found in table 5.1.<br />
In figure 5.11 the generated actuator force Fa is plotted. This is the external force required<br />
to keep the actuator at a fixed operating point za = z ′ a<br />
as a function of a supplied current<br />
offset ∆Ia = Ia −Ia(z ′ a). It is calculated by augmenting the static force equilibrium with<br />
an additional termFa(∆Ia,z ′ a ) <strong>and</strong> solving for it:<br />
Fa(∆Ia,z ′ a) = Fm(Ia(z ′ a)+∆Ia,z ′ a)+Fs(z0 +z ′ a −h). (5.17)<br />
The common relation between force <strong>and</strong> current in linear systems is through the motor<br />
constant defined in Newtons per Ampère, whereas (5.17) expresses this relation for the<br />
derived nonlinear system. Figure 5.11 shows that the relation between the current offset<br />
∆Ia <strong>and</strong> forceFa(∆Ia,z ′ a<br />
) is highly linear within the intended operating range. According<br />
to this figure, the generated force due to a change in current is only marginally different<br />
at the intended extreme operating points. In fact, when nonlinearities are neglected, it can<br />
be observed from figure 5.11 that the force per current unit – i.e. the motor constant – is<br />
approximately 0.2N/A.<br />
5.2.4 A dynamic actuator model<br />
The derived equations that describe the electromagnetic part of the actuator will be extended<br />
with the mechanical equations of motion to a (nonlinear) dynamic model. This model is<br />
linearized to obtain Bode plots of the actuator <strong>and</strong> linear electromechanical properties such<br />
as motor constant <strong>and</strong> coil inductance.<br />
The terminal voltageVa(t) over the actuator coil can be expressed as a function of timetin
5.2 The single actuator 105<br />
[s] using the flux linkage termλ(Ia,za) corresponding to the actuator coil as [59]:<br />
Va(t) = Ia(t)R+ ∂λ(Ia,za)<br />
∂t<br />
(5.18)<br />
In the expression for the flux linkage in (5.9), the magnetomotive force of the PM is expressed<br />
by that of a fictitious winding asNfIf . The latter can again be replaced by−Hcmlm<br />
of the PM, leading to:<br />
λ(Ia,za) = N <br />
NIaℜ−Hcmlmℜflcℜflm . (5.19)<br />
˜ℜ(za)<br />
Since this expression forλ(Ia,za) is a function of both currentIa <strong>and</strong> positionza, the partial<br />
derivative in (5.18) w.r.t. time t exp<strong>and</strong>s via the chain rule into:<br />
Va = IaRa + ∂λ ∂Ia ∂λ ∂za<br />
+<br />
∂Ia ∂t ∂za ∂t<br />
= IaR+ ∂λ<br />
Ia ˙ +<br />
∂Ia<br />
∂λ<br />
˙za<br />
∂za<br />
(5.20)<br />
where the dependence on t has been omitted for brevity. The first partial derivative represents<br />
the self-inductance voltage term, whereas the second occurs in product with the actuator<br />
velocity ˙za <strong>and</strong> is called the speed voltage. The latter is common to all electromechanical<br />
energy-conversion systems <strong>and</strong> is responsible for energy transfer between the mechanical<br />
system <strong>and</strong> the electrical system.<br />
Substitution of the flux linkageλ(Ia,za) of (5.19) into (5.20) then leads to:<br />
Va = IaRa + N2ℜ ˙<br />
˜ℜ(za)<br />
Ia +<br />
Nℜflc<br />
Agaµ0ℜ2 ˜ ℜ 2 (za) (Hcmlmℜflmℜ3 −NIaℜflcℜ2) ˙za (5.21)<br />
This equation describes the electromagnetic part of the system that generates the force<br />
Fm(za,Ia) in (5.12) on the suspended mass. The equation of motion of this mass-springdamper<br />
system can be expressed as:<br />
mac ¨za +ba˙za = Fm(za,Ia)+Fs(za +z0 −h), (5.22)<br />
where ˙za <strong>and</strong> ¨za are the first <strong>and</strong> second derivatives of za to time t respectively, ba is the<br />
mechanical viscous damping <strong>and</strong>mthe mass of the ferromagnetic moving core. Recall that<br />
the magnetic forceFm(Ia,za) is defined in (5.12) <strong>and</strong> the spring forceFs(zs) in (5.1). Together<br />
with (5.21) this equation forms a nonlinear dynamic actuator model. When assuming<br />
a state vectorx = [Ia za ˙za] T , the time derivative of this state vector can be expressed as:<br />
⎡ ⎤ ⎡ ˜ℜ(za)(Va−RaIa)<br />
Ia ˙<br />
⎣˙za<br />
⎦ ⎢ N<br />
= ⎣<br />
¨za<br />
2 ℜflc<br />
−<br />
ℜ(za) Aga µ0Nℜ2˜ ℜ(za)ℜ(za) (Hcmlmℜflmℜ3<br />
⎤<br />
−NIaℜflcℜ2) ˙za<br />
⎥<br />
˙za<br />
⎦<br />
1 (−Fm(za,Ia)−ba˙za −Fs(za +z0 −h))<br />
mac<br />
(5.23)<br />
where ˙ Ia was solved from (5.21) <strong>and</strong> ¨za from (5.22) after substitution of (5.12) <strong>and</strong> (5.1). All<br />
state derivatives except ˙za are nonlinear equations in terms of the state variables. However,<br />
if the effect of the nonlinearities on the actuator behavior is small, the nonlinear equations<br />
5
5<br />
106 5 The variable reluctance actuator<br />
will only complicate the design of a <strong>control</strong>ler. Therefore, this effect will be investigated<br />
in the next subsection through linearization of the nonlinear equations. This will also provide<br />
insight into (linear) dynamic properties such as inductance, resonance frequency <strong>and</strong><br />
damping.<br />
Linearization of the dynamic model<br />
Linearization is a widely used technique that enables the application of linear, frequency<br />
domain tools on nonlinear systems. However, it can only provide useful insight when the<br />
nonlinearities play a negligible role around a certain operation point or state. To verify this<br />
for the nonlinear system in (5.23), linearizations at several displacements za = z ′ a around<br />
the initial air gapz0 will be derived <strong>and</strong> their frequency response functions plotted.<br />
The coil terminal voltage Va serves as the system input <strong>and</strong> is assumed to be supplied by a<br />
voltage source. The system output is the displacementza, where no notational difference is<br />
made between the original <strong>and</strong> linearized system description as this is always clear from the<br />
context. The linearized system with state xl = [Ia za ˙za] T can thus be expressed as [69]:<br />
˙xl(t) = ∂˙x<br />
∂ x T<br />
<br />
<br />
<br />
<br />
<br />
x=xz ′ ,Va=Ia<br />
a z ′<br />
Ra<br />
a<br />
xl(t)+ ∂˙x<br />
<br />
<br />
<br />
<br />
∂ Va<br />
x=xz ′<br />
a<br />
Va(t),<br />
where the operating pointxz ′ a is chosen as[Iaz ′ z<br />
a<br />
′ a 0] T whereIaz ′ = Ia(z<br />
a<br />
′ a) <strong>and</strong> the current<br />
required for the actuator displacement za = z ′ a as plotted in figure 5.10. In the operating<br />
point used for linearization the velocity ˙za is assumed to be zero. The output of this system<br />
can then be chosen as displacementza, velocity ˙za <strong>and</strong>/or currentIa. After taking the partial<br />
derivatives, substitutingx = xz ′ a <strong>and</strong>Va = IaRa <strong>and</strong> omitting dependence ontfor brevity<br />
this leads to:<br />
⎡<br />
−Ra/La(z ′ a ) 0 −Ka(z ′ a )/La(z ′ a )<br />
Al<br />
⎤<br />
⎡<br />
1/La(z ′ a )<br />
˙xl = ⎣ 0 0 1<br />
Ka(z ′ a )/mac −ca(z ′ a )/mac<br />
⎦xl<br />
+ ⎣ 0 ⎦Va,<br />
(5.24)<br />
<br />
−ba/mac<br />
<br />
0<br />
<br />
whereca(z ′ a ) was defined in (5.16) <strong>and</strong> the elementsIa, za <strong>and</strong> ˙za of the state xl now form<br />
small signal variations around the operating point xz ′ a . Further, Ka(z ′ a ) <strong>and</strong> La(z ′ a ) are<br />
the motor constant <strong>and</strong> inductance at the operating point za = z ′ a respectively <strong>and</strong> can be<br />
expressed as:<br />
Ka(z ′ a ) = Nℜflcℜ2(NIaℜflcℜ2 −Hcmlmℜflmℜ3)<br />
˜ℜ µ0Aga<br />
2 (z ′ a )<br />
La(z ′ a) = N2 (ℜ1ℜ2 +ℜflcℜ2 +ℜflmℜm)<br />
˜ℜ(z ′ a )<br />
Bl<br />
,<br />
⎤<br />
(5.25)<br />
Figure 5.17 shows the values of the motor constant <strong>and</strong> the coil inductance as function of<br />
. Both the motor constant <strong>and</strong> the inductance decrease as the air<br />
the operation pointza = z ′ a<br />
gapz0+za increases, but for the latter the influence of the operating point is smaller. Bode<br />
plots of the actuator response are plotted for several operating pointsza = z ′ a in figure 5.12.
5.2 The single actuator 107<br />
Magnitude [m/V]<br />
Phase [deg]<br />
10 −4<br />
10 −5<br />
10 −6<br />
0<br />
−90<br />
−180<br />
z a ’ = −10.0 [µm]<br />
z a ’ = −5.0 [µm]<br />
z a ’ = 0.0 [µm]<br />
z a ’ = 5.0 [µm]<br />
z a ’ = 10.0 [µm]<br />
−270<br />
1000 1500 2000 2500 3000<br />
Frequency [Hz]<br />
Figure 5.12: Bode plots of the modeled transfer<br />
function between actuator voltage Va <strong>and</strong><br />
displacement za for various operating points<br />
za = z ′ a.<br />
This shows a first resonance frequency of the system around 2.04kHz, which increases as<br />
the air gapz0+z ′ a increases. It should be noted that the ratioRa/La between the electrical<br />
resistance <strong>and</strong> inductance of the actuator corresponds to a pole located at approximately<br />
2.1kHz, which is very close to the mechanical resonance frequency. Although this pole will<br />
affect the two poles corresponding to the system’s mechanics, this effect is small.<br />
Although the dynamic system equations were mainly derived to gain insight into the dynamic<br />
behavior of the actuator, the DC-gain of the linearized system also provides a direct<br />
relation between the supplied clamp voltage Va <strong>and</strong> the deflection za. This DC-gain H(0)<br />
can be derived from the state-space model in (5.24) by first rewriting it to transfer function<br />
form as:<br />
H(s) = 0 1 0 (sI3 −Al) −1 Bl,<br />
where the displacement za is chosen as the output <strong>and</strong> Al <strong>and</strong> Bl are the system matrices<br />
defined in (5.24). Further,s = jω is the complex Laplace variable <strong>and</strong>I3 the identity matrix<br />
of size 3×3. Subsequent substitution ofs = jω = 0 then yields the DC-gainH(0) as:<br />
H(0) = Ka(z ′ a )<br />
Raca(z ′ . (5.26)<br />
a)<br />
Observe that this gain depends on the coil resistance, the motor constant <strong>and</strong> the resulting<br />
actuator stiffness, which will all vary per actuator due to material <strong>and</strong> manufacturing tolerances.<br />
As a result, the DC-gain is expected to have a relatively large variation from actuator<br />
to actuator.<br />
A simplified case: no leakage flux<br />
The expressions describing the actuator behavior derived so far are complicated by the presence<br />
of the leakage flux around the coil <strong>and</strong> the PM. Although this leads to a more realistic<br />
model that is better able to describe the measurement results, it makes the physical interpretation<br />
of the expressions more difficult. Therefore, the case will be considered when the<br />
leakage flux is completely absent. This will provide a clearer underst<strong>and</strong>ing of the relations<br />
between quantities such as geometric dimensions, number of coil turns <strong>and</strong> inductance, motor<br />
constant, etc.<br />
The absence of leakage flux corresponds to the limit of all above electromagnetic equations<br />
forℜflm,ℜflc → ∞. For the fluxesφ2(Ia,za) <strong>and</strong>φ4(Ia,za) through the coil <strong>and</strong> the PM<br />
5
5<br />
108 5 The variable reluctance actuator<br />
respectively, this leads to:<br />
φ ′ 2(Ia,za) = lim<br />
ℜflm,ℜflc→∞ φ2(Ia,za) = NIa −Hcmlm<br />
,<br />
(za)<br />
µ0Agaℜ ′ 1<br />
φ ′ 4 (Ia,za) = lim<br />
ℜflm,ℜflc→∞ φ4(Ia,za) = NIa −Hcmlm<br />
,<br />
(za)<br />
whereℜ ′ 1 (za) denotes the sum of all remaining reluctances:<br />
µ0Agaℜ ′ 1<br />
ℜ ′ 1(za) = ℜga(za)+ℜr +ℜbc +ℜgr +ℜm.<br />
Note that since there is only a single flux path left, fluxesφ2 <strong>and</strong>φ4 are equal.<br />
Similarly, for ℜflm,ℜflc → ∞ the expression for the magnetic coenergy W(Ia,za) in<br />
(5.11) reduces to:<br />
W ′ (Ia,za) = lim<br />
ℜflm,ℜflc→∞ W(Ia,za) = (NIa −Hcmlm) 2<br />
2µ0Agaℜ ′ 1 (za)<br />
<strong>and</strong> that for the magnetic force in (5.12) to:<br />
F ′ m (Ia,za) = lim<br />
2µ0Aga (ℜ ′ 2<br />
1 (za))<br />
ℜflm,ℜflc→∞ Fm(Ia,za) = −(NIa −Hcmlm) 2<br />
= −1<br />
2 µ0Aga (φ ′ 2 (Ia,za)) 2 .<br />
The fluxes, the coenergy <strong>and</strong> the magnetic force are all proportional to the total magnetomotive<br />
force in the system <strong>and</strong> inversely proportional to the total, position dependent<br />
reluctance ℜ ′ 1(za). In fact, the fluxes are linearly proportional to the total magnetomotive<br />
force, whereas the coenergy <strong>and</strong> force are both quadratically proportional to this. The resulting<br />
cross product between the magnetomotive force NIa of the coil <strong>and</strong> the Hcmlm of<br />
the PM amplifies the effect of the coil on the electromagnetic force, which is an advantage<br />
of using a PM to pre-load the spring.<br />
Comparison of the derived limit equations for flux, magnetic coenergy <strong>and</strong> electromagnetic<br />
force with the original equations in (5.7d), (5.11) <strong>and</strong> (5.12) shows that leakage flux leads<br />
to additional weights on the magnetomotive forces of the coil <strong>and</strong> the PM as well as on their<br />
cross products.<br />
For the static actuator model, the absence of leakage flux leads to a simplified relation between<br />
currentIa <strong>and</strong> displacementza:<br />
I ′ a (za) = lim<br />
ℜflm→∞<br />
ℜflc→∞<br />
Ia(za) = Hcmlm<br />
N<br />
+ 1<br />
N<br />
<br />
−2(ℜ ′ 2<br />
1 (za))<br />
<br />
C1<br />
Emt3 m<br />
r2 Emtm<br />
zs +C2<br />
m r2 z<br />
m<br />
3 <br />
s ,<br />
wherezs is again used for brevity of notation. The relation between actuator stiffnessca(za)<br />
<strong>and</strong> displacementza becomes:<br />
c ′ a (za) = lim<br />
ℜflm,ℜflc→∞ ca(za),<br />
= − 2C2Emtm<br />
r2 m µ0Agaℜ ′ z<br />
1<br />
3 s<br />
3C2Emtm<br />
+<br />
r2 z<br />
m<br />
2 s + 2C1Emt3 m<br />
r2 µ0Agaℜ ′ 1<br />
zs + C1Emt3 m<br />
r2 , (5.27)<br />
m
5.2 The single actuator 109<br />
Siglab<br />
excitation<br />
Current<br />
amplifier<br />
actuator<br />
Laservibrometer<br />
velocity<br />
Figure 5.13: Experimental<br />
setup to measure the behavior<br />
of a single actuator using<br />
a Siglab TM system, a current<br />
source <strong>and</strong> a laser vibrometer.<br />
where the dependence ofℜ ′ 1 on the actuator deflectionza has been omitted for brevity. Observe<br />
that the actuator stiffness remains a third order polynomial in the membrane deflection<br />
zs, where only the first <strong>and</strong> third order terms depend on the magnetic circuit <strong>and</strong> are scaled<br />
by the total path reluctanceℜ ′ 1(za).<br />
When the leakage flux is neglected for the equations describing the dynamic actuator be-<br />
havior in (5.23), the expression for ˙<br />
Ia becomes:<br />
I ˙′<br />
a (za) = lim<br />
ℜflm,ℜflc→∞<br />
Ia(za) ˙ = −ℜ′ 1 (za)<br />
N2 <br />
RaIa −Va + N (Hcmlm −NIa)<br />
Agaµ0ℜ ′ 1 (za)<br />
<br />
˙za .<br />
The expression for the linearized system in (5.24) is unaffected, but the actuator stiffness<br />
ca(z ′ a) is given by c ′ a(z ′ a) in (5.27) <strong>and</strong> the motor constant Ka(z ′ a) <strong>and</strong> inductance La(z ′ a)<br />
becomeK ′ a (z′ a ) <strong>and</strong>L′ a (z′ a ) respectively, where:<br />
K ′ a(z ′ a) = −N (NIa −Hcmlm)<br />
(ℜ ′ 1 (z′ a)) 2 , L ′ a(z ′ a) = N2<br />
ℜ ′ 1 (z′ a) .<br />
Both the motor constant <strong>and</strong> the inductance are thus inversely proportional to the total reluctance<br />
ℜ ′ 1(z ′ a). Since the motor constant is inversely proportional to the square of this<br />
reluctance, a modest reduction in reluctance will lead to a significant improvement of the<br />
motor constant.<br />
5.2.5 Measurements <strong>and</strong> validation<br />
Several prototypes of single variable reluctance actuators have been manufactured <strong>and</strong> measurements<br />
have been performed to determine their (dynamic) behavior. The quasi-static<br />
behavior of the actuator is governed by the force-displacement characteristic as derived in<br />
(5.15), whereas its <strong>dynamics</strong> will be analyzed in terms of the parameters of the linearized<br />
system in (5.24) <strong>and</strong> its mechanical resonance frequency. The measurements were performed<br />
using the test setup as depicted in figure 5.13, consisting of a Siglab TM [35] system<br />
<strong>and</strong> a Polytec laser vibrometer. The Siglab system was used to generate an excitation signal<br />
that was fed to a current amplifier <strong>and</strong> also measured back at one of the Siglab inputs. The<br />
laser vibrometer was used to measure the velocity ˙za of the moving core <strong>and</strong> its output – an<br />
analog voltage – was also fed to a Siglab input. The sensitivity of the laser vibrometer was<br />
set to 25mm/s/V with an output range of -10...10V.<br />
To measure Frequency Response Functions (FRFs), a wide b<strong>and</strong>, white noise excitation<br />
signal was used with an Root Mean Square (RMS) level of ≈1.5mA <strong>and</strong> DC offsets ∆Ia<br />
varying between -15 <strong>and</strong> 15mA. The FRFs are estimated using the Siglab software with a<br />
Hanning window of 8192 samples <strong>and</strong> 50 averages without overlap. The results are shown<br />
in figure 5.14. Observe that the resonance frequency of the actuator lies around 2.1kHz <strong>and</strong><br />
varies approximately 2Hz per mA current offset.<br />
5
5<br />
110 5 The variable reluctance actuator<br />
Magnitude [m/s/A]<br />
Phase [deg]<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
90<br />
0<br />
∆ I a = −14mA<br />
∆ I a = −7mA<br />
∆ I a = 0mA<br />
∆ I a = 7mA<br />
∆ I a = 14mA<br />
−90<br />
1000 1500 2000 2500 3000<br />
Frequency [Hz]<br />
Figure 5.14: Measured frequency response functions<br />
of a single actuator for various current offsets∆Ia.<br />
10 1<br />
10 0<br />
Resonance frequency [kHz]<br />
Actuator stiffness [N/mm]<br />
Motor constant [N/A]<br />
Damping [mNs/m]<br />
10<br />
−15 −10 −5 0 5 10 15<br />
−1<br />
DC current offset ∆I [mA]<br />
a<br />
Figure 5.16: The viscous damping ba, resulting<br />
actuator stiffnessca, motor constantKa <strong>and</strong> undamped<br />
mechanical resonance frequency fe as a<br />
function of the DC current offset I ′ a as identified<br />
from measurement data.<br />
Magnitude [m/s/A]<br />
Phase [deg]<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
90<br />
0<br />
∆ I a = −14mA<br />
∆ I a = −7mA<br />
∆ I a = 0mA<br />
∆ I a = 7mA<br />
∆ I a = 14mA<br />
−90<br />
1000 1500 2000 2500 3000<br />
Frequency [Hz]<br />
Figure 5.15: Bode plots of the modeled transfer<br />
functions of a single actuator for various current<br />
offsets∆Ia.<br />
10 1<br />
10 0<br />
Resonance frequency [kHz]<br />
Actuator stiffness [N/mm]<br />
Motor constant [N/A]<br />
10<br />
−15 −10 −5 0 5 10 15<br />
−1<br />
DC current offset ∆I [mA]<br />
a<br />
Figure 5.17: The modeled actuator stiffness ca,<br />
motor constant Ka <strong>and</strong> resonance frequency fe<br />
as a function of the DC current offsetI ′ a.<br />
From the measured FRFs it is also possible to estimate the viscous damping, actuator stiffness<br />
<strong>and</strong> motor constant. This step requires the linearized model expressed in (5.24) to be<br />
adapted from voltage to current input, as used in the test setup. When the current Ia is<br />
prescribed, the first state (Ia) of the system vanishes <strong>and</strong> the system is determined by the<br />
equations for the acceleration ¨za <strong>and</strong> the velocity ˙za. This yields the following, second<br />
order state-space system:<br />
<br />
˙za 0 1<br />
=<br />
¨za −ca(z ′ a)/mac −ba/mac<br />
za<br />
˙za<br />
<br />
<br />
+<br />
0<br />
Ka(z ′ a)/mac<br />
<br />
Ia(t),
5.2 The single actuator 111<br />
where ˙za is chosen as the system output corresponding to the type of laser vibrometer measurement<br />
used. This system can be rewritten into transfer function form as:<br />
HI(s,z ′ Ka(z<br />
a ) =<br />
′ a )s<br />
macs2 +bs+ca(z ′ a )<br />
(5.28)<br />
<strong>and</strong> has an undamped mechanical resonance frequency fe that depends on the operating<br />
pointz ′ a :<br />
fe(z ′ <br />
1<br />
a ) =<br />
2π<br />
ca(z ′ a )<br />
To obtain estimates for Ka(z ′ a ), ba <strong>and</strong> ca(z ′ a ), first a parametric identification on the FRF<br />
will be performed using Matlab’sÒÚÖÕ×function. However, to be able to derive the<br />
desired properties from the estimated coefficients, two modifications must be made to the<br />
above transfer function. Firstly, note that the lowest term of the numerator polynomial<br />
is zero. To prevent the need for a constraint in the parametric identification procedure, the<br />
numerator is divided bys. This corresponds to time domain integration that must be applied<br />
to the measured velocity signal prior to parametric identification. The second change is due<br />
to the fact that of the four coefficients only three can be uniquely identified <strong>and</strong> a fourth<br />
must be given. Since the value of mac is well defined by manufacturing tolerances, this is<br />
further assumed to be known as listed in table 5.1. The transfer function whose coefficients<br />
will be estimated can thus be expressed as:<br />
mac<br />
.<br />
Ka(z ′ a )/mac<br />
HI(s,z ′ a ) =<br />
s2 +(ba/mac)s+ca(z ′ a )/mac<br />
. (5.29)<br />
The Matlab functionÒÚÖÕ×is then used to estimate the three unknown coefficients from<br />
whichKa(z ′ a ),ba <strong>and</strong>ca(z ′ a ) are determined. The obtained values are plotted together with<br />
the damped resonance frequency corresponding to the poles of (5.29) in figure 5.16.<br />
The following observations can be made by comparing the values obtained from the model<br />
in figure 5.17 <strong>and</strong> from the measurements in figure 5.16:<br />
• the measured resonance frequency is higher than modeled <strong>and</strong> since the moving mass<br />
is known, this implies that the actuator stiffness must be higher,<br />
• the stiffness <strong>and</strong> resonance frequency decrease with an increased axial air gap,<br />
• the measured motor constant is higher than modeled,<br />
• the measured viscous damping depends on the position of the core.<br />
To explain the differences found, a sensitivity analysis of the variable reluctance actuator is<br />
performed.<br />
5.2.6 Sensitivity analysis<br />
Figure 5.18 shows the sensitivity of the resonance frequencyfe, actuator stiffnessca, motor<br />
constant Ka <strong>and</strong> inductance La of the actuator w.r.t. the height h <strong>and</strong> the stiffness coefficients<br />
C1 <strong>and</strong> C2. This figure is obtained by evaluating the expressions for fe, Ka, ca <strong>and</strong><br />
La derived in the previous subsections while varying a single actuator property <strong>and</strong> keeping<br />
all others at their nominal values as listed in table 5.1. These values are marked by the thick,<br />
5
5<br />
112 5 The variable reluctance actuator<br />
10 0<br />
10 −1<br />
10 −2<br />
L a [mH]<br />
f e [kHz]<br />
c a [kN/m]<br />
K a [N/A]<br />
z 0 [mm]<br />
180 200 220 240 260 280 300<br />
Height h [µm]<br />
−0.1 0 0.1<br />
Stiffness coëficient C 1 [−]<br />
0.005 0.01 0.015 0.02 0.025<br />
Nonlinear stiffness coëficient C 2 [−]<br />
Figure 5.18: Sensitivity of the resonance frequency fe, resulting stiffness ca, motor constant Ka <strong>and</strong><br />
inductance La of the actuator w.r.t. the height h <strong>and</strong> the stiffness constants C1 <strong>and</strong> C2.<br />
The thick, dashed vertical line represents the nominal value of the parameter, as listed<br />
in table 5.1.<br />
dashed vertical lines. The results include the effect of the parameters on the initial air gap<br />
z0 on which all expressions implicitly depend. From figure 5.18 the following remarks are<br />
made:<br />
The heighth.<br />
An increase of the height h causes an increase of the initial air gap <strong>and</strong> a working<br />
point with less stiffness <strong>and</strong> lower motor constant.<br />
The stiffness coefficientC1.<br />
A change in the linear stiffness coefficientC1 of the membrane suspension will result<br />
in a change in initial air gap z0. The magnetic force depends on the reluctance of<br />
the magnetic circuit <strong>and</strong> therefore on the position of the moving core. In the equilibrium<br />
position, the spring force is equal to this magnetic force. The spring force is<br />
proportional to the deflection <strong>and</strong> the linear <strong>and</strong> nonlinear stiffness of the membrane<br />
suspension. If the linear stiffness coefficient decreases, the initial air gap z0 will decrease<br />
<strong>and</strong> the contribution of the nonlinear stiffness on the actuator stiffness will<br />
increase. This effect is attenuated by the stiffness due to the magnetic circuit which<br />
also increases for smaller air gaps. However, this effect is linear with the decrease air
5.2 The single actuator 113<br />
gap whereas the effect on the mechanical spring stiffness is quadratic. The net effect<br />
is an increase in actuator stiffness <strong>and</strong> resonance frequency.<br />
For a smaller air gap the magnetic reluctance drops, causing the magnetic flux, hence<br />
the forceFm <strong>and</strong> the motor constantKa to increase.<br />
The stiffness coefficientC2.<br />
A decrease in the nonlinear stiffness coefficientC2 of the membrane suspension will<br />
also result in a smaller initial air gap. In analogy with the linear stiffness coefficient<br />
C1 this leads to a larger deflection to maintain force equilibrium, but this deflection<br />
increase is accompanied by a lower spring stiffness <strong>and</strong> will therefore – in contradiction<br />
to the stiffness coefficientC1 – result in a lower overall stiffness.<br />
Again with the smaller air gap the magnetic reluctance drops <strong>and</strong> the motor constant<br />
increases.<br />
Figure 5.19 shows the sensitivity of the resonance frequencyfe, actuator stiffnessca, motor<br />
constantKa <strong>and</strong> inductanceLa w.r.t. the leakage flux reluctancesℜflc <strong>and</strong>ℜflm of the coil<br />
<strong>and</strong> the PM respectively <strong>and</strong> the radial air gap reluctanceℜga. The figure was obtained the<br />
same way as figure 5.18 <strong>and</strong> the thick, dashed vertical lines represent the parameter values<br />
listed in table 5.1. From figure 5.19 the following is observed:<br />
The coil leakage flux reluctanceℜflc.<br />
The leakage flux reluctance of the coil has no significant effect on the initial air gap<br />
z0, actuator stiffnessca or resonance frequencyfe. As the reluctanceℜflc of the coil<br />
leakage flux decreases to the order of the reluctance ℜbc of the coil core, the motor<br />
constant becomes affected. The major part of the flux generated by the coil will then<br />
flow into the leakage flux path.<br />
The coil inductanceLa decreases for increasingℜflc, since this inductance is proportional<br />
toN 2 / ˜ ℜ(za) <strong>and</strong> ˜ ℜ(za) is proportional toℜflc.<br />
The PM leakage flux reluctanceℜflm.<br />
If the reluctanceℜflm of the PM leakage flux is decreased, the attraction force on the<br />
ferromagnetic core decreases <strong>and</strong> the initial air gap will be larger. As illustrated by<br />
figure 5.8, a lower equilibrium force will result in a lower stiffness of the membrane<br />
suspension <strong>and</strong> resonance frequency. The motor constant Ka decreases when ℜflm<br />
decreases since the equilibrium force decreases <strong>and</strong> the air gap increases. As long as<br />
ℜflm does not significantly affect the total reluctance experienced by the coil, there<br />
is little change in coil inductance.<br />
The radial air gap reluctanceRga.<br />
The reluctance of the radial air gap forms a significant part of the total reluctance<br />
in the flux path through the axial air gap (φ1). This explains its significant effect of<br />
motor constant <strong>and</strong> initial air gapz0. When the total reluctance decreases, the flux, the<br />
force <strong>and</strong> the motor constant increase. The decrease of the axial air gap z0 explains<br />
the increase in actuator stiffness <strong>and</strong> resonance frequency.<br />
Figure 5.20 shows the sensitivity of the resonance frequencyfe, actuator stiffnessca, motor<br />
constantKa <strong>and</strong> inductanceLa of the actuator w.r.t. the axial air gap areaAga, the coercivity<br />
of the PM Hcm <strong>and</strong> the PM thickness lm. The figure was obtained the same way as figure<br />
5
5<br />
114 5 The variable reluctance actuator<br />
10 0<br />
10 −1<br />
10 −2<br />
L a [mH]<br />
f e [kHz]<br />
c a [kN/m]<br />
K a [N/A]<br />
z 0 [mm]<br />
10 7<br />
10 8<br />
10 9<br />
Coil leak flux reluctance R [1/H]<br />
flc<br />
10 7<br />
10 8<br />
10 9<br />
10<br />
Magnet leak flux reluctance R [1/H]<br />
flm 7<br />
10 8<br />
10 9<br />
Radial air gap reluctance R [1/H]<br />
ga<br />
Figure 5.19: Sensitivity of the resonance frequency fe, resulting stiffness ca, motor constant Ka <strong>and</strong><br />
inductance La of the actuator w.r.t. the leakage flux reluctances of the coil ℜflc <strong>and</strong><br />
the PM ℜflm <strong>and</strong> the radial air gap reluctance ℜga. The thick, dashed vertical line<br />
represents the parameter value listed in table 5.1.<br />
5.18 <strong>and</strong> the thick, dashed vertical lines represent the parameter values listed in table 5.1.<br />
From figure 5.20 the following is observed:<br />
The axial air gap areaAga.<br />
Observe that the area Aga is present in the expression for the magnetic force Fm in<br />
(5.12) <strong>and</strong> thus directly influences the initial air gapz0. However,Aga also affects the<br />
actuator properties indirectly via the reluctanceℜga. This reluctance enters quadratically<br />
in (5.12) <strong>and</strong> becomes the dominant term for small values ofAga. This leads to<br />
an increased magnetic force <strong>and</strong> a reduction of the nominal air gap width that explains<br />
the minimum in the graph for z0. The initial air gap also affects the motor constant,<br />
leading to a maximum in the graph for Ka that corresponds to the minimum for the<br />
air gap z0. The actuator stiffness decreases as z0 increases for small values of Aga.<br />
For larger values of Aga, the stiffness ca is dominated by the nonlinear mechanical<br />
spring stiffness that decreases as the air gap z0 increases. The decreased negative<br />
magnetic stiffness will not compensate for the mechanical spring stiffness reduction<br />
<strong>and</strong> an overall decrease in stiffness results.<br />
The magnetic field intensityHcm of the PM.
5.2 The single actuator 115<br />
10 0<br />
10 −1<br />
10 −2<br />
L a [mH]<br />
f e [kHz]<br />
c a [kN/m]<br />
K a [N/A]<br />
z 0 [mm]<br />
0.5 1 1.5<br />
Axial air gap area A ga [mm 2 ]<br />
−650 −600 −550 −500 −450<br />
PM coercive force H cm [kA/m]<br />
250 300 350<br />
PM thickness l m [µm]<br />
Figure 5.20: Sensitivity of the resonance frequency fe, resulting stiffness ca, motor constant Ka <strong>and</strong><br />
inductance La of the actuator w.r.t. the axial air gap radius rga, <strong>and</strong> the coercivity of<br />
the PM Hcm <strong>and</strong> thickness lm of the PM. The thick, dashed vertical line represents the<br />
parameter value listed in table 5.1.<br />
If the coercive force of the PM is increased the magnetic force increases. This increase<br />
leads to a smaller air gap with a higher actuator stiffness, resonance <strong>and</strong> motor<br />
constant.<br />
The thicknesslm of the PM.<br />
A thickness increase of the PM has the same effect as a coercive force increase.<br />
Besides sensitivity of the actuator properties fe, ca, Ka <strong>and</strong> La on the parameters h, C1,<br />
C2, ℜflc, ℜflm, ℜga, Aga, Hcm <strong>and</strong> lm, it is relevant to know what causes the possible<br />
differences in these parameters <strong>and</strong> how to minimize them. First of all, the predictability of<br />
the actuator properties is increased by the measurements performed to identify the coercivity<br />
Hcm of the PMs <strong>and</strong> the stiffness coefficients C1 <strong>and</strong> C2 of the membrane suspension.<br />
Besides this, the manufacturing tolerances <strong>and</strong> assembly tolerances play an important role:<br />
Manufacturing tolerances.<br />
All dimensions of the actuator are subject to manufacturing tolerances. Manufacturing<br />
tolerances are typically in the order of tens of µm. These dimensions directly<br />
determine all modeled magnetic reluctances, the mechanical stiffness coefficientsC1<br />
<strong>and</strong> C2 <strong>and</strong> the moving mass mac. Besides dimensions, the magnetic permeability<br />
5
5<br />
116 5 The variable reluctance actuator<br />
of the ARMCO can be influenced by stresses caused by the manufacturing process.<br />
This is likely to play a role for the reluctances of the baseplate, the moving core <strong>and</strong><br />
the radial air gap. As a result, all actuator properties will be affected <strong>and</strong> vary from<br />
actuator to actuator.<br />
Assembly tolerances.<br />
In addition to tolerances on the manufactured parts themselves, the design dimensions<br />
are affected by the assembly process. For instance, the thickness of the glue layers between<br />
(1) the PM <strong>and</strong> the coil insert <strong>and</strong> (2) between the baseplate <strong>and</strong> the membrane<br />
suspension <strong>and</strong> (3) between the moving core <strong>and</strong> the suspension membrane lead to<br />
a variation of the height h. Another example is the reluctance of the radial air gap,<br />
which is affected by the in-plane alignment of the moving core w.r.t. the pole shoes.<br />
Tolerances on the assembly process used are expected to be typically in the order of<br />
ten µm. In section 5.3 it will be shown how – by design <strong>and</strong> assembly – the effect of<br />
manufacturing <strong>and</strong> assembly tolerances on the actuator behavior is minimized.<br />
In the next section, the results from the modeled actuator (figure 5.17) <strong>and</strong> the measured<br />
actuator (figure 5.16) will be compared <strong>and</strong> analyzed based on insights of the sensitivity<br />
analysis. newpage<br />
5.2.7 Lessons learned<br />
Actuator stiffness ca <strong>and</strong> resonance frequencyfe<br />
The actuator stiffness ca is directly coupled to the mechanical resonance frequency<br />
fe via the mass mac. The stiffness values derived from the measurements are higher<br />
than expected from the model. With the use of figures 5.18 <strong>and</strong> 5.19 it is shown that<br />
the stiffness ca varies significantly for all considered parameters except for the coil<br />
leakage flux reluctance. This is caused mainly by its dependence on the initial air<br />
gap z0, which is affected by all parameters. Note from figure 5.9 that in general the<br />
stiffness increases asz0 decreases. Variation of parameters that lead to an increase in<br />
z0 will therefore lead to a decrease inca.<br />
The same effect is observed in the variation w.r.t. the operating point. In accordance<br />
to the model, the stiffness is found to change with the deflectionza (orIa): it increases<br />
for negativeza <strong>and</strong> decreases for positiveza.<br />
Motor constantKa<br />
The measured value (≈ 0.28N/A) for the motor constantKa is higher than the original<br />
design value of≈0.2N/A. As can be observed in figures 5.18 <strong>and</strong> 5.19, the motor<br />
constant shows considerable variation w.r.t. all analyzed parameters except the coil<br />
leakage flux reluctance ℜflc. Most noticeable is the strong dependence of the radial<br />
air gap reluctance ℜga. A higher motor constant can therefore be partially attributed<br />
to a lower reluctance of the radial air gap.<br />
Finally, note from figure 5.20 that there exists a value for the axial air gap area Aga<br />
where the motor constant has a maximum. This may be exploited in future designs to<br />
obtain a higher power efficiency.<br />
InductanceLa<br />
As can be observed in (5.25), the inductance La of the actuator is a function of the
5.3 The actuator module 117<br />
number of turns N in the coil <strong>and</strong> the reluctances in the magnetic circuit. Consequently,<br />
only parameters that have a significant effect on the total reluctance will<br />
cause a significant change in La. The dominant reluctances are ℜga(za) of the axial<br />
air gap <strong>and</strong>ℜgr of the radial air gap <strong>and</strong>ℜm of the PM.<br />
Dampingba<br />
In the model, the damping is considered a constant, but the measurement results in<br />
figure 5.16 indicate that the damping varies with the deflection of the moving core.<br />
A possible explanation for this effect is that a so-called squeeze film exists between<br />
the PM <strong>and</strong> the moving core. When the core moves, air is either expelled from or<br />
compressed between the two objects. Viscosity hampers the flow of air, which leads<br />
to both spring <strong>and</strong> damper behavior that depends on the distance between the two<br />
objects <strong>and</strong> the relevant time scale (i.e. frequency of motion) [114].<br />
Although differences between the model <strong>and</strong> measurements exist <strong>and</strong> the analysis results<br />
indicate that improvement of the motor constant in particular is well possible, the results are<br />
good enough to proceed with design <strong>and</strong> integration of actuators in grids.<br />
The design <strong>and</strong> realization of these actuator modules will be introduced in the next section.<br />
Measurement results, including the variation <strong>and</strong> spread in actuator properties, on all<br />
actuators of seven prototype grids will be presented in section 5.3.1.<br />
5.3 The actuator module<br />
In the transformation of the single actuator design into a grid actuator module, the philosophy<br />
is to design in layers that extend over many actuators <strong>and</strong> not to make many individual<br />
actuators that need to be placed <strong>and</strong> aligned individually. This reduces the number of parts<br />
<strong>and</strong> the complexity of assembly <strong>and</strong> improves the uniformity of the actuator properties.<br />
A hexagonal actuator layout is chosen since this gives the highest actuator areal density.<br />
The grids are also given a hexagonal shape to accommodate the assembly of large DMs<br />
from many actuator grids. The grid layout consists of a central actuator surrounded by a<br />
number of hexagonal ’rings’ of actuators. This approach results in a total of 7, 19, 37, 61,<br />
91 or 127 actuators for 1, 2, 3, 4, 5 or 6 rings respectively. For the prototype grids realized,<br />
four rings, corresponding to 61 actuators, are chosen. In this choice, several practical issues<br />
were taken into account, such as the size of the baseplate <strong>and</strong> its corresponding resonance<br />
frequency. A larger grid would require more thickness, or additional support points to avoid<br />
internal resonances. Further, the actuator coils are connected through flex foils. These will<br />
become more difficult to manufacture as the number of actuators per grid increases. When<br />
the line pattern is made on a single layer there is not enough area for more connections.<br />
Finally, the number 61 is close to a division factor of2 6 = 64, which is likely to be used in<br />
digital electronics for drivers <strong>and</strong> communication. The actuator grid is shown in figures<br />
5.21 <strong>and</strong> 5.22. Figure 5.23 shows the exploded view of the actuator grid. The main parts<br />
shown in the exploded view are summarized <strong>and</strong> will be discussed in detail:<br />
The baseplate<br />
The baseplate serves a the flux carrier for the magnetic circuits for 61 actuators <strong>and</strong><br />
is made from from ARMCO . The baseplates are cut from bar <strong>and</strong> their front <strong>and</strong><br />
5
5<br />
118 5 The variable reluctance actuator<br />
Figure 5.21: The st<strong>and</strong>ard actuator module with<br />
61 actuators seen from the front.<br />
Figure 5.22: The st<strong>and</strong>ard actuator module with<br />
61 actuators seen from the back.<br />
back surfaces are made plan parallel. The serrated circumference is made such that<br />
the actuator grids can be placed adjacent with a 0.3mm gap. The holes in the backside<br />
are made by milling, the circumference <strong>and</strong> pole shoe contours are made with wire<br />
Electrical Discharge Machining (EDM).<br />
The membrane suspension with moving cores<br />
The membrane suspension is made by laser cutting. The same sheet of rolled Havar <br />
material is used for all actuator grids in the same orientation to obtain uniformity for<br />
the spring characteristic. The moving cores are laser cut <strong>and</strong> made from a 0.3mm<br />
thick ARMCO sheet.<br />
Inserts with the SmCo5 PMs <strong>and</strong> coils<br />
The inserts complete the magnetic circuit. The inserts are produced on a Computer<br />
Numerical Control (CNC) lathe. After measurement <strong>and</strong> selection of the PMs described<br />
in [174], the PMs are glued on the inserts with Araldite 2020. The coils,<br />
made of 50µm copper wire with 500 turns, are fabricated separately from the insert<br />
<strong>and</strong> have pre-leaded ends. The electrical resistance of each coil is measured before<br />
placement. The inner radius of the coil is made slightly larger than the insert core<br />
to avoid damage to the electrical isolation when placed. The bottom of the inserts<br />
contains a hole <strong>and</strong> a slot to provide an axial feed through for the coil wires.<br />
The flex foil<br />
A flex foil is designed to connect the coil wires to the driver Printed Circuit Board<br />
(PCB). This flex foil design is shown in figure 5.22. Each of the three branches of<br />
the flex foil connects to a double row, 0.3mm pitch, connector. The central hexagonal<br />
part of the flex foil holds∅2.5mm holes through which the coil wires emerge. At the<br />
circumference of the holes, the two wires of that coil are soldered each on a copper<br />
pad. After soldering, a droplet of silicon glue is placed to encapsulate the fragile<br />
coil wires. A strain relief (figure 5.22) is added to avoid damage to the soldered<br />
connections.<br />
The actuator grid supports
5.3 The actuator module 119<br />
M1x6 (3x)<br />
Membrane suspension<br />
Moving cores (61x)<br />
Baseplate<br />
Coils (61x)<br />
SmCo Magnets (61x)<br />
Inserts (61x)<br />
Flex foil<br />
A-frames (3x)<br />
Strain relief <strong>and</strong> clamps (3x)<br />
Figure 5.23: An exploded view of the actuator grid shown in figures 5.21 <strong>and</strong> 5.22.<br />
5
5<br />
120 5 The variable reluctance actuator<br />
Figure 5.24: Bode plot of<br />
the measured FRFs s · H(za)<br />
from a single actuator in a<br />
grid, together with the fitted<br />
model <strong>and</strong> the original analytic<br />
model. The measured FRF of<br />
the single actuator as shown in<br />
figure 5.14 is also plotted for<br />
comparison.<br />
Magnitude [m/s/A]<br />
Phase [deg]<br />
10 3<br />
10 2<br />
10 1<br />
10 0<br />
10 −1<br />
10 −2<br />
180<br />
90<br />
0<br />
−90<br />
−180<br />
10 1<br />
Grid actuator measurement<br />
Fitted model<br />
Original model<br />
Single actuator measurement<br />
10 2<br />
Frequency [Hz]<br />
The actuator grid is connected with three A-frames to its support structure. Each<br />
A-frames is connected with one point to the baseplate <strong>and</strong> with two points to the<br />
support structure to avoid moments enforced on the baseplate. The A-frames are<br />
connected with M1 bolts. When the actuator grid is placed facedown on the table,<br />
with the membrane suspensions facing down, the bolt heads support the actuator grid<br />
<strong>and</strong> avoid damage to the suspension systems.<br />
Details on the manufacturing <strong>and</strong> assembly procedures of the actuator grids can be found in<br />
[174].<br />
5.3.1 Measurement results<br />
A batch of seven grids was realized using the procedure described in [174]. The only difference<br />
in nominal dimensions with respect to the single actuator is an additional 25µm for the<br />
height h. Although this is known to have a negative effect on the motor constant, actuator<br />
stiffness <strong>and</strong> resonance frequency, more margin is hereby built in for manufacturing errors<br />
<strong>and</strong> the risk on ferromagnetic cores to snap down on the PMs is avoided. The reduced motor<br />
constant, actuator stiffness <strong>and</strong> resonance frequency are estimated from figure 5.18 as<br />
Ka=0.17 N/A, ca=550N/m <strong>and</strong>fe=1980 Hz.<br />
For each grid all actuators are measured with the same experimental setup as shown in figure<br />
5.13. Figure 5.24 shows a typical transfer function with a parametric model fit of one of<br />
the actuators in the grid. As a reference, the figure also shows the nominal transfer function<br />
derived from the analytical model depicted previously in figure 5.15. Note that because<br />
of the velocity measurement of the experimental setup, the response for frequencies below<br />
the first resonance shows a +1 slope <strong>and</strong> a -1 slope above the resonance. Besides the first<br />
resonance frequency around 2kHz, a second resonance is visible at approximately 5kHz.<br />
This is the tip/tilt mode of the ferromagnetic core in its membrane suspension that can only<br />
be observed when the laser vibrometer is not perfectly aligned to the center of the moving<br />
core.<br />
10 3<br />
10 4
5.3 The actuator module 121<br />
Table 5.2: Average values <strong>and</strong> st<strong>and</strong>ard deviations of the actuator properties measured over all grid<br />
actuators.<br />
Property Ka ca ba fe<br />
Average 0.12N/A 471N/m 0.36mNs/m 1.83kHz<br />
Std.dev. 0.02N/A 48N/m 0.10mNs/m 95Hz<br />
Estimates for the actuator stiffness ca, resonance frequency fe, motor constant Ka <strong>and</strong><br />
viscous damping ba are obtained from the measurement data using the same procedure as<br />
described in section 5.2.5. The average <strong>and</strong> st<strong>and</strong>ard deviation values are listed in table<br />
5.2 <strong>and</strong> more detailed results are shown in figures 5.25, 5.26, 5.27 <strong>and</strong> 5.28. The values in<br />
the figures are sorted to provide insight into the statistical spread <strong>and</strong> differences in median<br />
values between grids. The mean actuator stiffness <strong>and</strong> resonance frequency are 471N/m <strong>and</strong><br />
1.83kHz respectively, which is lower than expected. In figure 5.25 it is shown that the spread<br />
in stiffness values within a grid is similar for all grids, but that the mean differs from grid to<br />
grid. This can be caused by manufacturing <strong>and</strong> assembly variations that directly affect all<br />
actuators in a module, such as baseplate or suspension membrane thickness variations.<br />
Figure 5.27 shows the values of the motor constants, which are lower than expected. The<br />
analytic model predictedKa=0.17N/A, whereas 0.12N/Ais measured. A possible explanation<br />
for the measurement results is that the leakage flux reluctances for the PM <strong>and</strong> coil are<br />
smaller in the actuator module baseplate than for the single actuator. The few high values<br />
for damping are explained by rests of glue in between the moving core <strong>and</strong> baseplate.<br />
5.3.2 Power dissipation<br />
To analyze the expected power dissipation of the actuator, only its static response is considered<br />
<strong>and</strong> assumed to be linear. The validation measurements have shown this to be an<br />
Actuator stiffness c a [N/m]<br />
650<br />
600<br />
550<br />
500<br />
450<br />
400<br />
350<br />
300<br />
Module 1<br />
Module 2<br />
Module 3<br />
Module 4<br />
Module 5<br />
Module 6<br />
Module 7<br />
10 20 30 40 50 60<br />
Actuator number [−]<br />
Figure 5.25: The identified actuator stiffnessca,<br />
for mac =3.6mg sorted for each measured actuator<br />
grid separately.<br />
Resonance frequency F e [Hz]<br />
2100<br />
2000<br />
1900<br />
1800<br />
1700<br />
1600<br />
1500<br />
Module 1<br />
Module 2<br />
Module 3<br />
Module 4<br />
Module 5<br />
Module 6<br />
Module 7<br />
10 20 30 40 50 60<br />
Actuator number [−]<br />
Figure 5.26: The identified actuator resonance<br />
fe, for mac =3.6mg sorted for each measured<br />
actuator grid separately.<br />
5
5<br />
122 5 The variable reluctance actuator<br />
Motor constant K a [N/A]<br />
0.2<br />
0.18<br />
0.16<br />
0.14<br />
0.12<br />
0.1<br />
0.08<br />
0.06<br />
Module 1<br />
Module 2<br />
Module 3<br />
Module 4<br />
Module 5<br />
Module 6<br />
Module 7<br />
10 20 30 40 50 60<br />
Actuator number [−]<br />
Figure 5.27: The fitted motor constants Ka, for<br />
mac =3.6mg sorted for each measured actuator<br />
grid separately.<br />
Viscous damping b a [mNs/m]<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
Module 1<br />
Module 2<br />
Module 3<br />
Module 4<br />
Module 5<br />
Module 6<br />
Module 7<br />
10 20 30 40 50 60<br />
Actuator number [−]<br />
Figure 5.28: The fitted damping ba, for<br />
mac =3.6mg sorted for each measured actuator<br />
grid separately.<br />
accurate approximation at least up to approximately 1600Hz, where effects such as viscous<br />
damping <strong>and</strong> eddy currents play a negligible role.<br />
The power dissipated by the currentIa through the actuator coil with resistance Ra can be<br />
expressed as:<br />
Pa = I 2 aRa <br />
Fa<br />
=<br />
Ka<br />
2<br />
Ra,<br />
where for the second step the linearized case Fa = KaIa was used. In chapter 2 the<br />
expected RMS actuator force was derived as 1mN, which was based on an actuator stiffness<br />
ca of 500N/m. Since the stiffness of the realized actuator was found to be very close to this<br />
value, the 1mN RMS force remains a valid basis to evaluate power dissipation. For 1mN<br />
RMS actuator force <strong>and</strong> a measured average motor constant Ka of 0.12N/A this yields an<br />
RMS actuator current Ia of approximately 8mA. Via the coil resistance Ra ≈ 39Ω, this<br />
corresponds to a power dissipation of approximately 3mW per actuator. This means that the<br />
realized actuators meet the specification in chapter 2 that the power dissipation must be in<br />
the order of mW’s.<br />
5.4 Conclusions<br />
The application of electromagnetic actuators for adaptive <strong>deformable</strong> <strong>mirror</strong>s has several<br />
advantages. Electromagnetic actuators can be designed with a limited stiffness, such that<br />
failure of an actuator will cause no hard point in the reflective surface <strong>and</strong> thus a small optical<br />
degradation compared to e.g. stiff piezo-electric elements. Other advantages over the<br />
latter type of actuators at the relatively low cost, low driving voltages <strong>and</strong> negligible hysteresis<br />
<strong>and</strong> drift.<br />
Variable reluctance type actuators were primarily chosen because of their high efficiency<br />
<strong>and</strong> low moving mass. A nonlinear mathematical model of the actuator was derived describing<br />
both its static <strong>and</strong> dynamic behavior based on equations from the magnetic, mechanic<br />
<strong>and</strong> electric domains. The model was linearized, leading to expressions for the actuator
5.5 Recommendations 123<br />
transfer function <strong>and</strong> linear electromechanical properties such as motor constant, coil inductance,<br />
actuator stiffness <strong>and</strong> resonance frequency.<br />
Single actuator prototypes were realized <strong>and</strong> transfer functions were estimated from measurement<br />
data, based on white noise current excitation. This was done for various operating<br />
points by adding static current offsets to the excitation signal, which showed that the effect<br />
of the nonlinearities is indeed small. The resonance frequency <strong>and</strong> the DC-gain of the<br />
transfer functions showed only marginal variation with respect to the operating point. This<br />
means that a <strong>control</strong> system will be able to use an LTI <strong>control</strong> law without sacrificing performance.<br />
The measured nominal resonance frequency is higher than modeled: 2.1kHz instead<br />
of 2kHz, corresponding to an actuator stiffness of≈680N/m. The measured motor constant<br />
is also higher than modeled 0.27N/A instead of 0.2N/A.<br />
Due to the satisfactory measurement results, the design for single actuators was applied with<br />
little modification to the design of st<strong>and</strong>ard hexagonal modules with 61 actuators. Only the<br />
nominal heighth of the moving core above the PM was increased by 25µm to limit the risk<br />
of the ferromagnetic cores snapping down onto the PM. Based on the results of a sensitivity<br />
analysis, this modification was expected to reduce the motor constant, actuator stiffness <strong>and</strong><br />
resonance frequency toKa=0.17N/A,ca=550N/m <strong>and</strong>fe=1980Hz respectively.<br />
In this sensitivity analysis the effect was determined of variation in a number of model parameters<br />
on several actuator properties (i.e. stiffness, resonance frequency, motor constant<br />
<strong>and</strong> inductance) as derived from the mathematical model. In particular, this indicated a<br />
strong influence of the radial air gap reluctance on the motor constant <strong>and</strong> actuator stiffness,<br />
which can explain the deviation between the modeled <strong>and</strong> measured properties.<br />
Seven actuator module prototypes were made of which all actuators were measured with<br />
the same setup as the single actuator prototypes. All actuators were found to be functional,<br />
indicating that the manufacturing <strong>and</strong> assembly process is reliable. From transfer function<br />
measurements, the motor constant, actuator stiffness <strong>and</strong> resonance frequency were identified.<br />
These properties showed slight deviations from the values derived from the model,<br />
but the statistical spread for the properties was small, stressing the reliability of the manufacturing<br />
<strong>and</strong> assembly process. The mean actuator stiffness <strong>and</strong> resonance frequency were<br />
471N/m <strong>and</strong> 1.83kHz respectively, which are very close to their design values of 500N/m<br />
<strong>and</strong> 1885Hz. The value derived from the model for the motor constant Ka was 0.17N/A,<br />
whereas on average only 0.12N/A was measured. This may be the result of leakage flux<br />
reluctances in the baseplate being lower – <strong>and</strong> thus the leakage flux being larger – than for<br />
the single actuator.<br />
Despite the motor constants of the actuators realized being lower than expected, the RMS<br />
power dissipation of the actuators is still low during operation <strong>and</strong> expected to be≈3mW.<br />
5.5 Recommendations<br />
In a redesign, the large influence of the radial air gap reluctance can be used to increase<br />
the motor constant <strong>and</strong> reduce power dissipation. A factor of two reluctance reduction will<br />
increase the motor constant to 0.37N/A <strong>and</strong> increase the actuator stiffness to 750N/m. This<br />
can be realized by a smaller gap width or by a larger gap area.<br />
A reduction of the axial air gap area will lead to a further increase in motor constant. With<br />
these changes, an improvement by a factor four is feasible. Since power dissipation is<br />
5
5<br />
124 5 The variable reluctance actuator<br />
inversely proportional to the squared motor constant, a power dissipation reduction by a<br />
factor 16 is achieved. A convenient side effect is the increased electronic damping of the<br />
mechanical resonance frequency. This is illustrated by the Bode plots in figure 5.29, in<br />
which he magnitude peak of the two systems is almost equal, whereas the DC-gain of the<br />
improved system is four times higher. The relative damping of the resonance has increased<br />
approximately 2.6 times, making it less limiting for the <strong>control</strong>ler performance.<br />
Magnitude [m/V]<br />
Phase [deg]<br />
10 −4<br />
10 −5<br />
10 −6<br />
0<br />
−90<br />
−180<br />
−270<br />
10 2<br />
K a = 0.19N/A<br />
K a = 0.76N/A<br />
Frequency [Hz]<br />
Figure 5.29: Bode plot of the voltage to position actuator model with the currently<br />
measured <strong>and</strong> four times higher motor constant. Note the<br />
increased relative damping of the resonance mode.<br />
10 3
ÔØÖ×Ü<br />
ÐØÖÓÒ×<br />
This chapter describes the design <strong>and</strong> realization of the actuator driver electronics<br />
<strong>and</strong> communication system. Pulse Width Modulation (PWM) based<br />
voltage drivers are selected <strong>and</strong> implemented in three Field Programmable Gate<br />
Arrays (FPGAs) for 61 actuators (one actuator module). A high base frequency<br />
of 61kHz <strong>and</strong> an additional analog 2 nd order low-pass filter is used to reduce<br />
the actuator position response ripple due to harmonics of the PWM signal to<br />
less than a quarter of the Least Significant Bit (LSB) of the setpoint. The driver<br />
electronics of each actuator module are contained on a single Printed Circuit<br />
Board (PCB), which is placed behind the actuator module to preserve the <strong>modular</strong><br />
concept. The FPGAs receive the voltage setpoints via an Low Voltage<br />
Differential Signalling (LVDS) communication link from the <strong>control</strong> system.<br />
As no commercial LVDS interface board was available for a st<strong>and</strong>ard PC, an<br />
ethernet-to-LVDS communications bridge is developed that translates ethernet<br />
packages into LVDS packages <strong>and</strong> vice versa. A single flat-cable connects up<br />
to 32 driver PCBs to this communications bridge.<br />
To determine the PWM frequency <strong>and</strong> evaluate the dynamic effects of the electronics,<br />
the actuator model from chapter 5 is extended with models for the<br />
communication <strong>and</strong> driver electronics. The communication is modeled as a<br />
pure delay <strong>and</strong> the driver electronics as an ideal voltage source with a linear,<br />
analog 2 nd order low-pass filter. The dynamic model is validated using white<br />
noise identification measurements on the actuator system. The system is evaluated<br />
on <strong>control</strong> aspects, showing the dependance of achievable b<strong>and</strong>width on<br />
the sampling frequency. Finally, the power dissipation of the FPGAs is evaluated<br />
<strong>and</strong> found to exceed the power dissipation in the actuator coils. Concepts<br />
are proposed <strong>and</strong> analyzed to reduce power dissipation in the digital electronics.<br />
Joint work with Roger Hamelinck [PhD]<br />
125
6<br />
126 6 Electronics<br />
6.1 Introduction<br />
In this chapter, the design <strong>and</strong> realization of the electronics required to generate the currents<br />
through the coils of the actuators is described. The actuators were described in chapter 5<br />
The electronics consist of two parts: a communication system <strong>and</strong> driver electronics. The<br />
communication system transmits the current value calculated by the <strong>control</strong> system to the<br />
driver electronics, which generate the actual currents. The main difficulty of the driver electronics<br />
is the required number of channels combined with the desired high power efficiency.<br />
To determine the values of certain design properties of the electronics <strong>and</strong> evaluate the effects<br />
of the <strong>dynamics</strong> of the electronics on the <strong>Adaptive</strong> Optics (AO) system performance,<br />
a first principles model is derived that is combined with the actuator model from chapter 5.<br />
After the realization of the electronics, the behavior of the full actuator system is analyzed<br />
by comparing it to the model.<br />
In the next two sections design concepts for the driver <strong>and</strong> communication electronics will<br />
be presented, followed by their implementation <strong>and</strong> realization in section 6.4. The first<br />
principles model is derived in section 6.5 <strong>and</strong> compared to measurement results in section<br />
6.7. Since the driver electronics <strong>and</strong> the actuators are located close to the Deformable<br />
Mirrors (DMs) reflective surface, power dissipation is an important design driver. Therefore,<br />
in section 6.8 the power dissipation of the prototype actuator system is analyzed <strong>and</strong><br />
compared to the design requirements. Finally, the main conclusions will be formulated.<br />
6.2 Driver electronics<br />
The driver electronics provide currents through the coils of the variable reluctance actuators.<br />
In the actuator, the current converts into forces that deform the <strong>mirror</strong> facesheet. In this<br />
section the requirements, design concepts <strong>and</strong> implementation of the driver electronics will<br />
be discussed.<br />
6.2.1 Requirements<br />
In section 2.3 the required actuator positioning resolution is derived as 5nm. From this<br />
value, the worst-case required range <strong>and</strong> resolution will be derived of the currents <strong>and</strong> voltages<br />
that must be provided by the driver electronics. Since the resolution requirement is in<br />
terms of displacement, the mechanical stiffness, motor constant <strong>and</strong> circuit resistance must<br />
be considered. The mechanical stiffness depends on the resulting deformation of the reflective<br />
surface. The minimal stiffness is used to determine the required resolutions <strong>and</strong> the<br />
maximal stiffness is used to determine the required ranges. Inertia forces <strong>and</strong> other dynamic<br />
effects are neglected. The stiffness is minimal when all actuators have the same displacement<br />
<strong>and</strong> the reflective membrane does not deform. The worst case force resolution for the<br />
actuator stiffnessca =583N/m can thus be derived as Fres = 2.9µN. Via a motor constant<br />
Ka of 0.19N/A (figure 5.11) <strong>and</strong> an actuator coil resistance Ra of 39.0Ω (table 5.1), this<br />
leads to required current <strong>and</strong> voltage resolutions of 15µA <strong>and</strong> 0.60mV respectively.<br />
The required current <strong>and</strong> voltage range follows from the maximum actuator force, motor<br />
constant <strong>and</strong> coil resistance. An estimate for the actuator force is obtained by summing the<br />
forces required to overcome the internal mechanical stiffness of the actuator <strong>and</strong> to over-
6.2 Driver electronics 127<br />
come the stiffness of the facesheet:<br />
Fa = caza +cfzia<br />
where the <strong>mirror</strong> facesheet stiffness cf is ≈6kN/m ([174]), the actuator stroke za = 10µm<br />
<strong>and</strong> the inter-actuator stroke zia = 0.36µm as derived in chapter 2. This results in a<br />
range of the required force of ±8mN. Via the motor constant <strong>and</strong> actuator coil resistance<br />
previously used, this leads to the required current <strong>and</strong> voltage range of±42mA <strong>and</strong>±1.6V<br />
respectively.<br />
This implies that the dynamic range (the total range divided by the resolution) of the driver<br />
electronics must be at least 2 · 8 · 10 −3 /(2.9 · 10 −6 ) ≈ 5.5 · 10 3 . For a digital driver<br />
system, this would require at least 13 bits of accuracy. This is a minimum <strong>and</strong> does not<br />
provide any margin for a stiffer facesheet, initial flattening <strong>and</strong> alignment of the <strong>mirror</strong> <strong>and</strong><br />
variation in motor constant between actuators. To account for this <strong>and</strong> acknowledge that<br />
the digital implementation in 16 bits is likely to be more efficient, this is the starting point<br />
in the design. The consequences for the driver electronics from this requirement will be<br />
elaborated in section 6.2.2.<br />
The DM actuators are designed to have a high efficiency. In combination with the low<br />
actuator force requirement, this minimizes power dissipation. As a result no active cooling is<br />
needed <strong>and</strong> vibrations introduced by such systems are avoided. Through natural convection,<br />
heat is convected without adding significantly to the wavefront disturbances that are to be<br />
corrected. By placing the driver electronics near the electromagnetic actuators the number<br />
of wires to the DM can be limited <strong>and</strong> spacious connectors avoided. With the short length<br />
of the wires, sensitivity to environmental loads (e.g. lead breakage, magnetic fields from<br />
nearby power sources, etc.) is reduced. However, it also means that the power dissipation<br />
of the driver electronics must be in the order of mW’s, similar to the actuators.<br />
Finally, the electronics should be compact in size, have low cost/actuator <strong>and</strong> preferably<br />
replaceable. Since the actuator grid is made extendable by means of the st<strong>and</strong>ard modules<br />
that hold 61 actuators each, the same should hold for the electronics. All drivers for a single<br />
actuator module should therefore be placed on a single Printed Circuit Board (PCB).<br />
6.2.2 Concepts<br />
For the driver electronics concept, two categories are distinguished: current <strong>and</strong> voltage<br />
sources. A current driver converts the digital setpoint from the <strong>control</strong> computer into a <strong>control</strong>led<br />
current through the electromagnetic actuator. A voltage driver converts this setpoint<br />
into a voltage over the actuator clamps, upon which the circuit resistance determines the<br />
current. The current may therefore vary due to <strong>dynamics</strong> or time variance in the electronic<br />
circuit. TheLa/Ra time constant is approximately 2.93mH/39.0Ω ≈ 75µs (section 5.2.5),<br />
which is small compared to the intended sampling time of 1ms. Therefore, an applied voltage<br />
will result in a current without significant delay.<br />
The DC-gain of the actuator system including driver electronics depends on the actuator’s<br />
motor constant Ka, stiffness ca <strong>and</strong> – in the case of a voltage source – the resistance R of<br />
the actuator circuit. All three properties vary from actuator to actuator <strong>and</strong> vary with temperature,<br />
causing slow gain variations. A current source will compensate for variations in<br />
the resistanceRa, but variations ofKa <strong>and</strong>ca must still be compensated by the AO <strong>control</strong><br />
6
6<br />
128 6 Electronics<br />
I1<br />
b1<br />
b15<br />
I15<br />
b16<br />
!b16<br />
actuator<br />
!b16<br />
Figure 6.1: A design based on current sources that can be efficiently<br />
implemented in an ASIC using current <strong>mirror</strong>s. The<br />
"!" denotes a logicalÒÓØ.<br />
b16<br />
I ref<br />
I out<br />
Load<br />
V out<br />
Figure 6.2: Schematic of a single<br />
current <strong>mirror</strong>. The reference current<br />
Iref – here determined by a resistor<br />
– is <strong>mirror</strong>ed to the output current<br />
Iout.<br />
system. Therefore, the conceptual advantage of a current driver over a voltage driver is<br />
small, but concrete designs of both driver types will be discussed before a choice is made.<br />
Current <strong>mirror</strong>s<br />
In practice, a current source regulates its voltage output based on measurement of the actual<br />
current. This feedback can be done with a linear amplifier circuit, but for this application this<br />
approach has several drawbacks. Firstly, the circuit needs two supply voltages to generate<br />
positive <strong>and</strong> negative currents. Secondly, the required number of components is relatively<br />
high since no st<strong>and</strong>ard ASICs are available that compactly house a large number of accurate,<br />
efficient <strong>and</strong> low power linear amplifiers. Finally, the amplifier obtains its setpoint current<br />
from an analog voltage input that must be generated from the digital value of the <strong>control</strong><br />
system by an additional component such as a Digital to Analog Convertor (DAC).<br />
These drawbacks can be circumvented using a design based on current <strong>mirror</strong>s, which is<br />
schematically represented in figure 6.1. This design holds 15 current <strong>mirror</strong>s <strong>and</strong> can be efficiently<br />
implemented in CMOS technology. A single current <strong>mirror</strong> (figure 6.2) consists of<br />
two parts: in the first part a reference current is generated that is <strong>mirror</strong>ed with a certain ratio<br />
to the second part that includes the load. The reference currents of the current <strong>mirror</strong>s will<br />
be permanently flowing, whereas the <strong>mirror</strong>ed load currents can be switched according to<br />
the setpoint bits. The physical dimensions of the current <strong>mirror</strong>’s two transistors determine<br />
the ratios between the reference <strong>and</strong> load currents. These ratios can be designed to minimize<br />
the permanently flowing reference currents <strong>and</strong> thus optimize power efficiency. The total<br />
current can be constructed with 15 fixed current <strong>mirror</strong>s <strong>and</strong> a sign-switch corresponding<br />
to the Most Significant Bit (MSB) of the setpoint. The sign-switch is obtained using the<br />
full bridge configuration as depicted in figure 6.1, where the four switches are driven by the<br />
MSB of the setpoint <strong>and</strong> its inverse. This design is relatively simple, has a linear setpoint to<br />
current characteristic <strong>and</strong> many of these circuits can be implemented in a single Integrated<br />
Circuit (IC). However, it requires an ASIC, which is expensive to design <strong>and</strong> manufacture.<br />
Its power efficiency is comparable to that of linear amplifiers <strong>and</strong> dependent on the ratio<br />
between the Root Mean Square (RMS) <strong>and</strong> Peak To Valley (PTV) currents: the crest factor.<br />
The current sources regulate their output voltage, leading to internal voltage drops <strong>and</strong> thus
6.2 Driver electronics 129<br />
dissipation. When neglecting internal current paths of the current sources, the total power<br />
consumptionPtot can be expressed in terms of the supply voltageVcc <strong>and</strong> the desired actuator<br />
currentI asPtot = VccI. The powerPload dissipated in the load with resistanceRload<br />
is equal to Pload = I 2 Rload <strong>and</strong> can be expressed relative to Ptot as:<br />
Pload<br />
Ptot<br />
= I Rload<br />
Vcc<br />
= I<br />
.<br />
Imax<br />
This means that for low currents, almost all power is dissipated by the current source <strong>and</strong><br />
for the maximum currentImax = Vcc/Rload all power is consumed by the load.<br />
For the RMS current of 8mA (section 5.3.2), a load resistance of 39.0Ω <strong>and</strong> a supply voltage<br />
of 1.6V, this yields a power efficiency of approximately 20%. This low efficiency could be<br />
improved by adapting the supply voltageVcc in accordance with the desired actuator current<br />
by dividing the output range over a number of supply voltages. However, as this leads to an<br />
even more complex design that can only be realized in an ASIC, a Pulse Width Modulation<br />
(PWM) voltage driver with a higher power efficiency is considered as an alternative in the<br />
next section.<br />
Pulse Width Modulation<br />
The absence of a feedback path to regulate a current simplifies the design in comparison<br />
with linear amplifiers although this advantage is limited when compared to the current <strong>mirror</strong><br />
design. A regulated voltage source that generates an analog voltage for the actuator from<br />
the digital setpoint value is essentially a DAC. There exist many types of DACs, but the high<br />
accuracy <strong>and</strong> low power consumption required for this application limit the options. For instance,<br />
the low accuracy of the resistors of the common resistor ladder network DAC limits<br />
the useful accuracy of this type of converter to 8 bits or less.<br />
For high accuracy applications the PWM principle is often used, in which a digital output<br />
is modulated between high <strong>and</strong> low states to yield a desired average Direct Current (DC)<br />
value of the output voltage. The desired output voltage is translated into a duty cyclerPWM,<br />
which is the time fraction that the digital output is high during a certain time periodTPWM.<br />
This time period forms the base frequencyfPWM = 1/TPWM of the PWM.<br />
The advantage of a PWM based voltage source over the proposed <strong>mirror</strong> concept is twofold.<br />
Firstly the PWM generators can be implemented in Field Programmable Gate Arrays<br />
(FPGAs), which reduces the number of components <strong>and</strong> does not require the expensive<br />
<strong>and</strong> complex design <strong>and</strong> realization of an ASIC. Moreover, its power efficiency is superior<br />
to the current <strong>mirror</strong> driver because it has no internal voltage drop that leads to dissipation.<br />
Dissipation is limited to switching losses <strong>and</strong> indirect losses of the PWM signal generator<br />
<strong>and</strong> does not significantly depend on the desired output voltage. Finally, PWM design <strong>and</strong><br />
implementation is well understood, which limits development risks. A drawback of PWM<br />
is that it outputs a signal with high-frequency components, which causes a corresponding<br />
ripple on the system output. Using the Fourier series expansion shown in appendix C, for a<br />
constant duty cyclerPWM ∈ [0...1] the PWM output voltage ˜ VPWM that modulates between<br />
0 (low) <strong>and</strong>Vcc (high) can be expressed as the following infinite sum of cosines:<br />
˜VPWM(t) = Vcc<br />
<br />
rPWM +<br />
∞<br />
n=1<br />
sin(nπrPWM)<br />
nπ<br />
cos(2πnfPWMt)<br />
<br />
, (6.1)<br />
6
6<br />
130 6 Electronics<br />
Table 6.1: Properties of industrial communication st<strong>and</strong>ards as derived from data sheets of available<br />
driver ICs.<br />
St<strong>and</strong>ard B<strong>and</strong>width Multi-drop Predefined protocol<br />
USB2 480Mb/s no yes, high overhead<br />
FireWire 800 800Mb/s no yes, high overhead<br />
CAN 1Mb/s yes yes<br />
Gigabit Ethernet 1Gb/s no yes, high overhead<br />
RS-485 (Profibus) 40Mb/s yes no<br />
LVDS 655Mb/s a yes no<br />
a According to LVDS st<strong>and</strong>ard as defined in ANSI/TIA/EIA-644-A<br />
whereVcc is the switched supply voltage. The spectrum of this PWM signal thus only contains<br />
power at frequencies2πnfPWM forn = 1,2,...∞. These harmonic frequencies must<br />
be sufficiently attenuated by the <strong>dynamics</strong> of the driven system, such that the remaining ripple<br />
on the system’s output is within the accuracy margin of a quarter of the Least Significant<br />
Bit (LSB). To achieve this, not only the base frequency fPWM can be suitably chosen, but<br />
also the system’s response to the PWM signal harmonics can be tailored using additional<br />
filters.<br />
As this drawback can thus well be h<strong>and</strong>led, the driver electronics for the DM actuators will<br />
be based on PWM.<br />
6.3 Communication electronics<br />
The communication electronics send the actuator setpoints from the <strong>control</strong> computer to the<br />
driver electronics. The communication link should have low latency to allow a high <strong>control</strong><br />
b<strong>and</strong>width (e.g. low phase lag) <strong>and</strong> a high reliability <strong>and</strong> b<strong>and</strong>width to allow a large number<br />
of actuators to be quickly updated. In addition, the power dissipation, flexibility <strong>and</strong><br />
costs are relevant. If for example, the number of actuators, the b<strong>and</strong>width or the number of<br />
setpoint bits changes, the communication link <strong>and</strong> protocol should allow adaptation. Furthermore<br />
a protocol that can be chosen freely <strong>and</strong> with low overhead is preferred. To limit<br />
development costs, the choice is limited to industrial st<strong>and</strong>ards, such as RS-485 (Profibus),<br />
USB2 (Universal Serial Bus), ethernet, LVDS (Low Voltage Differential Signalling), CAN<br />
(Controller Area Network) <strong>and</strong> FireWire. A few relevant properties of these st<strong>and</strong>ards are<br />
listed in table 6.1. When an update rate of 1kHz <strong>and</strong> 16 bit setpoint values are assumed,<br />
the minimum b<strong>and</strong>width for 5000 actuators is: 1000 · 5000 · 16 = 80Mb/s. With a protocol<br />
overhead of 10% <strong>and</strong> latency limited to one quarter of the sampling time (250µs), a<br />
minimum b<strong>and</strong>width of (5000·1000·16·1.1/0.250) ≈350Mb/s is obtained. For the prototypes<br />
developed with actuator numbers up to 427 actuators, approximately 30Mb/s would<br />
already suffice, but with future, larger, systems the CAN bus <strong>and</strong> RS-485 are no option.<br />
Since the driver electronics will be placed on modules <strong>and</strong> located close to the DM the<br />
power consumption of the transceivers must be as small as possible <strong>and</strong> for practical reasons<br />
the number of wires leading to the modules should be small. Both arguments suggest
6.4 Implementation <strong>and</strong> realization 131<br />
Vpwm<br />
Ll Rl<br />
Low-pass filter coil<br />
La<br />
Cl Ra<br />
Ve<br />
Actuator<br />
Rc<br />
VpwmB<br />
Figure 6.3: The analog electronic<br />
circuit, consisting of a<br />
coarse <strong>and</strong> a fine PWM generator,<br />
an analog low-pass filter<br />
<strong>and</strong> the actuator.<br />
the use of a multi-drop topology in which one transmitter communicates to many receivers<br />
on the same bus. The modules are given a unique identification code to allow messages<br />
to be passed to specific modules. For such topologies the modules do not require a power<br />
dissipating termination resistor <strong>and</strong> the number of communication wires is independent of<br />
the number of receiving modules. However, the power efficiency <strong>and</strong> speed of the communication<br />
link are not only determined by its hardware alone. If the method requires a<br />
specific protocol with a high overhead, both speed <strong>and</strong> power efficiency are reduced <strong>and</strong><br />
flexibility for future upgrades limited. Both arguments favor the development of a custom<br />
protocol. Development costs of such a protocol will be limited, as a master-slave structure<br />
with a small comm<strong>and</strong> set will suffice <strong>and</strong> throughput is more important than guaranteed<br />
transmission.<br />
The Low Voltage Differential Signalling (LVDS) st<strong>and</strong>ard was chosen for the serial communication.<br />
In contrast to USB, this allows for a high b<strong>and</strong>width multi-drop topology for<br />
which low-power transceiver ICs are commercially available. Each transceiver dissipates<br />
only 15mW <strong>and</strong> requires no termination resistor. A custom communication protocol can be<br />
designed that has a small overhead compared to e.g. the USB, FireWire or ethernet protocols.<br />
Two LVDS wire pairs can be used to keep the protocol as simple as possible: one<br />
comm<strong>and</strong> line <strong>and</strong> one return line.<br />
6.4 Implementation <strong>and</strong> realization<br />
In this section the implementation of the chosen design concepts for the driver <strong>and</strong> communication<br />
electronics will be discussed.<br />
6.4.1 PWM implementation<br />
For several reasons the PWM voltage drivers will be implemented in an FPGA. Firstly,<br />
because this leads to a compact design with few components because the FPGA can house<br />
many PWM generators. Moreover, no expensive ASIC has to be designed <strong>and</strong> realized <strong>and</strong><br />
it allows modifications to the implementation through a software update.<br />
As derived in section 6.2.1, the driver electronics require a dynamic range of 16 bits.<br />
The PWM driver electronics will be designed such that the ripple magnitude due to the<br />
harmonic component in the PWM signal at the frequencyfPWM is less than a quarter of the<br />
system’s response to the least significant bit for any duty cycle rPWM. Let H(s) denote the<br />
transfer function between the PWM voltage ˜ VPWM <strong>and</strong> the position za of a single actuator.<br />
6
6<br />
132 6 Electronics<br />
Figure 6.4: H-bridge construction<br />
to allow the PWM<br />
voltage VPWM of the coarse<br />
PWM to be both positive <strong>and</strong><br />
negative.<br />
Vcc<br />
a<br />
b<br />
VPWM<br />
Observe from (6.1) that the worst case magnitude of the first harmonic (n = 1) occurs for<br />
rPWM = 0.5 <strong>and</strong> is equal toVcc/π. The design condition can thus be formulated as:<br />
π |H(2πjfPWM)| < 1Vcc<br />
4216|H(0)|<br />
Vcc<br />
<strong>and</strong> thus:<br />
|H(2πjfPWM)| < π<br />
218|H(0)|. (6.2)<br />
In section 5.2.4, the frequency response functionH(jω) has been derived for the case that<br />
the PWM output is directly connected to the actuator coil. A Bode plot ofH(jω) is plotted<br />
in figure 6.10, which shows that thefPWM for which (6.2) is satisfied lies above 100kHz.<br />
When implemented in an FPGA, the PWM generator will consist of a 16 bit counter <strong>and</strong> a<br />
comparator. The counter value is increased by one at every FPGA clock cycle <strong>and</strong> resets<br />
to zero at the beginning of each PWM time period. The comparator compares the counter<br />
value to a value corresponding to the setpoint. The PWM output is high if the counter is<br />
higher than this value <strong>and</strong> low otherwise. The counting <strong>and</strong> thus clock frequency of the<br />
FPGA can be expressed as:<br />
fFPGA = fPWM2 Nb (6.3)<br />
where Nb is the number of bits of the counter. The clock frequency of currently available<br />
FPGAs is limited to approximately 200MHz, which implies that for Nb = 16 the base frequency<br />
fPWM is limited to approximately 3kHz. The dynamic power dissipation of digital<br />
electronics is for most designs linearly correlated with the clock frequency, which is an important<br />
drive to keep the base frequency as low as possible.<br />
To keep fFPGA below 200MHz while implementing 16bit PWM generators, two modifications<br />
are made. Firstly, an analog2 nd order low-pass filter is added to reduce the system<br />
response magnitude at high frequencies <strong>and</strong> secondly the PWM is split into a fine part consisting<br />
of 5 bits <strong>and</strong> a course part of 11 bits. For 11 bits, the PWM base frequency can be<br />
increased to approximately 95kHz.<br />
The analog low-pass filter consists of the inductor with inductanceLl <strong>and</strong> a capacitor with<br />
capacitance Cl (figure 6.3). It is given a b<strong>and</strong>width of 5kHz that is high enough to have<br />
a negligible influence on the behavior of the system up to the mechanical resonance, but<br />
low enough to reduce the required PWM base frequencyfPWM to less than 95kHz. Assume<br />
that above the resonance at 2kHz the magnitude response of the actuator between driving<br />
voltage <strong>and</strong> position decays a factor 1000 per decade. Further, the response of the analog<br />
low-pass filter decays a factor 100 per decade above 5kHz. The minimum PWM frequency<br />
c<br />
d
6.4 Implementation <strong>and</strong> realization 133<br />
A<br />
B<br />
Vpwm<br />
0<br />
1<br />
0<br />
1<br />
0<br />
Vcc<br />
−Vcc<br />
〈Vpwm〉 = 0<br />
time<br />
〈Vpwm〉 > 0<br />
Figure 6.5: Comparison of the traditional <strong>and</strong> BD modulation schemes. The latter is represented by<br />
the black, solid lines <strong>and</strong> the first by the gray, dashed lines that are shifted slightly to the<br />
top-right for clarity.<br />
fPWM that satisfies (6.2) is then solved from:<br />
3 fPWM<br />
·<br />
2000<br />
A<br />
B<br />
Vpwm<br />
Vcc<br />
0<br />
1<br />
0<br />
1<br />
0<br />
−Vcc<br />
2 fPWM<br />
= 2<br />
5000<br />
18<br />
yielding fPWM ≈ 35kHz. Without the low-pass filter this would be 128kHz, such that this<br />
filter reduces the required PWM base frequency almost a factor 4.<br />
However, this reduced base frequency is only achievable for currently available FPGAs<br />
when the number of bits is less or equal to 12. Therefore, the PWM has been split into<br />
two parts: the 11-bit course PWM provides VPWM whereas the 5-bit fine PWM provides<br />
VPWMB. The latter can re-use the lowest 5 bits of the course PWMs counter <strong>and</strong> is connected<br />
through an appropriately chosen resistor Rc to one of the actuator clamps (figure 6.3).<br />
Evaluating (6.3) for Nb = 11 bits, the required FPGA clock frequency becomes approximately<br />
125MHz. The number of bits of the PWM is split into unequal parts on purpose, as<br />
the resistance of Rc is in practice inaccurate <strong>and</strong> causes an output bias that increases with<br />
the magnitude of the fine PWMs highest bit.<br />
To send positive <strong>and</strong> negative currents through the actuator coil, the PWM must provide<br />
positive <strong>and</strong> negative voltages. This is achieved with an H-bridge construction as shown in<br />
figure 6.4. This construction has only been applied for the coarse PWM. Due to the limited<br />
range of the fine PWM, the added value of a sign change does not make up for the extra<br />
FPGA pins <strong>and</strong> PCB connections. The PWM signals <strong>control</strong> the switchesa,b,c, <strong>and</strong>dsuch<br />
that current flows either via a <strong>and</strong> d or via b <strong>and</strong> c. Care must be taken that both a <strong>and</strong>bas<br />
well as c <strong>and</strong> d are never closed at the same time as this forms a short-circuit. By defining<br />
the PWM output ’low’ as the closing of a <strong>and</strong> d <strong>and</strong> the PWM output ’high’ as the closing<br />
ofb <strong>and</strong>c, the effective voltage over the actuator coil can be varied between−Vcc <strong>and</strong>Vcc.<br />
However, in practice the mean actuator voltage will be approximately zero, which for this<br />
approach corresponds to a duty cycle of 50% (figure 6.5). This means that the voltages over<br />
the coils of the actuator <strong>and</strong> the low-pass filter will continuously vary, resulting in small,<br />
but significant dissipative currents. These can be prevented by the use of a BD modulation<br />
time<br />
6
6<br />
134 6 Electronics<br />
Clock<br />
LVDS<br />
Actuator <strong>control</strong>ler<br />
200 MHz<br />
125 MHz<br />
LVDS<br />
converters<br />
Master<br />
FPGA<br />
125 MHz clock<br />
data-bus<br />
address-bus<br />
<strong>control</strong> lines<br />
Slave<br />
FPGA 1<br />
Slave<br />
FPGA 2<br />
Figure 6.6: Actuator <strong>control</strong>ler architecture.<br />
H-bridges<br />
.<br />
.<br />
Actuators<br />
1-31<br />
scheme.<br />
For the BD modulation scheme, the a −d <strong>and</strong> b − c switches are driven by two different,<br />
but related PWM signals A <strong>and</strong> B, whereas otherwise these would be complementary (i.e.<br />
A=ÒÓØB). For a zero effective voltage, both signals have a duty cycle of 50% <strong>and</strong> are fully<br />
in-phase (figure 6.5). In this situation neither thea−d path nor theb−c path will ever conduct<br />
<strong>and</strong> cause dissipation. If a positive voltage is desired then the period of PWM signal A<br />
is increased whereas that of B is decreased <strong>and</strong> for a negative voltage vice versa. For both<br />
modulation methods, the signals A <strong>and</strong> B drive the switches according to a = A, b =ÒÓØA,<br />
c = B <strong>and</strong> d =ÒÓØB. Therefore, switches b <strong>and</strong> d have always the opposite state of a <strong>and</strong> c<br />
respectively to prevent short-circuits.<br />
A second advantage of the BD modulation method is that the output voltage swing is only<br />
Vcc, whereas for the traditional modulation this is2Vcc (figure 6.5). Consequently, the magnitudes<br />
of the harmonics in the frequency spectrum of the traditionally modulated signal are<br />
twice as high as suggested in (6.1).<br />
6.4.2 FPGA implementation<br />
In section 6.2.1 it is explained that for <strong>modular</strong>ity of the DM system, the driver electronics<br />
should be made in PCB modules containing 61 driver circuits <strong>and</strong> connect to a single<br />
actuator module. To implement the 61 PWM generators <strong>and</strong> the LVDS communication<br />
protocol, three Altera Cyclone II (EP2C8) FPGAs are present on each electronics module.<br />
One master FPGA h<strong>and</strong>les the LVDS protocol <strong>and</strong> two identical slaves implement 32 PWM<br />
generators each. The functionality is not realized in a single FPGA to limit the risk of the<br />
number of available logic cells or electrical connections being insufficient to implement the<br />
required functionality. Due to the two-level PWM solution discussed in section 6.2.2, each<br />
actuator requires FPGA connections for each of the four H-bridge switches <strong>and</strong> one for the<br />
fine PWM signal. This results in 5 connections in total per actuator <strong>and</strong> thus 305 connections<br />
for 61 actuators.<br />
The master FPGA decodes the LVDS signal using 5-times over-sampling (200MHz) <strong>and</strong><br />
interprets the comm<strong>and</strong>s. If required, information is sent to or requested from the slaves<br />
via a 16-bit parallel data bus. The slaves each have one counter that is increased with the<br />
frequency of an externally supplied 125MHz clock. There the 11-bit counter signal is fed<br />
to 32 comparator circuits that generate the PWM signals. These circuits are divided into<br />
.<br />
32-61<br />
.
6.4 Implementation <strong>and</strong> realization 135<br />
Figure 6.7: PCB (top <strong>and</strong> bottom) containing the analog filter electronics for 61 actuators. (left) PCB<br />
(top <strong>and</strong> bottom) containing the master FPGA, DC-DC convertors <strong>and</strong> LVDS drivers.<br />
(right)<br />
Figure 6.8: Top-view of the encased LVDS communications<br />
bridge.<br />
Figure 6.9: Seven PCBs connected to the LVDS<br />
bridge via the multi-drop flat-cable. The bridge<br />
is connected to the laptop via ethernet.<br />
four blocks of eight circuits to prevent a large fan-out of connection wires that bring the<br />
counter to the comparators. Such fan-out limits the switching speed <strong>and</strong> leads to undesired<br />
dissipation.<br />
Nevertheless, as will be discussed at the end of this chapter, the dissipation of the three<br />
FPGAs is dominant over the RMS dissipation in the actuator coils. In section 6.8.1 several<br />
design concepts will be proposed to reduce this.<br />
Figure 6.7 shows the double-sided PCB with 61 drivers <strong>and</strong> the PCB with the master FPGA,<br />
the DC/DC convertors <strong>and</strong> the LVDS drivers. The connector board that connects to the<br />
three flex foil flaps on one side <strong>and</strong> the analog electronics PCB on the other side is shown<br />
in the lower right photo in figure 7.2.<br />
6
6<br />
136 6 Electronics<br />
6.4.3 The ethernet to LVDS bridge<br />
At the time of design, no general purpose PC expansion card was available to provide an<br />
off-the-shelf PC with two LVDS connections <strong>and</strong> a fully customized communication protocol.<br />
Therefore, a communications bridge was conceived that bridges the 100Mb/s ethernet<br />
connection of a PC with the custom LVDS connection. The LVDS bridge must relay messages<br />
received over the ethernet connection to the LVDS on the other side <strong>and</strong> vice versa.<br />
The bridge should be reliable <strong>and</strong> add little latency – i.e. the delay between reception <strong>and</strong><br />
transmission of the first bit of a data package – to the delay of the two-step communication<br />
chain. To limit development time, the LVDS bridge is based on an Altera NIOS-II FPGA<br />
development board (figure 6.8). This board is extended with an ethernet PHY that implements<br />
the MAC layer of the ethernet protocol in hardware to limit latency. A second plug-in<br />
PCB contains the LVDS driver ICs. The NIOS FPGA implements a processor that executes<br />
an open source Internet Protocol (IP) stack that has been optimized for latency. As with any<br />
communication type, transmission errors may occur for which detection methods are usually<br />
implemented. However, for real-time application it is more important to limit latency than<br />
to detect or recover rarely occurring errors. Therefore, the User Datagram Protocol (UDP)<br />
protocol has been chosen (appendix E) for the ethernet communication, whose checksums<br />
to detect faulty data have been disabled or are ignored.<br />
6.5 Modeling<br />
The actuator <strong>and</strong> its electronic circuit are modeled to determine a suitable base frequency<br />
for the PWM signals <strong>and</strong> to check whether both the actuator <strong>and</strong> the electronics behave as<br />
designed. Furthermore it allows validation of the full DM system including its reflective<br />
facesheet (chapter 7) <strong>and</strong> serves as input for a <strong>control</strong>ler synthesis procedure.<br />
Recall the analog electronic circuit depicted in figure 6.3. Let the circuit be driven by the<br />
PWM voltage VPWM. The effect of the fine PWM signal that connects to the system at a<br />
different location – leading to different <strong>dynamics</strong> – will further be neglected. The actuator<br />
has been modeled in section 5.2.4, leading to the linearized system in (5.24) on page 106.<br />
The2 nd order analog low-pass filter consists of coilLl with internal resistance Rl in series<br />
with capacitor Cl that is connected in parallel with the actuator. From Kirchoff’s laws it<br />
follows that:<br />
VPWM = VLl +VRl +Va, <strong>and</strong> IRl = ICl +Ia,<br />
whereVRl <strong>and</strong>VLl denote the potentials overRl <strong>and</strong>Ll <strong>and</strong>ICl <strong>and</strong>IRl the currents through<br />
Cl <strong>and</strong>Rl respectively. They are defined through the following constitutional equations:<br />
VRl<br />
= RlIRl , VLl = Ll ˙<br />
IRl , ICl = Cl ˙ Va.<br />
The system will be modeled in the state-update form,<br />
˙x = Ax+BVPWM, (6.4)<br />
with state vectorx(t) = [Ia(t) za(t) ˙<br />
za(t) Va(t) IRl (t)]T . The time derivatives of the state<br />
elements can be derived from the constitutional equations together with the two Kirchhoff
6.5 Modeling 137<br />
Table 6.2: Properties of the components of the 2 nd order analog low-pass filter.<br />
equations, leading to:<br />
Parameter Value Unit Parameter Value Unit<br />
Ll 220 µH Cl 4.7 µF<br />
Rl 2.7 Ω Rc 16.2 kΩ<br />
R ′ l 2.4 Ω R ′ a 3 Ω<br />
˙<br />
IRl = (VPWM −RlIRl −Va)/Ll,<br />
˙Va = (IRl −Ia)/Cl.<br />
These two equations can be combined with the previously derived actuator system equation<br />
in (5.24) <strong>and</strong> expressed in the state update form of (6.4) as:<br />
⎡<br />
˙<br />
⎤<br />
Ia<br />
⎢za<br />
⎢<br />
˙ ⎥<br />
⎢ ¨za ⎥<br />
⎢ ⎥<br />
⎣ ˙Va<br />
⎦<br />
˙ IRl<br />
=<br />
⎡<br />
−Ra/La 0 −Ka(z<br />
⎢<br />
⎣<br />
′ a )/La 1/La 0<br />
0 0 1 0 0<br />
Ka(z ′ a )/mac<br />
⎤⎡<br />
⎤<br />
Ia<br />
⎥⎢<br />
za ⎥<br />
⎥⎢<br />
⎥<br />
−ca/mac −ba/mac 0 0 ⎥⎢<br />
za ⎥⎢<br />
˙ ⎥<br />
−1/Cl 0 0 0 1/Cl ⎦⎣Va<br />
⎦<br />
0 0 0 −1/Ll −Rl/Ll IRl<br />
+<br />
⎡ ⎤<br />
0<br />
⎢ 0 ⎥<br />
⎢ 0 ⎥<br />
⎣ 0<br />
1/Ll<br />
(6.5)<br />
The output signals that will be used for analysis <strong>and</strong> testing are the actuator displacement<br />
za(t) = [0 1 0 0 0]x(t) <strong>and</strong> the voltage Va(t) = [0 0 0 1 0]x(t) that can be measured over<br />
actuator coil. Let the transfer function between the PWM voltageVPWM(t) <strong>and</strong> the actuator<br />
positionza(t) be denotedH(s). For properties of the actuator <strong>and</strong> the electronics as in table<br />
6.2, figure 6.10 shows the Bode plots of the resulting transfer functions. Figure 6.10 also<br />
shows the Bode plot of the transfer function when only the current-<strong>control</strong>led mechanical<br />
system is considered. The static relation Ia = Va/Ra is used to scale the corresponding<br />
transfer function in (5.28) on page 111 <strong>and</strong> allow comparison with the full mechatronic<br />
system. When omitting the nominal operating pointz ′ a, this yields the transfer function:<br />
Magnitude [m/V]<br />
Phase [deg]<br />
10 −4<br />
10 −6<br />
10 −8<br />
10 −10<br />
0<br />
−90<br />
−180<br />
−270<br />
−360<br />
−450<br />
H mech (s)<br />
H v (s)<br />
H(s)<br />
18 bit suppression<br />
10 3<br />
Hm(s) = HI(s)/Ra =<br />
macRas2 . (6.6)<br />
+baRas+caRa<br />
10 4<br />
Frequency [Hz]<br />
10 5<br />
Ka<br />
Figure 6.10: Bode diagram of three transfer<br />
functions: Hm(s) from (6.6) (only the mechanics),<br />
Hmv(s) from (5.26) (mechanics including<br />
the actuator coil) <strong>and</strong> H(s) defined in section<br />
6.5 (mechanics with actuator coil <strong>and</strong> low-pass<br />
filter).<br />
⎦ VPWM.<br />
6
6<br />
138 6 Electronics<br />
Observe in the Bode plot that the electronics provide a small amount of additional damping<br />
of the mechanical resonance, but have negligible influence on the low-frequent actuator<br />
behavior.<br />
The PWM base frequency<br />
In section 6.2.2 it was discussed that an 18-bit attenuation of the PWM ripple is desirable.<br />
To achieve this with an FPGA based implementation, an additional2 nd order low-pass filter<br />
was added to the design.<br />
The effect of this filter is illustrated in figure 6.10, which contains Bode plots of the actuator<br />
system with <strong>and</strong> without the filter as described by (6.5) <strong>and</strong> (5.24) respectively. The 18bit<br />
ripple attenuation is shown in the magnitude plot of figure 6.10 as the dash-dotted line.<br />
Observe that the application of the low-pass filter reduces the PWM frequency requirement<br />
from approximately 128kHz to approximately 35kHz.<br />
Although higher than necessary for DMs with Pyrex facesheets, the base frequency fPWM<br />
is set at 61kHz. This is done because in future developments the replacement of the Pyrex<br />
<strong>mirror</strong> facesheet by beryllium is foreseen, dem<strong>and</strong>ing a higher base frequency. The specific<br />
stiffness of beryllium is more than five times higher , which allows for thinner facesheets<br />
<strong>and</strong> thus less mass per actuator. With the same actuator stiffness this increases the system’s<br />
eigenfrequency <strong>and</strong> decreases the attenuation of the PWM ripple.<br />
For the foreseen update rate of 1kHz the base frequency of 61kHz provides 61 times oversampling<br />
of the setpoint signal. This means that cross-harmonics in the PWM output voltage<br />
VPWM resulting from non-constant setpoint signals can be neglected.<br />
Serial communication<br />
The serial communication via both ethernet (UDP) <strong>and</strong> LVDS will introduce a certain delay<br />
τc of the <strong>control</strong> output. This delay should be as small as possible <strong>and</strong> its variation (jitter)<br />
should be restricted to a negligibly small fraction of the delay itself. The communication<br />
latency is in the Laplace domain modeled as Hτc(s) = e −τcs . Due to the definition of the<br />
communication protocol (appendix E) <strong>and</strong> its serial nature, the latency τc will be different<br />
for each 61-actuator module.<br />
6.6 Evaluation of <strong>control</strong> aspects<br />
In this section the actuator <strong>and</strong> electronics design will be evaluated from the perspective of<br />
closed-loop <strong>control</strong> of a single actuator. In section 5.3.1 it is shown that the damping of<br />
the first mechanical resonance frequency of the actuators is low. Further, the resonance frequency<br />
is higher than the sampling frequency of the foreseen <strong>control</strong> system. The latter is<br />
mainly limited by the Charge Coupled Device (CCD)-based Shack-Hartmann sensor used<br />
for feedback. As derived in chapter 2, the required b<strong>and</strong>width for position <strong>control</strong> of the<br />
DM is 200Hz. If this is to be achieved for the full DM system in which there is interaction<br />
between actuators, then it should also be achievable with the single actuator considered<br />
here. However, the lowly damped <strong>and</strong> potentially aliased resonance may limit the achievable<br />
b<strong>and</strong>width. For this reason, damping should be increased in future designs.<br />
The <strong>control</strong> <strong>and</strong> sampling aspects of the actuator system will be evaluated using the Aliased
6.6 Evaluation of <strong>control</strong> aspects 139<br />
r + e u<br />
z y<br />
C Hτc Hzoh H G<br />
−<br />
ideal<br />
sampler<br />
Figure 6.11: Schematic of the <strong>control</strong> loop for position <strong>control</strong> of a single actuator. The grey area<br />
contains the continuous time parts of the loop.<br />
Frequency Response Function (AFRF) as proposed in [94, 163]. Recall H(s) to denote the<br />
transfer function between PWM voltageVPWM(t) <strong>and</strong> actuator positionza(t). The <strong>control</strong>ler<br />
C(za) will be implemented in discrete time with sampling timeTs <strong>and</strong> send its voltage setpoint<br />
via the communication link with latencyτc to the driver electronics. The PWM based<br />
driver electronics form a Zero Order Hold (ZOH) mechanism that can be expressed in the<br />
Laplace domain as Hzoh(s) = (1 − e −sTs )/s [62]. Finally, the sensor input of the <strong>control</strong>ler<br />
is considered to be CCD-based, similar to the Shack-Hartmann sensor. It integrates<br />
photons over the exposure time τe. When ignoring measurement noise, a continuous time<br />
measurement ˜y(t) of the signalza(t) at timetcan be modeled as [100]:<br />
˜y(t) = 1<br />
τe<br />
t<br />
t−τe<br />
which can be transformed to the Laplace domain as:<br />
za(τ)dτ,<br />
˜y(s) = G(s)za(s), where G(s) = 1−e−sτe<br />
.<br />
sτe<br />
It is assumed that the continuous time signal ˜y(t) is then sampled by an ideal sampler,<br />
yielding the discrete time signal y(k) for k = nTs, n = 0,1,2,.... Schematically, this<br />
yields the <strong>control</strong> loop depicted in figure 6.11.<br />
Since the loop contains both discrete time <strong>and</strong> continuous time components, the loop gain<br />
can be written as:<br />
L(s,z) = Ls(s)C(z),<br />
where<br />
Ls(s) = G(s)H(s)Hzoh(s)Hτc(s) = 1−e−sτe<br />
sτe<br />
H(s) 1−e−sTs<br />
e<br />
s<br />
−sτc .<br />
Note thatz denotes the discrete time Laplace variable. Since this transfer function depends<br />
on boths<strong>and</strong>z, use will be made of the AFRF, which is a discrete-time system that models<br />
both the continuous time components <strong>and</strong> the aliasing effects introduced by the ideal sampler.<br />
This means that the continuous time systems are discretized to the sampling frequency<br />
fs = 1/Ts of the <strong>control</strong>ler, causing the frequencies in the signal ˜y(t) above the Nyquist<br />
frequency fN = fs/2 to be mapped to frequencies 0 < f < fN . The result is a linear,<br />
discrete time system from which performance <strong>and</strong> stability properties can be derived using<br />
common Linear Time-Invariant (LTI) tools [163].<br />
The AFRF L ∗ (z) of the loop gain L(s,z) is the open-loop transfer function between the<br />
6
6<br />
140 6 Electronics<br />
Magnitude [m/V]<br />
Phase [deg]<br />
10 −3<br />
10 −4<br />
10 −5<br />
10 −6<br />
10 −7<br />
0<br />
−180<br />
−360<br />
10 1<br />
10 2<br />
10 3<br />
Frequency [Hz]<br />
f s = 500 [Hz]<br />
f s = 1000 [Hz]<br />
f s = 1500 [Hz]<br />
f s = 2000 [Hz]<br />
f s = 10000 [Hz]<br />
H(s)<br />
Figure 6.12: The AFRF of the single actuator<br />
system for various sampling frequencies when<br />
assuming τc = 120µs.<br />
10 4<br />
10 5<br />
Figure 6.13: Nyquist plots of the loop gain<br />
L ∗ (z) based on the AFRF of the single actuator<br />
system <strong>and</strong> the <strong>control</strong>lerC(z) from (6.8) for<br />
various sampling frequencies <strong>and</strong> τc = 120µs.<br />
Table 6.3: Achievable b<strong>and</strong>widths for an integrator type <strong>control</strong>ler for various sampling frequencies.<br />
Sampling frequency [Hz] 500 1000 1500 2000 10000<br />
B<strong>and</strong>width [Hz] 98 198 159 213 196<br />
Optimalα[V/m] (×1000) 170 170 86 84 15<br />
Optimalγ [V/m] (×1000) -24 -19 -15 -7 -16<br />
sampled output y(k) <strong>and</strong> the discrete time reference signal r(k). It can be obtained using<br />
thez-transform denoted by the operatorZ as:<br />
L ∗ −ste 1−e<br />
(z) = Z(Ls(s,z))C(z) = Z H(s)<br />
ste<br />
1−e−sTs<br />
e<br />
s<br />
−sτc<br />
<br />
C(z).<br />
When assuming the exposure time τe to be equal to one sampling time – i.e. τe = Ts – <strong>and</strong><br />
using thatZ(1−e −sTs ) = 1−z −1 , this reduces to:<br />
L ∗ (z) = (1−z −1 )(1−z −1 <br />
H(s)<br />
)Z<br />
s2 e<br />
Ts<br />
−sτc<br />
descriptions are listed in [62], but this conversion is also implemented in Matlab’s<br />
<br />
C(z). (6.7)<br />
The involvedz-transform can be derived by first transforming the continuous time transfer<br />
function into pole/residue form (partial fraction expansion). This describes the system as a<br />
sum of first order systems, which have an equivalentz-domain description. These equivalent<br />
comm<strong>and</strong> [130], which can also account for arbitrary time-delays τc. In figure 6.12 the<br />
AFRFs are plotted for various sampling frequencies when assumingτc = 120µs. Here it is<br />
clear that the effect of aliased <strong>dynamics</strong> is very small.<br />
To answer the question whether the b<strong>and</strong>width of 200Hz is achievable, let C(z) be a<br />
discrete time proportional plus integrator <strong>control</strong>ler with integrator gainα<strong>and</strong> proportional
6.7 Testing <strong>and</strong> validation 141<br />
gainγ:<br />
C(z) =<br />
α<br />
+γ (6.8)<br />
1−z −1<br />
As discussed in chapter 3, an integrator structure is often used for the <strong>control</strong>ler of an AO<br />
system. The proportional gain term is added to increase the phase margin around the b<strong>and</strong>width.<br />
Note that computational delays are here neglected.<br />
Figure 6.13 shows Nyquist plots of the loop gain L ∗ (z) for various sampling frequencies<br />
after tuning the parametersα <strong>and</strong> γ for highest b<strong>and</strong>width (0dB crossing of the magnitude<br />
of the open-loop gainL ∗ (z)) while considering a modulus margin [62] of 0.5. The achieved<br />
b<strong>and</strong>widths are listed in table 6.3. Observe that the b<strong>and</strong>width varies with the sampling<br />
frequency fs <strong>and</strong> for fs =2000Hz it is with 213Hz the highest. The reason that a higher<br />
sampling frequency does not lead to a higher b<strong>and</strong>width is that the phase delay due to the<br />
low sampling frequency rotates the phase response at the resonance frequency away past<br />
the (-1,0) point. Consequently, the integrator gain α can be further increased without compromising<br />
the modulus margin.<br />
As discussed in chapter 2 the achievable b<strong>and</strong>width depends strongly on the chosen sampling<br />
frequency fs, but it has here been shown to be amplified by the low damping of the<br />
mechanical resonance. This should be taken into account for future design revisions. Nevertheless,<br />
for fs = 2000Hz the desired <strong>control</strong> b<strong>and</strong>width of 200Hz can be achieved while<br />
respecting a modulus margin of 0.5.<br />
6.7 Testing <strong>and</strong> validation<br />
The electronics <strong>and</strong> the actuator grids were tested. The dynamic response of the actuators<br />
was first measured using a <strong>control</strong>led current source. These results were shown in section<br />
5.3.1. In this section, first the test results of the communication between a PC <strong>and</strong> the LVDS<br />
bridge <strong>and</strong> between the PC <strong>and</strong> the driver modules are shown (figure 6.9) followed by a full<br />
system test. Here the dynamic response of the actuators is measured. At the end of the<br />
section the power dissipation of the electronics will be discussed.<br />
6.7.1 Communications tests<br />
To measure the latency of the communications bridge, two of its debug lines were connected<br />
to a logic analyzer. The first line is high while a UDP packet is being received <strong>and</strong> the second<br />
while an LVDS packet is being transmitted. A second computer was used to send burst<br />
packets (appendix E). These are the most relevant in practice <strong>and</strong> contain 16-bit setpoint updates<br />
for all 61 actuators corresponding to 1024 bits in total. To minimize ethernet protocol<br />
overhead, each UDP burst packet can contain up to eight LVDS burst packets (appendixs D<br />
<strong>and</strong> E).<br />
Measurements taken by the logic analyzer show that the transmission time of a UDP burst<br />
packet can be expressed as:<br />
τudp ≈ 4.7·10 −6 +10.24·10 −6 Nm<br />
where Nm is the number of actuator modules within the packet. The constant part is due<br />
to ethernet protocol overhead <strong>and</strong> the approximately 10µs per module corresponds to 1024<br />
6
6<br />
142 6 Electronics<br />
bits at a rate of 100Mb/s.<br />
Further, the measurements show a time delay of approximately 85µs with a variation (jitter)<br />
of less than 10µs after reception of the UDP packet, before transmission of an LVDS<br />
packet. During this time the bridge processes the packet, splits it into LVDS packets <strong>and</strong><br />
copies it to the transmit buffer. Since calculation of the UDP checksum takes a significant<br />
time – approximately(10Nm)µs – this checksum is sacrificed for speed <strong>and</strong> ignored in the<br />
current implementation. Transmission of the 1024 functional bits over the 40Mb/s LVDS<br />
connection with 16-bit data words separated by 18 pause bits, one start-bit <strong>and</strong> one stop-bit<br />
should takeτlvds ≈ 28.8·10 −6 Nm, which is confirmed by the measurements.<br />
The total communication latencyτc can thus be expressed as:<br />
τc = τudp +85·10 −6 +τlvds = 89.7·10 −6 +39·10 −6 Nm<br />
Since the communication chain consists of two sequential, buffered links, the maximum<br />
update rate is determined by the slowest link, in this case the LVDS line. This rate equals<br />
1/28.8 · 10 −6 /Nm, which for Nm = 1 <strong>and</strong> Nm = 7 is approximately 35 <strong>and</strong> 5kHz respectively.<br />
However, since the LVDS bridge may drop incoming packets during its 85µs<br />
processing time, in practice this latency adds directly to the ethernet latency. For the case<br />
that Nm = 1, this makes the ethernet latency dominant <strong>and</strong> reduces the maximum update<br />
rate to approximately10kHz.<br />
6.7.2 Parasitic resistance measurements<br />
Before the full actuator systems will be tested, first several properties of the electronics<br />
are measured. Deviations of the expected values measured in the next section can then be<br />
properly attributed to either the electronics or the mechanics.<br />
In practice – due to wiring – the resistances Rl <strong>and</strong> Ra are assumed to increase by R ′ l <strong>and</strong><br />
R ′ a respectively. Resistance measurements of the actuator coils show on average the design<br />
value of 39.0Ω, but the average resistance measured over the capacitor Cl is found to be<br />
42Ω, indicating that R ′ a ≈ 3Ω. The resistance R ′ l will be estimated from a few additional<br />
measurements. Firstly, the PWM setpoints are set such that the measured voltage over<br />
the capacitor Cl is VCl =1.001V. After reconnecting the actuators this yielded an average<br />
voltage drop over the capacitor ofV=0.903V. Using the fact that the current throughRl,R ′ l ,<br />
Ra <strong>and</strong>R ′ a is equal <strong>and</strong> the sum of the voltage drops is equal to the PWM voltage of 1.001V,<br />
the resistance R ′ l is estimated as R′ l = 2.4Ω. A 1Ω part of the latter can be attributed to<br />
a safety resistor present in the design for short-circuit protection, whereas the rest must be<br />
attributed to wiring <strong>and</strong> connector resistance.<br />
Although the supply voltage variations will not have an effect for power dissipation – this<br />
will be compensated by a <strong>control</strong>ler – the power dissipation increases linearly with the<br />
= 3 + 2.4 =<br />
resistance of the current path. The total parasitic resistance of R ′ a + R ′ l<br />
5.4Ω will therefore lead to an undesired increase in power dissipation on the driver PCB of<br />
(R ′ a +R′ l )/(Ra +Rl)·100% = 5.4/(39.0+2.7)·100% ≈ 13%.<br />
6.7.3 Actuator system validation<br />
Whereas dynamic measurements were performed on the single actuator prototypes using a<br />
Siglab system, the setup depicted in figure 6.14 will be used for testing <strong>and</strong> model validation
6.7 Testing <strong>and</strong> validation 143<br />
xPC<br />
excitation Ethernet LVDS<br />
UDP switch bridge<br />
UDP LVDS<br />
Electronics<br />
module<br />
actuator<br />
Laservibrometer<br />
velocity<br />
position<br />
Figure 6.14: Setup used to perform the actuator response measurements using the custom built electronics.<br />
of the grid actuators. A Matlab TM xPC-target computer is used to generate a white noise<br />
sequence <strong>and</strong> send it in UDP burst packets (appendix E) over an ethernet connection. The<br />
sequence is logged internally to be used for analysis later. The ethernet connection goes<br />
via a switch to allow both the xPC target <strong>and</strong> the dedicated electronics to be <strong>control</strong>led <strong>and</strong><br />
configured by a host computer. Although this doubles the ethernet latency, this is not critical<br />
for the open-loop validation measurements.<br />
The LVDS bridge then converts the packets into LVDS packets to be sent to the electronics<br />
module corresponding to the targeted actuator. Both the position <strong>and</strong> velocity response of<br />
the actuator are measured using a polytec laser vibrometer. This outputs the measurements<br />
as analog voltages that are fed back to the xPC target using a National Instruments Analog<br />
to Digital Convertor (ADC) card (NI-6025E) that does not contain any anti-aliasing filters.<br />
The measurements are performed for update rates of 1, 3, 5 <strong>and</strong> 10kHz to be able to evaluate<br />
the effects of sampling <strong>and</strong> aliasing.<br />
Let the discrete time frequency response function between the PWM voltage output<br />
<strong>and</strong> the actuator positionza that includes the effects of sampling <strong>and</strong> digital communication<br />
Magnitude [m/V]<br />
Phase [deg]<br />
10 −4<br />
10 −6<br />
0<br />
−180<br />
−360<br />
−540<br />
−720<br />
10 1<br />
Measurements<br />
Fitted models<br />
Original model<br />
F s = 1kHz<br />
10 2<br />
F s = 3kHz<br />
Frequency [Hz]<br />
F s = 10kHz<br />
Figure 6.15: Bode plot of empirical frequency<br />
response function estimates ˆH ∗ p,Ts (f) together<br />
with the parametric fit H ∗ p,Ts (za, ˆ θ) <strong>and</strong> the<br />
nominal model H ∗ p,Ts (za,θ0) at 1, 3 <strong>and</strong> 10kHz<br />
sampling frequencies.<br />
10 3<br />
Magnitude [m/Vs]<br />
Phase [deg]<br />
10 0<br />
10 −1<br />
10 −2<br />
10 −3<br />
10 −4<br />
180<br />
0<br />
−180<br />
−360<br />
−540<br />
−720<br />
10 1<br />
F s = 1kHz F s = 3kHz<br />
10 2<br />
Frequency [Hz]<br />
F s = 10kHz<br />
Figure 6.16: Bode plot of empirical frequency<br />
response function estimates ˆHv,Ts(f) together<br />
with the parametric fit H ∗ v,Ts (za, ˆ θ) <strong>and</strong> the<br />
nominal model H ∗ v,Ts (za,θ0) at 1, 3 <strong>and</strong> 10kHz<br />
sampling frequencies.<br />
10 3<br />
6
6<br />
144 6 Electronics<br />
Actuator stiffness c a [N/m]<br />
650<br />
600<br />
550<br />
500<br />
450<br />
400<br />
350<br />
300<br />
Module 1<br />
Module 2<br />
Module 3<br />
Module 4<br />
Module 5<br />
Module 6<br />
Module 7<br />
10 20 30 40 50 60<br />
Actuator number [−]<br />
Figure 6.17: Identified actuator stiffness ca<br />
when assuming mac =3.6mg sorted on value<br />
for each measured actuator grid separately. The<br />
value predicted by the original model is 583N/m.<br />
Resonance frequency f e [Hz]<br />
2100<br />
2000<br />
1900<br />
1800<br />
1700<br />
1600<br />
1500<br />
Module 1<br />
Module 2<br />
Module 3<br />
Module 4<br />
Module 5<br />
Module 6<br />
Module 7<br />
10 20 30 40 50 60<br />
Actuator number [−]<br />
Figure 6.18: Resonance frequenciesfe when assuming<br />
mac =3.6mg sorted on value for each<br />
measured actuator grid separately.<br />
be denoted H ∗ p,Ts (za,θ,τc), where the subscript Ts indicates the corresponding sampling<br />
time <strong>and</strong> the vectorθ contains the physical parametersmac,ba,ca,La,Ra,Ka,Ll,Rl <strong>and</strong><br />
Cl. Similarly,H ∗ v,Ts (za,θ,τc) denotes the transfer function to the actuator velocityza. ˙ The<br />
effect of the sampling performed by the NI ADC card can be modeled by assuming a zero<br />
order hold on the excitation signal <strong>and</strong> applying the z-transform similar to the procedure<br />
described in section 6.6, yielding:<br />
H ∗ p,Ts (z,θ,τc)<br />
Tss 1−e<br />
= ZTs<br />
s H(s,θ)e−τcs<br />
<br />
= (1−z −1 −τcs<br />
)ZTs H(s)e /s ,<br />
Motor constant K a [N/A]<br />
0.2<br />
0.18<br />
0.16<br />
0.14<br />
0.12<br />
0.1<br />
0.08<br />
0.06<br />
Module 1<br />
Module 2<br />
Module 3<br />
Module 4<br />
Module 5<br />
Module 6<br />
Module 7<br />
10 20 30 40 50 60<br />
Actuator number [−]<br />
Figure 6.19: The motor constants Ka when assuming<br />
mac =3.6mg sorted on value for each<br />
measured actuator grid separately. The value<br />
predicted by the original model is 0.19N/A.<br />
Viscous damping b a [mNs/m]<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
Module 1<br />
Module 2<br />
Module 3<br />
Module 4<br />
Module 5<br />
Module 6<br />
Module 7<br />
10 20 30 40 50 60<br />
Actuator number [−]<br />
Figure 6.20: The viscous damping ba when assuming<br />
mac =3.6mg sorted on value for each<br />
measured actuator grid separately.
6.7 Testing <strong>and</strong> validation 145<br />
Inductance L a [mVs/A]<br />
3.6<br />
3.4<br />
3.2<br />
3<br />
2.8<br />
2.6<br />
2.4<br />
2.2<br />
2<br />
Module 1<br />
Module 2<br />
Module 3<br />
Module 4<br />
Module 5<br />
Module 6<br />
Module 7<br />
10 20 30 40 50 60<br />
Actuator number [−]<br />
Figure 6.21: Identified actuator coil inductance<br />
La when assumingmac =3.6mg sorted on value<br />
for each measured actuator grid separately. The<br />
value used for the original model is 2.93mH.<br />
Latency τ c [µs]<br />
170<br />
165<br />
160<br />
155<br />
Module 1<br />
Module 2<br />
Module 3<br />
Module 4<br />
Module 5<br />
Module 6<br />
Module 7<br />
10 20 30 40 50 60<br />
Actuator number [−]<br />
Figure 6.22: The communication latency τc<br />
sorted on value for each measured actuator grid<br />
separately.<br />
Table 6.4: Average values <strong>and</strong> st<strong>and</strong>ard deviations of the actuator system properties measured over<br />
all grid actuators using the setup depicted in figure 6.14.<br />
Property Ka ca ba fe La<br />
Average 0.11N/A 473N/m 0.30mNs/m 1.83kHz 3.0mH<br />
Std.dev. 0.02N/A 46N/m 0.11mNs/m 91Hz 0.2mH<br />
H ∗ v,Ts (z,θ,τc) = ZTs<br />
1−e Tss<br />
s<br />
sH(s,θ)e −τcs<br />
<br />
= (1−z −1 −τcs<br />
)ZTs H(s)e ,<br />
where H(s,θ) denotes the transfer function H(s) based on the physical parameters<br />
in the vector θ. Let the empirical transfer function estimates of H ∗ p,Ts (za,θ,τc) <strong>and</strong><br />
H∗ v,Ts (za,θ,τc) be denoted ˆ H∗ p,Ts (f) <strong>and</strong> ˆ H∗ (f) respectively, where f is the fre-<br />
v,Ts<br />
quency. These estimates together with the corresponding coherence functions Cp,Ts(f)<br />
<strong>and</strong> Cv,Ts(f) were obtained from 10s of input-output data logged by the xPC target using<br />
Welch’s averaged periodogram method with a block size of 2048 samples with 70% overlap<br />
<strong>and</strong> a Hanning window. A typical set of estimates is shown in figures 6.15 <strong>and</strong> 6.16 for<br />
actuator 10 of grid 1.<br />
The model parametersca/mac,Ka/mac,ba/mac,La <strong>and</strong> the latencyτc will be identified<br />
from the empirical transfer function estimates. The other model parameters Rl, Ll <strong>and</strong> Cl<br />
are assumed to be accurately known. A single set of parameters is fit against eight measurements<br />
series: four different sampling frequencies times two measurement outputs (position<br />
<strong>and</strong> velocity). The optimization is performed w.r.t. the cost functionJp +Jv, in which:<br />
Jp = <br />
<br />
Ts∈Ts f∈Fm(Ts)<br />
⎛<br />
⎝ Ĉp,Ts(f)<br />
<br />
<br />
ˆ <br />
<br />
<br />
<br />
H<br />
Hp,Ts(f) <br />
∗ p,Ts (e2πjf )− ˆ ⎞2<br />
<br />
<br />
Hp,Ts(f) ⎠<br />
,<br />
6
6<br />
146 6 Electronics<br />
Jv = <br />
<br />
Ts∈Ts f∈Fm(Ts)<br />
⎛<br />
⎝ Ĉv,Ts(f)<br />
<br />
<br />
ˆ <br />
<br />
<br />
<br />
H<br />
Hv,Ts(f) <br />
∗ v,Ts (e2πjf )− ˆ ⎞2<br />
<br />
<br />
Hv,Ts(f) ⎠<br />
.<br />
The set Ts consists of the sampling times corresponding to 1, 3, 5 <strong>and</strong> 10kHz update rates<br />
<strong>and</strong> the sets Fm(Ts) contain the frequencies at which the transfer functions were estimated<br />
for the sampling timeTs. The values of the coherence functionsCp,Ts(f) <strong>and</strong>Cv,Ts(f) are<br />
used as weights to include the reliability of the transfer function estimates in the parametric<br />
optimization problem. To remove the bias introduced by the system magnitude response,<br />
the inverse of this response is applied as a second weight. In the optimization the properties<br />
of the electric components that form the analog low-pass filter are taken from table 6.2 <strong>and</strong><br />
include the parasitic resistancesR ′ l <strong>and</strong>R′ a<br />
. The z-transform is implemented using Matlab’s<br />
function, which also accounts for the latencyτc. Since this cost function is non-linear<br />
w.r.t. the parameters to be estimated, the optimization is performed using Matlab’s nonlinear<br />
least squares solverÐ×ÕÒÓÒÐÒ.<br />
Examples of the estimation result for an arbitrary actuator are depicted as the gray dashed<br />
lines in figures 6.15 <strong>and</strong> 6.16, showing only very slight deviations between the model <strong>and</strong><br />
the measurements. Moreover, it should be noted that both figures are based on the same<br />
set of parameter estimates. The average values <strong>and</strong> st<strong>and</strong>ard deviations of the estimates for<br />
Ka, ca, ba <strong>and</strong> La when assuming mac =3.6mg are listed in table 6.4. These values agree<br />
well with the results obtained using the current source in the previous chapter as listed in<br />
table 5.2 on page 121, which indicates the robustness of the measurement <strong>and</strong> identification<br />
process. Further, all estimates are plotted in figures 6.19, 6.17, 6.20 <strong>and</strong> 6.21. For these<br />
plots the values are sorted per actuator module for better insight into their statistical spread<br />
<strong>and</strong> correspond well to the values estimated using the current source setup in section 5.3.1<br />
depicted in figures 5.27, 5.25 <strong>and</strong> 5.28 on page 122. Each module has a few actuators with<br />
significantly different properties, but only a single actuator is malfunctioning. Moreover,<br />
the variation between actuators is not linked to the location of the actuator in the grid. This<br />
is illustrated for the resonance frequency <strong>and</strong> motor constant in the figures in appendix F.<br />
These show the corresponding values for all actuator grids in relation to the location of the<br />
actuator in the module.<br />
Although the resistance in the current path of an actuator affect this system’s DC-gain, it<br />
cannot be separately estimated. The parasitic resistances R ′ a <strong>and</strong> R′ l<br />
are only practically<br />
measurable for a few actuators per module. Measurements for approximately 20 actuators<br />
provided the average value used for the estimation of the actuator parameters. The several<br />
percent resistance variation affects the parameter estimates, which together with estimation<br />
errors explains the differences with the values obtained from current source measurements<br />
shown in section 5.3.1.<br />
In addition to the results shown in section 5.3.1, the voltage excitation allows to estimate the<br />
actuator inductance La. Observe from figure 6.21 that also the average of the inductance<br />
differs significantly per actuator module, which is most significant for module 5. This module<br />
also has a relatively low average motor constant (figure 6.19). Based on the sensitivity<br />
analysis performed on the actuator design in section 5.2.6, such variation can for instance<br />
be attributed to an increased radial air gap reluctanceℜgr. This reluctance depends strongly<br />
on the radial air gap width, which is e.g. equally affected for all actuators of a module by<br />
the radius of the mill used in the baseplate milling process.
6.7 Testing <strong>and</strong> validation 147<br />
Displacement [µm]<br />
30<br />
20<br />
10<br />
0<br />
−10<br />
−20<br />
−30<br />
−40<br />
Response measurement<br />
Linear approximation<br />
Linearity error<br />
−3 −2 −1 0 1 2 3<br />
Voltage setpoint [V]<br />
3<br />
2<br />
1<br />
0<br />
−1<br />
−2<br />
−3<br />
−4<br />
Error [µm]<br />
Figure 6.23: Non linearity <strong>and</strong> hysteresis<br />
measurements performed on a<br />
single actuator excited by a 4Hz sine<br />
excitation voltage with 3.3V amplitude.<br />
The response is plotted against<br />
the excitation for single (thick solid<br />
line) <strong>and</strong> the deviation from a linear<br />
response (dashed line) is plotted as<br />
the thin solid line on the right vertical<br />
axis.<br />
Finally, the estimated latency varies between approximately 160 <strong>and</strong> 170µs, of which<br />
89.7 + 39 = 128.7µs can be attributed to the serial communication (section 6.7.1). Another<br />
8µs can be attributed to communication between the master <strong>and</strong> slave FPGAs <strong>and</strong><br />
also 8µs to the implemented PWM update method that yields an average latency of half a<br />
period of the PWM base frequency. The remaining 20µs are likely caused by overhead in<br />
the XPC target computer, in which ethernet communication is performed by a background<br />
process <strong>and</strong> not strictly real-time.<br />
6.7.4 Nonlinear behavior<br />
So far only the linear system <strong>dynamics</strong> were considered, whereas it was shown in section<br />
5.2.3 that both the mechanical <strong>and</strong> magnetic stiffnesses of the actuator are nonlinear functions<br />
of current <strong>and</strong> deflection. This is particularly true for large deflections, when the<br />
difference with the operating point used for linearization becomes significant. However,<br />
measurement results show that no significant (i.e. measurable) hysteresis is present (figure<br />
6.23).<br />
Measurements were performed to quantify both effects for the single actuator system.<br />
The previously described setup depicted in figure 6.14 was used to excite the system with<br />
a low frequent sine signal <strong>and</strong> measure its deflection response. Differential measurement<br />
capabilities of the laser vibrometer were used to limit drift due to e.g. air motion <strong>and</strong> reduce<br />
external disturbances such as floor vibrations. The excitation frequency is chosen at 4Hz<br />
such that only the system’s static behavior (stiffness) plays a role <strong>and</strong> not its resonances.<br />
An amplitude of 3.3V corresponding to the maximum available input voltage is used. The<br />
sampling frequency of the measurements is chosen as 10kHz to minimize effects of aliasing<br />
<strong>and</strong> the results depicted in figure 6.23 have been compensated for the discussed latencies<br />
that yield a spurious hysteresis loop. The figure shows that hysteresis is negligible <strong>and</strong> of<br />
similar order of magnitude as the drift of the laser vibrometer. A linear function is fit to the<br />
response <strong>and</strong> shown as the dashed line. The difference with the actual response is plot as<br />
the dash-dotted line against the right y-axis. Although for large deflections the nonlinear<br />
actuator stiffness becomes visible, for the intended±10µm deflection, the actuator linearity<br />
error is less than 5%.<br />
6
6<br />
148 6 Electronics<br />
6.8 Power dissipation<br />
As stressed in chapter 2, power dissipation forms a design driver for the DM system. In this<br />
section, the power dissipation measured will be discussed.<br />
For the analysis only the static response of the system is considered, which is assumed to<br />
be linear w.r.t. the PWM voltage setpoint. The validation measurements have shown this to<br />
be an accurate approximation at least up to approximately 1600Hz. Non-static dissipative<br />
effects such as viscous damping <strong>and</strong> eddy currents play a negligible role in this frequency<br />
range. The remaining power dissipation of the single actuator system consists of several<br />
parts. Firstly, there is the power dissipated as a direct consequence to the actuator current.<br />
This flows through the actuator coil with a resistanceRa+R ′ a <strong>and</strong> the coil of the analog lowpass<br />
filter with resistanceRl +R ′ l . For the static case, the corresponding power dissipation<br />
can be expressed as:<br />
Pa = I 2 a(Ra +R ′ a +Rl +R ′ 2 Fa<br />
l) = (Ra +R<br />
Ka<br />
′ a +Rl +R ′ l),<br />
where for the second step the system was assumed to be linear such thatFa = KaIa.<br />
For the expected RMS actuator force of 1mN, a measured average motor constant Ka ≈<br />
0.12N/A <strong>and</strong> Ra + R ′ a + Rl + R ′ l ≈ 39 + 3 + 2.7 + 2.2 ≈ 46.9Ω, this corresponds to<br />
3.2mW per actuator. For the design values in tables 5.1 <strong>and</strong> 6.2 this power dissipation<br />
would be 1.4mW, which means that the actual dissipation will be approximately 2.3 times<br />
higher than expected.<br />
Besides direct dissipation of the electronics, there is also indirect dissipation. This consists<br />
of the power dissipated by the FPGAs to generate the PWM signals <strong>and</strong> h<strong>and</strong>le the communication<br />
<strong>and</strong> dissipation of the Field Effect Transistor (FET) switches of the H-bridges,<br />
LVDS drivers <strong>and</strong> voltage converters. These contributions have been quantified by measuring<br />
the supply current to a single electronics module for various configurations using a<br />
Fluke digital multi-meter.<br />
The static power dissipation of the three FPGAs is provided by the manufacturer as approximately<br />
40mW. The summed power dissipated by the master FPGA, the voltage converter<br />
<strong>and</strong> the LVDS driver has been obtained by measuring the supply current with only the master<br />
print connected. The power dissipated by the slave FPGAs that generate the PWM signals<br />
has been obtained by measuring the supply current with the resulting PWM outputs disabled.<br />
Losses in the analog part of the electronics <strong>and</strong> due to the switching of the H-bridge<br />
were obtained by measuring the supply current for various actuator setpoints, but without<br />
the actuators being connected. This prevents DC currents from flowing <strong>and</strong> allows the measuring<br />
of parasitic effects only. Further, the difference in supply current to the case that<br />
the actuators are connected can be attributed to actuator currents <strong>and</strong> resulting dissipation.<br />
Finally, for all measurements the DC-DC convertor was assumed to have an efficiency of<br />
85%, leading to the results plotted in figure 6.24.<br />
For an output voltage of 0V the dissipation consists only of the mentioned indirect losses,<br />
whereas for non-zero voltages the dissipation is proportional to the square of the voltage<br />
setpoint divided by the total resistance. The results in the figure confirm this resistance to<br />
be around 40Ω. The RMS voltage setpoint expected in practice is derived from the expected<br />
RMS actuator force of 1mN derived in chapter 2 by division by the motor constant<br />
≈ 46.9Ω, yield-<br />
Ka ≈ 0.12N/A <strong>and</strong> multiplication by the total resistanceRa+R ′ a +Rl+R ′ l
6.8 Power dissipation 149<br />
Power per actuator [mW]<br />
35<br />
30<br />
25<br />
20<br />
15<br />
10<br />
5<br />
DCDC−convertor<br />
Actuators<br />
Analog electronics<br />
Slave initialization<br />
Slaves<br />
Master<br />
0<br />
−0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8<br />
PWM voltage V [V]<br />
pwm<br />
Figure 6.24: Measured<br />
power dissipation of the<br />
single actuator system including<br />
driver <strong>and</strong> communication<br />
electronics. Results<br />
are split into the contributions<br />
of several components.<br />
ing 0.39V. Observe in figure 6.24 that for this voltage setpoint value the power dissipated<br />
in the FPGAs dominates the power dissipated in the actuator to generate force. A reduction<br />
of the total power consumption will thus be most effectively achieved by a reduction<br />
of the power consumption of the FPGAs. The FPGA implementation has been analyzed in<br />
detail by William de Bruijn, who also proposed design modifications to improve the power<br />
efficiency. His work has been documented in detail in [42], but only the main findings <strong>and</strong><br />
design proposals will be summarized.<br />
6.8.1 Optimizing the FPGA power efficiency<br />
The driver electronics developed for the DM system use three FPGAs to implement 61<br />
PWM signal generators <strong>and</strong> the LVDS communication with a <strong>control</strong> computer. The power<br />
dissipated by an FPGA can be divided into static <strong>and</strong> dynamic dissipation. The first is<br />
dissipated regardless of the program loaded or configuration, but depends highly on the<br />
specific IC. For each of the Altera ICs this dissipation is approximately 40mW. The dynamic<br />
dissipation depends on the program loaded <strong>and</strong> can be approximated as:<br />
Pdyn = Psc(fclk)+αCV 2<br />
cc fclk, (6.9)<br />
where Psc is the short-circuit power dissipation that is linearly proportional to the clock<br />
frequency fclk. Further, α denotes the switching activity (the average number of 0 to 1<br />
transitions per clock cycle), Vcc the supply voltage <strong>and</strong> C the total capacitance. The latter<br />
is a measure for the amount of hardware (transistors, interconnection wires, etc.) in use.<br />
The static power dissipation <strong>and</strong> the supply voltage Vcc are determined by the choice for<br />
a particular FPGAs. Both are likely to decrease for future models due to technological<br />
progress. A reduction of the dynamic power dissipation can be achieved by reducing the<br />
capacitance C <strong>and</strong> the total number of state switches per second expressed by the product<br />
αfclk. Such a reduction requires more insight into the distribution of the dissipation over<br />
6
6<br />
150 6 Electronics<br />
Slaves<br />
Master<br />
Clocks<br />
RAM<br />
LVDS receiver<br />
PWM counter<br />
Comparators<br />
Other<br />
0 10 20 30 40 50 60 70 80 90 100<br />
Fraction of total power dissipation [%]<br />
Figure 6.25: The power usage of the master <strong>and</strong> slave FPGAs subdivided into several functional<br />
categories.<br />
specific parts of the current FPGA program.<br />
To determine the contribution of code parts, the existing FPGA programs were first ported to<br />
a Xylinx University Program (XUP) development board. This development board contains<br />
a single Xylinx chip instead of the three Altera’s, but measurements confirmed that the<br />
dynamic FPGA power dissipation is of the same order of magnitude: approximately 550mW<br />
instead of 673mW. The latter is derived from figure 6.24 by summing the contributions of the<br />
master, the slaves <strong>and</strong> the initialization thereof atVPWM = 0. These measurements were used<br />
as a benchmark to compare with the output of the XPower simulation tool [195] by Xylinx<br />
that uses the post-place <strong>and</strong> routing net-lists of the ModelSim PE simulator [131]. Since the<br />
simulation results corresponded well to the measurements [42], the more detailed results<br />
of the XPower tool were trusted to provide an accurate subdivision of the total dissipation.<br />
This is shown graphically in figure 6.25 for the master <strong>and</strong> slave FPGAs. Observe that the<br />
dissipation of the master FPGA can be mainly attributed to a RAM module <strong>and</strong> that the<br />
dissipation of both FPGA types is for over 50% attributable to clock signals. This means<br />
that power dissipation is reduced by:<br />
1. Removing the R<strong>and</strong>om Access Memory (RAM) module.<br />
2. Reducing the clock frequencyfclk.<br />
3. Reducing the amount of clocked hardware.<br />
The RAM module is currently used to buffer the incoming LVDS messages, which is not<br />
necessary. According to (6.9) the dynamic dissipation is linearly proportional to the clock<br />
frequency fclk. A reduction is not possible for all functionality of the FPGA, since this is<br />
linked to the PWM base frequency (section 6.2.2), but for some parts it is. The dissipation<br />
due to clock signals can be significantly reduced using an asynchronous design in which<br />
logic cells are synchronized using local h<strong>and</strong>shakes instead of global clock signals. Such<br />
an asynchronous design can be efficiently written in a parallel language such as Tangram<br />
or HASTE [93], but the achievable reduction in dissipation depends highly on the software<br />
tools used for mapping this code onto a specific FPGA. For maximum effect the design<br />
should be mapped to an ASIC instead of an FPGA. Nevertheless, three new designs were<br />
proposed that have been conceptually analyzed <strong>and</strong> evaluated using both simulations <strong>and</strong>
6.8 Power dissipation 151<br />
implementations on the XUP development board. This has lead to the following observations:<br />
• The master functionality can be efficiently expressed in an asynchronous language.<br />
After implementation without RAM module, a reduction in the power dissipation of<br />
the master FPGA of 40% was measured.<br />
• Although asynchronous designs have a high potential for power reduction, this cannot<br />
be fully exploited by implementation on FPGAs, because these devices are inherently<br />
synchronous <strong>and</strong> must in fact simulate asynchronous programs.<br />
• The slave functionality was most efficiently implemented using a recursive PWM<br />
driver designs. This design has no central counter, but uses a counter for each channel.<br />
A power reduction is achieved via a recursive counter design in which the bits are<br />
clocked at their rate of change (lower bits have a higher clock frequency).<br />
• A combination of a recursive counter implementation for each PWM driver <strong>and</strong> the<br />
asynchronous master lead to a reduction in overall power dissipation of approximately<br />
29%. This reduction is significantly larger when the design is mapped to an ASIC.<br />
6.8.2 Cooling<br />
The improvements proposed in section 6.8.1 are not yet implemented. The power dissipation<br />
in the FPGAs will therefor still exceed the power dissipation of the coils by far. To<br />
avoid this heat to be transferred to the ambient air with detrimental air flow in the path of<br />
light as a result, a possibility for active cooling is added. This cooling system consists of an<br />
aluminium fin which is placed between the master <strong>and</strong> two slave FPGAs. The aluminium<br />
fan is connected with a the aluminium block, which holds the cooling channels. The cooling<br />
liquid cools the block <strong>and</strong> thereby the aluminium fan. In case of, for example, an electronic<br />
failure, the PCB together with the aluminium fin can be disconnected from the cooling<br />
system without disconnection of the fluid system. The flow is regulated by a commercial<br />
Central Processing Unit (CPU) cooling system. The aluminium block is suspended in a thin<br />
plate to allow dimensional change of the PCB as well as tolerances on the PCB size. In<br />
figures 6.27 <strong>and</strong> 6.26 the active cooling system for the FPGAs is shown.<br />
Figure 6.26: The cooling system seen from the<br />
back.<br />
Figure 6.27: The cooling system seen from the<br />
front.<br />
6
6<br />
152 6 Electronics<br />
6.9 Conclusions<br />
The electronics for the prototype DMs consist of two parts: the communication electronics<br />
that supply the setpoints as computed by the <strong>control</strong> system <strong>and</strong> the driver electronics that<br />
generate the corresponding actuator voltages. The requirements for both part are derived.<br />
Both current <strong>and</strong> voltage drivers are considered. Current <strong>mirror</strong>s were not implemented because<br />
of their complexity to produce them <strong>and</strong> the lower efficiency, especially with the large<br />
dynamic range required. Since theLa/Ra time of the actuator is short: 75µs, the advantage<br />
of current <strong>control</strong> over voltage <strong>control</strong> is limited. The motor constantKa, stiffnessca <strong>and</strong> –<br />
in the case of a voltage source – the resistance R of the actuator circuit, will vary from actuator<br />
to actuator <strong>and</strong> vary with temperature, causing slow gain variations. A current source<br />
will compensate for variations in the resistance R, but variations of Ka <strong>and</strong> ca must still<br />
be compensated by the AO <strong>control</strong> system. PWM based voltage drivers are chosen because<br />
of their high efficiency <strong>and</strong> capability to be implemented in large numbers with only a few<br />
electronic components.<br />
A LVDS based serial communication bus was chosen for its low power consumption<br />
(15mW/transceiver), high b<strong>and</strong>width (up to 655Mb/s) <strong>and</strong> consequently low latency, low<br />
communication overhead <strong>and</strong> extensive possibilities for customization. The driver electronics<br />
for 61 actuators are located on a single, multi-layer PCB <strong>and</strong> consist of FPGAs to<br />
generate the PWM signals, FETs for the H-bridge switches <strong>and</strong> a coil/capacitor pairs that<br />
form 2 nd order low-pass filters. The FPGAs that generate the PWM signals also <strong>control</strong><br />
two LVDS communication connections – one up- <strong>and</strong> one downlink – to receive setpoint<br />
updates <strong>and</strong> to report status information. A 16-wire flat-cable connects up to 32 electronics<br />
modules to a custom designed communications bridge, which translates ethernet packages<br />
into LVDS packages <strong>and</strong> vice versa. The ethernet side of the communications bridge is<br />
connected to the <strong>control</strong> computer at a speed of 100Mbit/s <strong>and</strong> uses the UDP protocol to<br />
minimize overhead <strong>and</strong> latency.<br />
The actuator model from chapter 5 was extended with models for the communication <strong>and</strong><br />
driver electronics. The communication is modeled as a pure delay <strong>and</strong> the driver electronics<br />
as a voltage source with an analog 2 nd order low-pass filter. The model is used<br />
to select a suitable PWM base frequency for which the position response from the voltage<br />
ripple due to higher harmonics of the PWM signal is less than a quarter of the LSB<br />
of the setpoint. This frequency should be higher than 40kHz for the DMs with Pyrex<br />
facesheets, but is set at 61kHz to be suited for the replacement of these facesheets by<br />
beryllium. The actuator model including its communication <strong>and</strong> electronics was validated<br />
by measurements. The measurements include communication tests, static <strong>and</strong> dynamic<br />
response measurements <strong>and</strong> power dissipation measurements. It is shown that the<br />
communication latency is well represented by τc = 89.7 · 10 −6 + 39 · 10 −6 Nm, where<br />
Nm is the number of the actuator grid. With the actuator response measurements, actuator<br />
properties as stiffness (ca =473±46N/m), motor constant (Ka =0.11±0.02N/A),<br />
damping (ba =0.30±0.11mNs/m), inductance(La =3.0±0.2mH) <strong>and</strong> resonance frequency<br />
(fe =1.83±91Hz) are verified. These properties showed some variation between actuators,<br />
but this could not be attributed to the location of the actuator in the grid (appendix F).<br />
The time domain response of an actuator to a 4Hz sine voltage was used to determine hysteresis<br />
<strong>and</strong> semi-static nonlinear response of the actuator. This showed the first to be negligible<br />
<strong>and</strong> the second to remain below 5% for the intended±10µm stroke.
6.9 Conclusions 153<br />
Finally, power dissipation was measured. Unintended resistances in the paths between the<br />
voltage source <strong>and</strong> the actuator, combined with the lower motor constant showed to lead<br />
to 2.3 times higher power consumptions of the actuators: 3.2mW instead of 1.4mW. Measurements<br />
also showed that in the expected operating range, the total power dissipation is<br />
dominated by indirect losses in the FPGAs. An alternative FPGA implementation is investigated.<br />
A reduction of 40% in the master FPGA <strong>and</strong> 29% in the slave FPGAs is thereby<br />
achieved.<br />
6
ÔØÖ×ÚÒ<br />
ËÝ×ØÑÑÓÐÒÒ<br />
ÖØÖÞØÓÒ<br />
The developed actuator modules (chapter 5) <strong>and</strong> electronics (chapter 6) are<br />
integrated with the reflective facesheet [174] to form a complete Deformable<br />
Mirror (DM) system. The static <strong>and</strong> dynamic system behavior is modeled <strong>and</strong><br />
compared to measurement results. The reflective <strong>deformable</strong> facesheet, which<br />
couples all actuators, is modeled with a biharmonic plate equation <strong>and</strong> an analytic<br />
solution for the surface shape under a regular actuator grid is found.<br />
The model is used to derive the actuator influence functions. The static model<br />
is extended with lumped masses to include the dynamic behavior. From the<br />
model, the transfer functions, impulse response functions <strong>and</strong> mode shapes are<br />
derived. The verification of the static behavior of the DM system is done using<br />
an interferometer setup. The dynamic system identification is performed using<br />
white noise excitation on the actuators <strong>and</strong> displacement <strong>and</strong> velocity measurement<br />
of the <strong>mirror</strong> facesheet with a laser vibrometer. With these measurements<br />
the model modal analysis is compared with the measurements.<br />
sections 7.2, 7.3 <strong>and</strong> 7.4 are joint work with Roger Hamelinck<br />
155
7<br />
156 7 System modeling <strong>and</strong> characterization<br />
7.1 Introduction<br />
In this chapter the developed actuator modules (chapter 5) <strong>and</strong> electronics (chapter 6) will be<br />
combined with a reflective facesheet (for design considerations, see [174]) into a complete<br />
prototype Deformable Mirror (DM) system. First, the integration of these parts is described,<br />
after which the behavior of the DM system will be analyzed. An analytical model for the<br />
reflective facesheet is derived that – combined with the DC-gain of the single actuator system<br />
provides a static model for the DM system. This model describes the actuator influence<br />
functions that will be compared with measurements on the DM system. Further, measurement<br />
results are presented that show the initial flatness of the DM <strong>and</strong> its ability to form<br />
Zernike mode shapes. Finally, the Direct Current (DC) model is used to determine the expected<br />
average power dissipation of the DM when correcting Kolmogorov type wavefront<br />
disturbances.<br />
An analytical dynamic model for the system is then derived based on the available model<br />
for a single actuator system from chapter 6. From this model the expected resonance frequencies<br />
<strong>and</strong> modal shapes are derived that will be compared to measurement results on<br />
the DM system. Finally, the dynamic behavior of the DM system will be evaluated w.r.t.<br />
discrete time <strong>control</strong> aspects.<br />
7.2 DM integration<br />
A single actuator grid with 61 actuators is integrated with a 100µ thick ∅50mm Pyrex<br />
facesheet <strong>and</strong> a single Printed Circuit Board (PCB) to form the first prototype. In a second,<br />
larger prototype, 7 actuator modules are placed on a reference base <strong>and</strong> connected to a<br />
single,∅150mm Pyrex facesheet.<br />
7.2.1 Integration of the 61 actuator <strong>mirror</strong><br />
In figure 7.1 the integration the first DM prototype is shown at different stages. The actuator<br />
struts are first connected to the back of the <strong>mirror</strong> facesheet <strong>and</strong> then connected to<br />
the actuator grid. With the struts attached, only the out-of-plane DOFs of the facesheet are<br />
constrained. The in-plane DOFs are still free <strong>and</strong> will be constrained by the three folded<br />
leafsprings shown in figure 7.1. The folded leafsprings are placed in their aluminium mount<br />
<strong>and</strong> glued to the backside of the facesheet. Figure 7.2 shows the DM with the folded leafsprings<br />
in place <strong>and</strong> figure 7.3 shows the 61 prototype DM including its electronics <strong>and</strong><br />
protective cover.<br />
7.2.2 Integration of the 427 actuator <strong>mirror</strong><br />
The second DM prototype with 427 actuators is assembled similar to the single actuator<br />
grid DM. First the 7 actuator grids are placed on a reference base (figures 7.4 <strong>and</strong> 7.5). The<br />
corrugated edges of the actuator grids are separated by 0.3mm. The base is made from a<br />
40mm thick aluminium block, perforated with 7 large holes (∅30mm) <strong>and</strong> a pattern of small<br />
holes to accommodate the M2-bolts to attach the A-frames that connect to the actuator grids.<br />
These bolts are mounted from the back. The base itself is placed vertically <strong>and</strong> supported<br />
by three larger A-frames.
7.2 DM integration 157<br />
The PCBs with the driver electronics are placed in one box (figure 7.6). The PCBs are<br />
mounted similar to figures 6.27 <strong>and</strong> 6.26. Via the slits in the front plate, the flex foils<br />
connect to the connector boards <strong>and</strong> PCBs. Figure 7.7 shows the electronic box connected<br />
to the actuator grids.<br />
The electronic box is decoupled before the <strong>mirror</strong> facesheet is connected to the actuator<br />
grids by means of the actuator struts <strong>and</strong> small droplets of glue. First the connection<br />
struts are glued to the backside of the <strong>mirror</strong> facesheet, during which the <strong>mirror</strong> facesheet<br />
is supported by a porous air bearing. After curing of the glue, the <strong>mirror</strong> facesheet with<br />
the struts is glued to the actuator modules. Details on the procedure can be found in [174].<br />
Finally, the folded leafsprings needed to constrain the in-plane DOFs are placed (figure<br />
7.8) <strong>and</strong> the flex foils are connected to the electronics (figure 7.9). The <strong>mirror</strong> is now fully<br />
assembled, except for the protective cover ring. Unfortunately, while placing the cover ring<br />
the <strong>mirror</strong> got damaged before any measurements could be obtained from the completed<br />
DM system. The results presented in the sequel of this thesis originate from the single<br />
actuator module prototype shown in figure 7.3.<br />
Figure 7.1: The DM prototype with 61 actuators shown during final assembly. The upper left figure<br />
shows 61 struts attaching the <strong>mirror</strong> facesheet to the actuator module. The module is<br />
connected to the (black) base with three A-frames. The flex foil is fed through a central<br />
hole in the base. On the right, the folded leafsprings that constrain the facesheet’s inplane<br />
DOFs are shown prior to assembly. In the lower left, one of the folded leafsprings<br />
is located a little below the <strong>mirror</strong> facesheet, before it is translated to make the glued<br />
connection with the facesheet.<br />
7
7<br />
158 7 System modeling <strong>and</strong> characterization<br />
7.3 Static system validation<br />
In this section, the static behavior of the DM is modeled, providing a description for<br />
the actuator influence functions. This involves modeling of the <strong>mirror</strong> facesheet <strong>and</strong><br />
combining this with the static model of the actuator system from chapter 6. The actual<br />
influence functions of the DM system are measured using a Wyko interferometer to which<br />
the modeled influence functions are compared. This was done for all 61 actuators in<br />
the ∅50mm DM. Further, the influence matrix derived from the measurements is used<br />
to form the <strong>mirror</strong> facesheet into the first 28 Zernike mode-shapes including the piston<br />
term that represents the best flattened <strong>mirror</strong> [86, 87]. The measured shapes are compared<br />
to the perfect Zernike modes <strong>and</strong> to the least square fit based on the DC model derived.<br />
Figure 7.2: The 61 actuator DM. The protective cover is not shown, to see the inner parts. The<br />
connector board, described in section 6.4.2 is shown in the lower right photo.
7.3 Static system validation 159<br />
7.3.1 Modeling<br />
Figure 7.3: The 61 actuator DM including<br />
its electronics.<br />
First a model for the reflective facesheet is derived, leading to an expression for the influence<br />
function matrix. Finally, this matrix is used in an algorithm to calculate the actuator<br />
comm<strong>and</strong>s that provide a facesheet that best approximates a certain Zernike shape in a least<br />
squares or absolute-error sense.<br />
Facesheet modeling<br />
The <strong>mirror</strong> facesheet is modeled as a circular plate with free edges, subjected to point forces.<br />
Although the facesheet has a large diameter to thickness ratio, the facesheet is still a plate<br />
with significant bending stiffness, particularly on the spatial scale of the actuator pitch. In<br />
contrast to a true membrane, there is no pre-tension from which it derives its stiffness <strong>and</strong><br />
resonance frequency. Since the connection struts are only 100µm thick – which is small<br />
in comparison to the pitch – the forces exerted on the reflective facesheet are considered<br />
to be point-forces. The in-plane stiffness is provided at the circumference by three folded<br />
Figure 7.4: Seven actuator modules placed on<br />
the reference base. The actuator grids are separated<br />
by 0.3mm.<br />
Figure 7.5: The backside of the reference base<br />
with the 7 actuator modules mounted. The flex<br />
foils are visible through the larger holes.<br />
7
7<br />
160 7 System modeling <strong>and</strong> characterization<br />
Figure 7.6: The 7 PCBs with driver electronics assembled in the electronics box. The cooling, similar<br />
to figure 6.27 is visible. Via the slits in the front plate, the flex foils connect to the<br />
connector boards <strong>and</strong> PCBs.<br />
leaf springs at 120 ◦ intervals. The out-of-plane stiffness that these springs contribute is<br />
negligible in comparison to the facesheet <strong>and</strong> can therefore be neglected . The edge of the<br />
facesheet is thus considered to be free.<br />
Let r (i) <strong>and</strong> ρ (j) for i = 1...Nr <strong>and</strong> j = 1...Na be complex values corresponding to<br />
coordinates in the complex plane. The deflection zf(r (i)) at coordinate r (i) of a circular,<br />
Figure 7.7: The DM ready for testing. Each actuator is tested individually (section 6.7.3). After<br />
testing the facesheet is assembled.
7.3 Static system validation 161<br />
Figure 7.8: Detail of one of the three folded leafsprings<br />
that constrain the in-plane DOFs.<br />
Figure 7.9: The assembled DM with 427 actuators.<br />
Figure 7.10: The broken DM.<br />
thin plate of Hookean material, with radiusrf <strong>and</strong> free edge conditions due to a point force<br />
F (j) located at ρ (j) can be derived from the biharmonic plate equation [177] in terms of the<br />
Laplacian operator∇ 2 <strong>and</strong> the plate’s flexural rigidityDf as:<br />
where ∇ 2 (r) =<br />
δ 2<br />
δRe 2 (r) +<br />
∇ 4 zf(r) = F (j)r2 f<br />
,<br />
Df<br />
δ 2<br />
δIm 2 (r) , <strong>and</strong> Df =<br />
Eft 3 f<br />
12(1−ν 2 f ).<br />
Further, Re(r) <strong>and</strong> Im(r) denote the real <strong>and</strong> imaginary parts of r respectively, Ef is the<br />
plate material’s Young’s modulus, νf its Poisson ratio <strong>and</strong> tf its thickness. The deflection<br />
Z(r (i)) can be expressed analytically in terms ofF (j) as [122]:<br />
Z(r (i)) = F (j)r2 f<br />
W(r (i),ρ (j))+wp +wxRe(r (i))+wyIm(r (i)),<br />
16πDf<br />
7
7<br />
162 7 System modeling <strong>and</strong> characterization<br />
Table 7.1: Dimensions <strong>and</strong> material properties of the reflective facesheets of the 61 <strong>and</strong> 427 actuator<br />
DM prototypes.<br />
Parameter Value Unit<br />
rf for the 61 actuator DM 25.4 mm<br />
rf for the 427 actuator DM 76.2 mm<br />
Ef 64 GPa<br />
ρf 2230 kg/m 3<br />
tf 100 µm<br />
νf 0.2 -<br />
wherewp,wx <strong>and</strong>wy denote the rigid body motions in the out-of-plane direction <strong>and</strong> around<br />
thex- <strong>and</strong>y axes respectively. The functionW(r (i),ρ (j)) is defined as:<br />
W(r (i),ρ (j)) = ̺ (i,j)̺ ∗ (i,j)<br />
+<br />
<br />
ln(̺ (i,j))+ln(̺ ∗ (i,j) )+1–νf<br />
<br />
ln(1–r(i)ρ<br />
3+νf<br />
∗ (j) )+ln(1–r∗ (i) ρ (j)) <br />
(1–νf) 2<br />
(1+νf)(3+νf) r (i)r ∗ (i) ρ (j)ρ ∗ (j) +<br />
8(1+νf)<br />
(1–νf)(3+νf)<br />
+k(r (i)ρ ∗ (j) )+(1–r∗ (i) ρ (j))ln(1–r ∗ (i) ρ (j))+k(r ∗ (i) ρ (j))<br />
where the superscript ∗ denotes the complex conjugate <strong>and</strong><br />
̺ (i,j) = r (i) −ρ (j) <strong>and</strong> k(x) =<br />
x<br />
0<br />
<br />
(1–r (i)ρ ∗ (j) )ln(1–r (i)ρ ∗ (j) )<br />
<br />
,<br />
ln(1−ς)<br />
dς = −dilog(1−x).<br />
ς<br />
This analytic expression allows the spatial grids to be discretized without loss of accuracy.<br />
The values of the geometric <strong>and</strong> material parameters for the 61 <strong>and</strong> 427 actuator DM prototypes<br />
can be found in table 7.1.<br />
Influence function modeling<br />
The shape of the influence functions depend on the stiffness of the facesheet <strong>and</strong> actuators<br />
<strong>and</strong> the lay-out of the actuator grid. Linearity is assumed to allow linear superposition of<br />
multiple point forces. For convenience, matrix-vector notation is used, where matrices are<br />
set in a bold typeface.<br />
Let z f,(i) = Z(r (i)), z a,(i) = Z(ρ (i)) <strong>and</strong> F ρ,(j) = F (j) be elements of the vectors zf , za<br />
<strong>and</strong> Fρ respectively. Similarly, the coordinatesr (i) <strong>and</strong> ρ (j) form the i th <strong>and</strong> j th elements<br />
of the vectors r <strong>and</strong> ρ <strong>and</strong> Ω rρ,(i,j) = w(r (i),ρ (j)) the elements of the matrix Ωrρ. The<br />
facesheet deflectionzf can then be expressed as:<br />
zf = ΩrρFρ +Urwpxy, (7.1)<br />
where Ur = [1 Nr×1 Re(r) Im(r)] <strong>and</strong> wpxy = [wp wx wy] T . The reflective facesheet is<br />
supported by actuators with effective mechanical stiffnesses that exert forces denoted by the
7.3 Static system validation 163<br />
vectorFa. Since these stiffnesses can be considered linear (section 6.7.4 on page 147), the<br />
following force equilibrium must be satisfied at the actuator locations in the vectorρ:<br />
Fa −Caza −Fρ = 0, (7.2)<br />
where Ca is a diagonal matrix whose i th diagonal element is the stiffness ca of actuator<br />
i <strong>and</strong> it is assumed that the facesheet deflection at the actuator locations is equal to the<br />
actuator deflectionza.<br />
Since the rigid body modes are not constrained by the free edge condition of the plate, the<br />
moments due to the net plate forces Fρ around the x <strong>and</strong> y axes <strong>and</strong> the net force in the<br />
out-of-plane direction should be zero. This leads to the extra condition U T ρ Fρ = 0, where<br />
Uρ is defined similar to Ur as Uρ = [1 Nr×1 Re(ρ) Im(ρ)]. When (7.1) is evaluated only<br />
on the actuator grid – i.e. r = ρ – it can be expressed as za = ΩρρFρ +Uρwpxy, where<br />
Ωρρ = Ωrρ| r=ρ . Together with the rigid body constraint this can be written in matrix form<br />
as: <br />
za<br />
=<br />
0<br />
which can be inverted to:<br />
Fρ<br />
wpxy<br />
<br />
=<br />
<br />
Ωρρ Uρ<br />
UT <br />
Fρ<br />
,<br />
ρ 0 wpxy<br />
Ωρρ Uρ<br />
U T ρ 0<br />
−1 <br />
za<br />
=<br />
0<br />
Km<br />
Kz<br />
<br />
za, (7.3)<br />
in which the matrices Km <strong>and</strong> Kz are implicitly defined. Substitution of this result for Fρ<br />
into the force equilibrium in (7.2) then yields:<br />
Fa −Caza −Kmza = 0<br />
<strong>and</strong> thus the facesheet deflectionza at the actuator positions is related to the actuator forces<br />
Fa as<br />
za = (Km +Ca) −1 Fa, (7.4)<br />
The static forceFa of a certain actuator due to a supplied Pulse Width Modulation (PWM)<br />
voltageVPWM can be expressed as the quotient of the motor constantKa <strong>and</strong> the total electric<br />
resistance: Fa = Ka/(Ra +R ′ a +Rl +R ′ l )VPWM. This can be written in vector notation<br />
for all actuators as:<br />
Fa = Ka(Ra +R ′ a +Rl +R ′ l )−1 VPWM, (7.5)<br />
where the ith diagonal elements of the diagonal matrices Ka, Ra, R ′ a, Rl <strong>and</strong> R ′ l are the<br />
values of the corresponding (regularly typefaced) symbols for all actuators i = 1...Na.<br />
Further, the vector VPWM stacks the PWM voltages VPWM of all actuators. Substitution of<br />
(7.5) into (7.4) then leads to:<br />
VPWM, (7.6)<br />
<br />
za = (Km +Ca) −1 Ka(Ra +R ′ a +Rl +R ′ l) −1<br />
Bρ<br />
where Bρ is the influence matrix that links PWM voltages to facesheet deflection at the<br />
actuator locations. The plate deflections due to point forces at positionsρ can also be evaluated<br />
over the arbitrary grid with complex coordinate vector r. The results from (7.3) can<br />
7
7<br />
164 7 System modeling <strong>and</strong> characterization<br />
LVDS<br />
Ethernet<br />
LVDS<br />
bridge<br />
Driver<br />
electronics<br />
UDP<br />
DM<br />
Ethernet<br />
switch<br />
setpoint<br />
UDP<br />
PC<br />
Wyko 400<br />
interferometer<br />
Intelliwave TM<br />
software<br />
Figure 7.11: The measurement setup which is used to measure the influence functions of the DM.<br />
be substituted into the plate equation in (7.1) together with (7.4), yielding:<br />
zf = (ΩrρKm +UrKz)Fa<br />
= (ΩrρKm +UrKz)Bρ VPWM, (7.7)<br />
<br />
Bf<br />
whereBf is the influence matrix that links PWM voltages to deflections at an arbitrary grid<br />
of points on the facesheet.<br />
7.3.2 Measurements <strong>and</strong> results<br />
This section describes the setup <strong>and</strong> procedure used to measure the <strong>mirror</strong>s influence functions<br />
<strong>and</strong> low order Zernike modes.<br />
Interferometric measurement setup<br />
The verification of the static behavior of the DM system is done using an interferometer<br />
setup. A Wyko 400 interferometer available at TNO Science <strong>and</strong> Industry measures the<br />
surface shape of the DM. Intelliwave TM software is used to perform the reconstruction of<br />
the actual wavefront from the measured fringe patterns. Figure 7.11 shows the schematic<br />
of the measurement setup, where a PC sends desired setpoint comm<strong>and</strong>s VPWM via the<br />
ethernet/Low Voltage Differential Signalling (LVDS) communication link to the DM. All<br />
shapes <strong>and</strong> measurements in the coming sections are considered w.r.t. an arbitrary grid as<br />
determined by the interferometer’s Charge Coupled Device (CCD) camera. Since the interferometer<br />
cannot observe the piston mode corresponding to a non-zero average deflection, it<br />
is assumed that all measurementsˆz are piston-free. When considering the zero-mean, white<br />
measurement noisen, this allows the measurementsˆzf ∈ R Nw to be expressed as:<br />
ˆzf = Pzf +Pn, (7.8)<br />
where the rank deficient matrix P = I −pp T projects out the piston term denoted by the<br />
vectorp whose elements are all equal to1/ √ Nw s.t. p T p = 1.
7.3 Static system validation 165<br />
Assuming the static response of the DM to be linear, let the shape zf of the DM facesheet<br />
be expressed as:<br />
zf = Bf,wVPWM +zf,0,<br />
wherez f,(0) is the initial unactuated shape of the DM facesheet <strong>and</strong> the matrix Bf,w is the<br />
influence matrix Bf w.r.t. the measurement grid of the Wyko interferometer. Substitution<br />
of this expression for zf into (7.8) then yields the measurement corresponding to a certain<br />
actuator comm<strong>and</strong>VPWM as:<br />
ˆzf = PBf,wVPWM +Pzf,0 +Pn (7.9)<br />
This measurement equation will be used in the following two subsections to estimate the<br />
influence function matrixBf,w <strong>and</strong> fit the DM facesheet to a desired set of shapes.<br />
Influence function measurements<br />
As described in section 7.3.1, the influence functions are the static responses of the DM to<br />
actuator comm<strong>and</strong>s. Analytically they are expressed over the actuator grid by the matrix<br />
Bρ in (7.6) <strong>and</strong> over an arbitrary grid asBf in (7.7). In this section the method is described<br />
that is used to measure the influence functions of the DM prototypes. The most obvious<br />
method to determine the influence functions is to individually poke each actuator, measure<br />
the response <strong>and</strong> then compute the influence function. Multiple measurements must be used<br />
per actuator to reduce the measurement noise <strong>and</strong> at least two different comm<strong>and</strong> values<br />
are required to determine the influence function as a linear relation between comm<strong>and</strong> <strong>and</strong><br />
deflection. More comm<strong>and</strong> values <strong>and</strong> measurements can be used to distinguish any nonlinear<br />
behavior. Each Wyko measurement takes approximately 8 seconds including data<br />
processing. When several comm<strong>and</strong>s are used for each actuator <strong>and</strong> the number of actuators<br />
is large (e.g. 427) the measurements would take several hours to complete. And even then,<br />
each influence function has to be estimated from only a few measurements, still leading to a<br />
high sensitivity to measurement noise. Better <strong>and</strong> more efficient methods can be used. For<br />
instance, in [92, 109] columns of scaled Hadamard matrices [22] are used as the actuator<br />
comm<strong>and</strong> vectors. This setpoint choice will minimize the mean st<strong>and</strong>ard deviation of the<br />
estimation error of the influence matrix due to measurement noise <strong>and</strong> thus requires fewer<br />
measurements.<br />
All elements of a Hadamard matrixQn ∈ Rn×n are either 1 or -1 <strong>and</strong> the matrix is orthogonal<br />
s.t. QnQT n = nI. Although it is yet unknown whether Hadamard matrices exist for all<br />
n ∈ N + , algorithms are available for specific dimensions. When an algorithm is unavailable<br />
for n = Na, it is argued in [109] that virtual actuators can be added that do not influence<br />
the DM shape, but allow the use of a larger Hadamard matrix of size Nav > Na that does<br />
exist at the cost of additional measurements.<br />
Accordingly, for the 61 actuator DM prototype a 64 × 64 Hadamard matrix is used that<br />
is generated by Matlab’sÑÖfunction. For the 427 actuator DM, the 428 × 428<br />
Hadamard matrix derived in [111] can be used. In the procedure described below, the influence<br />
matrix ˆ Br is estimated that includes influence functions of both the real <strong>and</strong> virtual<br />
actuators. After estimation, the columns corresponding to the virtual actuators are ignored.<br />
By individual scaling of the rows of the Hadamard comm<strong>and</strong> matrix it is possible to compensate<br />
for the lower stiffness at the edges of the DM that would lead to a larger deflection<br />
7
7<br />
166 7 System modeling <strong>and</strong> characterization<br />
Figure 7.12: The 61 influence functions of the DM prototype shown in figure 7.3. Each influence<br />
function is downsized <strong>and</strong> placed on the location of the corresponding actuator in the<br />
grid.<br />
than at the center. This is done for optimal use of the measurement range of the interferometer<br />
<strong>and</strong> avoid large deflections that would exceed the measurement range.<br />
Let measurements be expressed as in (7.9) using actuator setpoints VPWM taken as the<br />
columns of the matrixV that consist of the vectorsv (i) s.t.<br />
V = [v (0) v (1) ... v (Nav)] ∈ R Nav×Nav+1 .
7.3 Static system validation 167<br />
Herev (0) = 0 is a zero voltage comm<strong>and</strong> vector that is included to allow direct estimation of<br />
the unactuated shapez (0). Further,v (i) = Λq (i) fori = 1...Nav are the comm<strong>and</strong> vectors<br />
based on the Hadamard matrixQNav, whereQNav = [q (1)...q (Nav)]. The diagonal matrix<br />
Λ scales the rows of the Hadamard matrix QNav <strong>and</strong> provides the mentioned individual<br />
setpoint gains for all actuators. This matrix will be derived from the DM system model.<br />
For the influence functions identification, (7.9) becomes:<br />
ˆz f,(i) = Pz f,(0) + ˜ Bf,wv (i) +Pn (i),<br />
where ˜ Bf,w = PBf,w. Before stating the estimation problem, let all measurements be<br />
expressed in matrix form as:<br />
ˆZf = ˜ Bf,wV+Pz f,(0)1 T + Ñ = <br />
<br />
Bf,w ˜ 0 ΛQNav<br />
Pzf,(0) 1 1<br />
X<br />
T<br />
<br />
+<br />
<br />
Υ<br />
Ñ,<br />
where<br />
ˆ Zf = [ˆz f,(0) ˆz1...ˆzNav],<br />
Ñ = P[n (0) n (1) ... n (Nav)],<br />
<strong>and</strong> all elements of the column vectors1 <strong>and</strong>0are equal to 1 <strong>and</strong> 0 respectively.<br />
The matrixXof unknowns is estimated as:<br />
<br />
ˆX = argmin Tr ˆZf −XΥ ˆZ T<br />
f −Υ<br />
X T X T<br />
,<br />
= ˆ <br />
T<br />
ZfΥ ΥΥ T −1<br />
,<br />
= ˆ <br />
−<br />
Zf<br />
1<br />
Nav 1TQT NavΛ−1 <br />
1<br />
,<br />
0<br />
1<br />
Nav QT Nav Λ−1<br />
where the second step follows from a completion of squares argument <strong>and</strong> the last follows<br />
from the orthogonality of the Hadamard matrixQNav. The sought estimate of the influence<br />
matrix forms the first Na columns of ˆ X <strong>and</strong> that of the unactuated shape Pz f,(0) the last<br />
column.<br />
Although the Hadamard matrix approach has minimal sensitivity to measurement noise, it<br />
assumes linearity of the DM system <strong>and</strong> does not lead to an overdetermined set of equations<br />
from which to estimate the influence matrix. As a result, neither the quality of the estimate<br />
nor the linearity of the DM can be verified using criteria such as the Variance Accounted<br />
For (VAF). Such information can be obtained by extending the set of measurements, e.g.<br />
by repeating the measurements using scaled versions of the actuator setpoints.<br />
Finally, it should be noted that the same procedure can be used to determine the influence<br />
functions from measurements of a Shack-Hartmann sensor.<br />
Influence function results<br />
The above described procedure is used on the 61 actuator DM prototype, where the comm<strong>and</strong><br />
vector scaling matrixΛwas obtained as:<br />
<br />
γwdiag(Bρ)<br />
Λ =<br />
−1 <br />
0<br />
,<br />
0 I<br />
7
7<br />
168 7 System modeling <strong>and</strong> characterization<br />
Here, the bottom-right identity matrix corresponds to the three virtual actuators <strong>and</strong> is of<br />
size 3 × 3. The scalar γw determines the range of the facesheet deflections <strong>and</strong> thus the<br />
measurement range of the interferometer. A value γw = 2µm has been used. The matrix<br />
Bρ was computed for this DM from (7.6) by substitution of the relevant parameter values in<br />
tables 5.1, 6.2 <strong>and</strong> 7.1. Further, diag(·) denotes the diagonal operator that sets all elements<br />
of the matrix between brackets to zero except for the diagonal entries.<br />
The estimated influence functions are shown in figures 7.12, 7.13 <strong>and</strong> 7.14. Figure 7.12<br />
shows the 61 influence functions downsized <strong>and</strong> placed on the location of the corresponding<br />
actuator. The facesheet deflection due to a unit voltage increases for actuators near the edge<br />
of the facesheet, which is the result of the decreased stiffness due to the facesheet boundary<br />
<strong>and</strong> the smaller number of surrounding actuators. As can be expected from a hexagonal<br />
actuator layout, a 60 ◦ symmetry is observed.<br />
Figures 7.13 <strong>and</strong> 7.14 show the cross sections over two indicated axes of the measured <strong>and</strong><br />
modeled influence functions. From the right figure the actuator coupling η can be well<br />
observed as the ratio between the the deflection value at a radius of 6mm (a single actuator<br />
spacing) from the maximum <strong>and</strong> the maximum deflection value. For the central actuator<br />
this leads to η ≈ 0.52, whereas for the edge actuators this reduces to η ≈ 0.3 due to the<br />
reduced facesheet stiffness at the edge. Although the influence function measured for the<br />
center actuator matches almost perfectly to the one derived from the model, the errors vary<br />
per actuator. This is attributed to the measured variation between actuators in properties that<br />
determine its DC-gain, i.e. motor constantKa, stiffnessca <strong>and</strong> the total electrical resistance<br />
Ra +R ′ a +Rl +R ′ l .<br />
Zernike mode measurements <strong>and</strong> results<br />
With superposition of the influence functions, the DM facesheet can be fit to a desired shape.<br />
This shape may be entirely flat, in which case the actuators must compensate for the initial<br />
unflatness of the DM. Here it is only possible to correct for spatial frequencies up to the<br />
Displacement [µm]<br />
2.5<br />
2<br />
1.5<br />
1<br />
0.5<br />
0<br />
Measurements<br />
Model<br />
−20 −10 0 10 20<br />
Radius [mm]<br />
Figure 7.13: The cross-section of the five influence<br />
functions of figure 7.12 along the x-axis.<br />
The thick lines represent the functions as derived<br />
with the model.<br />
Displacement [µm]<br />
2.5<br />
2<br />
1.5<br />
1<br />
0.5<br />
0<br />
Measurements<br />
Model<br />
−20 −10 0 10 20<br />
Radius [mm]<br />
Figure 7.14: The cross-section of nine influence<br />
functions in figure 7.12 along the axis rotated 30degree<br />
counter-clockwise. The thick lines represent<br />
the functions as derived with the model.
7.3 Static system validation 169<br />
Nyquist frequency, which is determined by the actuator spacing. Higher order deformations,<br />
e.g. caused by the shrinkage of the glue that connects the actuator struts to the reflective<br />
facesheet, or initial waviness in the polished facesheet, cannot be corrected.<br />
The flat surface shape corresponds to the first Zernike mode: piston. With the influence<br />
function superposition also higher order Zernike modes are fit. The shape errors can be<br />
minimized for both the Peak To Valley (PTV) <strong>and</strong> RMS norms, corresponding to theℓ1 <strong>and</strong><br />
ℓ2 norms respectively.<br />
Let a shape measurement be denoted by the vector ˆzf as defined in (7.9) <strong>and</strong> subject to<br />
white measurement noise. Let the comm<strong>and</strong> vector ˆ Vℓ for which the difference between<br />
the measured facesheet shape ˆzf <strong>and</strong> the desired shape zd is minimized w.r.t. an arbitrary<br />
normℓ:<br />
ˆVℓ = arg min ˆzf −zd<br />
V<br />
ℓ<br />
PWM<br />
<br />
<br />
= arg min PBfVPWM +Pzf,(0) +Pn−zd .<br />
ℓ<br />
V PWM<br />
In practice the PWM voltage is limited to Vmax = 3.3V , leading to the constrained optimization<br />
problem:<br />
<br />
<br />
ˆVℓ = arg min PBfVPWM +Pzf,(0) +Pn−zd <br />
V ℓ<br />
PWM<br />
subject to −Vmax ≤ VPWM ≤ Vmax.<br />
The effects of the measurement noise n can be reduced by taking ˆz f,(0) as the average of<br />
several measurements. Since the actual influence matrixBf is not known, the productPBf<br />
shall be replaced by the piston removed estimate ˆ Bf .<br />
For the ℓ1 norm (i.e. minimization of the PTV value), the optimization problem becomes a<br />
linear programming problem:<br />
ˆVℓ1 = arg min γ subject to<br />
VPWM ,γ<br />
−γ ≤ ( ˆ BfVPWM +Pz f,(0) +Pn−zd) ≤ γ <strong>and</strong> −Vmax ≤ VPWM ≤ Vmax.<br />
For the ℓ2 norm, the optimization problem becomes a quadratic programming problem:<br />
<br />
ˆVℓ2 = arg min <br />
VPWM ˆ <br />
<br />
BfVPWM +Pzf,(0) +Pn−zd <br />
ℓ2<br />
subject to −Vmax ≤ VPWM ≤ Vmax.<br />
However, due to e.g measurement noisen<strong>and</strong> nonlinear behavior of the DM prototypes, the<br />
DM model in (7.9) with PBf replaced by ˆ Bf will not be correct. The estimated comm<strong>and</strong><br />
vectors ˆ Vℓ1 <strong>and</strong> ˆ Vℓ2 will thus neither minimize PBfVPWM + Pz f,(0) − zdℓ for the ℓ1<br />
norm, nor for the ℓ2 norm. Therefore, an iterative process with iteration indexmis used to<br />
derive the vector ˆ V (m)<br />
ℓ :<br />
ˆV (m+1)<br />
ℓ<br />
= ˆ V (m)<br />
ℓ<br />
+ arg min<br />
∆V PWM<br />
<br />
<br />
ˆ Bf∆VPWM +ˆz (m)<br />
f −zd<br />
<br />
<br />
,<br />
ℓ<br />
7
7<br />
170 7 System modeling <strong>and</strong> characterization<br />
Figure 7.15: The first 28 Zernike modes made with the DM from figure 7.3. The inset shows the<br />
RMS fitting errors w.r.t. the desired shape for the model <strong>and</strong> the measurements.
7.3 Static system validation 171<br />
Figure 7.16: The assembled dummy <strong>mirror</strong><br />
placed vertically in front of the Wyko 400 interferometer.<br />
Flattening force [mN]<br />
1.5<br />
1<br />
0.5<br />
0<br />
−0.5<br />
−1<br />
−1.5<br />
50 100 150 200 250 300 350 400<br />
Actuator number [−]<br />
Figure 7.17: Forces required to obtain the flattened<br />
shape of the 427 struts glue experiment in<br />
figure 7.18. Actuators are sorted on value for insight<br />
into the statistical spread.<br />
where the minimization is subject to−Vmax − ˆ V (m)<br />
ℓ ≤ ∆VPWM ≤ Vmax − ˆ V (m)<br />
ℓ .<br />
Each new setpoint yields the measurement:<br />
ˆz (m)<br />
f = PBf ˆ V (m)<br />
ℓ +Pz f,(0) +Pn (m) .<br />
Here again the effect of measurement noise can be reduced by taking ˆz (m)<br />
f as the average<br />
of several measurements.<br />
The above procedure has been applied for the 61 actuator DM prototype, using an<br />
∅44mm circular aperture area that corresponds the inscribed circle of the hexagonal actuator<br />
grid. The results for the first 28 Zernike modes with an RMS amplitude of 400nm<br />
are shown in figure 7.15. Since the DM has only a limited number of regularly spaced actuators,<br />
there will be a fitting error. The inset in figure 7.15 shows the RMS errors w.r.t.<br />
a perfect Zernike mode together with the numerically evaluated fitting error of the derived<br />
static model.<br />
Since the latter is not subject to measurement noise, nonlinearities or initial unflatness,<br />
this forms the error contribution only due the limited number of actuators. From the piston,<br />
tip <strong>and</strong> tilt mode it becomes clear that the higher order unflatness is ≈25nm RMS. This<br />
unflattness can not be compensated by the (limited) number of actuators. For the higher<br />
order Zernike modes this unflattness can result in better or in worse fitting.<br />
Flattening of the∅150mm dummy <strong>mirror</strong><br />
Since the 427 actuator DM prototype was broken during the last assembly step, it was not<br />
possible to measure its initial unflatness, static or dynamic behavior. Also the effect of<br />
combining multiple actuator modules into a DM system with a single continuous facesheet<br />
could not be analyzed. However, a dummy <strong>mirror</strong> (figure 7.16) was assembled from an<br />
identical facesheet <strong>and</strong> 427 struts glued using the same procedure as used for the broken<br />
7
7<br />
172 7 System modeling <strong>and</strong> characterization<br />
Figure 7.18: Measured shape in µm of the 427 struts glue experiment (left). Shape in in µm after<br />
correction using hypothetical actuator influence functions (right). The unsupported edge<br />
area of the facesheet is not shown as this provides no insight into the achievable flatness.<br />
DM prototype. The shape of this <strong>mirror</strong>, measured with the Wyko 400 interferometer is<br />
depicted in figure 7.18 on the left <strong>and</strong> shows a PTV unflatness of approximately 4.5µm.<br />
Since the results of the above influence function measurements for the 61 actuator DM<br />
indicate that the influence functions are well predicted by the model, it is inferred that the<br />
same is true for the 427 actuator DM. With the influence matrix Bf evaluated with the<br />
parameters from tables 5.1, 6.2 <strong>and</strong> 7.1, the dummy <strong>mirror</strong> is flattened fictitiously. The<br />
flattening procedure, described in the previous section, is here not applicable, since the<br />
facesheet is here supported by 427 struts <strong>and</strong> a rigid reference plate instead of soft actuators.<br />
The shape of interest is the initial unflatness of the facesheet, when it would have been glued<br />
to soft actuators instead of to a stiff base. The soft actuators will deflect due to stresses in the<br />
facesheet, whereas the stiff base does not. Therefore, the flattening must thus be performed<br />
after ’replacing’ the stiff base with soft actuators.<br />
This is analyzed as follows. Let the vector ˆzf,b contain the deflections measured at the<br />
strut positions <strong>and</strong>ˆzf,a the deflections after placing the facesheet on the soft actuators. The<br />
force Fa that the actuators must exert on the facesheet to keep the shape ˆzf,b is obtained<br />
using (7.3) as Fa = Kmˆzf,a. This force is generated by deflection of the actuators from<br />
their measured ˆzf,b positions via the effective actuator stiffness as Fa = Ca(ˆzf,b − ˆzf,a).<br />
Combination of both equations yields:<br />
ˆzf,a = (Km +Ca) −1<br />
<br />
T<br />
Caˆzf,b. (7.10)<br />
Observe that T is the influence matrix that links force to deflection <strong>and</strong> Ca is a diagonal<br />
matrix with the actuator stiffnesses. Since application of the first part to a vector has a lowpass<br />
filter effect <strong>and</strong> the latter part is only a linear scaling, the transformation from ˆzf,b to
7.3 Static system validation 173<br />
ˆzf,a is a smoothing operation. The strength of this smoothing effect depends on the widths<br />
of the influence functions <strong>and</strong> thus on the facesheet <strong>and</strong> actuator stiffnesses.<br />
At this point the fictitious DM with initial unflatnessˆzf,a is flattened w.r.t. an RMS criterion<br />
by applying a set of suitable fictitious actuator forcesFf a. The forcesFf a that minimize the<br />
unflatness w.r.t. the weighting function ˆzf,a − TFa2 F are found using a completion of<br />
squares argument as:<br />
F f a = (TT T) −1 T T ˆzf,a = Caˆzf,b,<br />
where the second step uses the definition ofTin (7.10). The resulting forcesF f a are plotted<br />
in figure 7.17. Note that these forces are independent of the influence functions <strong>and</strong> just<br />
overcome the actuator spring forces corresponding to the deflectionˆzf,b. Figure 7.18 shows<br />
the hypothetical unflatness that remains when fictitious, soft actuators exert the force F f a .<br />
For this dummy <strong>mirror</strong> the residual RMS unflatness is approximately 20nm, requiring an<br />
RMS actuator force of 0.4mN. With the average values of the actuator properties Ka, Ra,<br />
Rl as measured in the previous chapter, the latter corresponds to an RMS actuator voltage<br />
of 0.17V which is approximately 5% of the available maximum voltage.<br />
7.3.3 Power dissipation<br />
Based on the measured influence functions <strong>and</strong> flattening comm<strong>and</strong>s, the expected average<br />
power dissipation to correct atmospheric wavefront distortions, is estimated. For realistic<br />
results, a trade-off is made between power dissipation <strong>and</strong> performance in terms of fitting<br />
error. The RMS power Pa, dissipated by the actuators, can be expressed in terms of the<br />
actuator voltage setpoint <strong>and</strong> the total resistanceRa+R ′ a +Rl+R ′ l estimated in sections<br />
6.7.2 <strong>and</strong> 6.7.3 as:<br />
Pa = (Ra +R ′ a +Rl +R ′ l )−1 V 2 PWM (t) , (7.11)<br />
where (·) 2 denotes the element-wise square <strong>and</strong> t the time. The actuator voltages consists<br />
of the static voltagesVf required for the initial flattening of the DM <strong>and</strong> the dynamic voltages<br />
Vd(t) required to correct atmospheric wavefront disturbances. The value V 2 PWM (t)<br />
in (7.11) can be expressed accordingly as (Vd(t)+Vf) 2 . This reduces to V 2 d (t) +V 2 f<br />
when the atmospheric wavefront distortion <strong>and</strong> thusVd is a zero-mean signal that is therefore<br />
not correlated with the constant signal Vf . Application of these simplifications to<br />
(7.11) yields:<br />
Pa = (Ra +R ′ a +Rl +R ′ l )−1V 2 d (t) +V 2 f . (7.12)<br />
An estimate for the vector Vf with flattening voltages is obtained from the measurements<br />
described above. To quantify V2 d (t) , consider the atmospheric wavefront disturbance<br />
to have a Von Karmann spatial spectrum (section 2.1.2) with covariance matrix<br />
Cφ = φ(t)φ T <br />
(t) . Here the vectorφ(t) denotes the wavefront distortion over a fine grid<br />
over the telescope aperture in radians. The matrix Cφ is numerically approximated using<br />
the approach described in [95] with the modification that the Kolmogorov structure function<br />
has been replaced by the Von Karmann structure function [98] corresponding to the power<br />
spectrum given in (2.4) on page 28:<br />
D vk<br />
5 <br />
3<br />
1/3 2 <br />
7/3<br />
r r r r<br />
φ (r) = 6.88 1−1.485 +5.383 −6.381 .<br />
r0 L0 L0 L0<br />
7
7<br />
174 7 System modeling <strong>and</strong> characterization<br />
The resulting covariance matrix is an approximation, since the continuous spatial integrals<br />
are replaced by numerical sums over a discrete grid of points within the telescope aperture.<br />
The incoming wavefront is corrected by reflection on the DM. Let the DM shape zf(t) be<br />
expressed by (7.6), based on the estimated influence matrix ˆ Bf . Although inertial forces<br />
are neglected, the static model accurately describes the DM facesheet deflections, since its<br />
first resonance frequency lies far above the <strong>control</strong> b<strong>and</strong>width. A fine grid is used to model<br />
the wavefront distortions <strong>and</strong> the DM influence functions. Hereby a realistic estimate of<br />
the fitting error <strong>and</strong> power dissipation is made. When considering the open-loop <strong>control</strong>led<br />
case, let the actuator voltages be chosen as the minimizing argument of a quadratic cost<br />
function that weights both fitting error <strong>and</strong> <strong>control</strong> effort. Let the fitting error, in meters, be<br />
expressed as:<br />
efit(t) = λ<br />
2π φ(t)− ˆ BfVPWM(t)<br />
where λ is the wavelength of the incoming light. The optimal actuator comm<strong>and</strong> vector<br />
Vd(t) is chosen as:<br />
efit(t) Vd(t) = arg min<br />
VPWM (t)<br />
T efit(t) 2 <br />
+γ<br />
F<br />
V T d (t)2 <br />
F<br />
= λ<br />
2π ˆ B +<br />
f φ(t), (7.13)<br />
where ˆ B +<br />
f = (ˆ B T f ˆ Bf +γ 2 I) ˆ B T f is a regularized pseudo inverse of ˆ Bf <strong>and</strong>γ is a weighting<br />
factor for the <strong>control</strong> effort. For this comm<strong>and</strong> vector the fitting errorefit(t) becomes:<br />
f )<br />
<br />
ˆB<br />
<br />
−<br />
f<br />
φ(t) (7.14)<br />
efit = (I− ˆ Bf ˆ B +<br />
<strong>and</strong> the comm<strong>and</strong> signal covariance matrixCVd can be expressed as:<br />
CVd = Vd(t)V T d(t) = <br />
λ 2 ˆB +<br />
2π f Cφ<br />
T ˆB +<br />
f ,<br />
where the second step follows after substitution of (7.13). The diagonal elements of the<br />
covariance matrix CVd now form the vector V2 <br />
d . Similarly, using (7.14) the fitting error<br />
covariance matrixCfit can be expressed as:<br />
Cfit = efit(t)e T fit (t) = <br />
λ 2 ˆB −<br />
2π f Cφ<br />
from which the RMS fitting errorσfit,d can be derived as:<br />
σfit,d =<br />
<br />
Tr(Cfit)<br />
,<br />
nf<br />
T ˆB −<br />
f ,<br />
where nf is the number of grid points used. To provide insight in the power dissipation,<br />
based on measurements on a 61 actuator prototype, several parameters must be scaled.<br />
Firstly, in chapter 2 the number of actuators for an 8 meter telescope is 5000, which implies<br />
that for the same actuator density, the 61 actuator DM prototype should be used on a 0.9m<br />
telescope. Secondly, a continuous facesheet type DM can only prescribe the slope of the
7.3 Static system validation 175<br />
Power [W]<br />
10 −2<br />
10 −3<br />
10 −4<br />
10 −5<br />
10 −8<br />
10 −6<br />
10 −7<br />
RMS fitting error [nm]<br />
Illuminated area<br />
Edge<br />
Uncorrected error<br />
10 −6<br />
Figure 7.19: Relation between the fitting error<br />
σfit,d <strong>and</strong> the average power dissipation per actuator<br />
based on the measured influence matrix of<br />
the 61 actuator DM prototype. Power dissipation<br />
is differentiated between actuators in- <strong>and</strong> outside<br />
the ∅32mm illuminated area (λ=550nm).<br />
Power [W]<br />
10 −2<br />
10 −3<br />
10 −4<br />
10 −5<br />
10 −8<br />
10 −6<br />
10 −7<br />
RMS fitting error [m]<br />
Illuminated area<br />
Edge<br />
Uncorrected error<br />
10 −6<br />
Figure 7.20: Relation between the fitting error<br />
σfit,d <strong>and</strong> the average power dissipation per actuator<br />
based on the modeled influence matrix of<br />
a 427 actuator DM. Power dissipation is differentiated<br />
between actuators in- <strong>and</strong> outside the<br />
∅102mm illuminated area (λ=550nm).<br />
Table 7.2: The expected average power dissipation in mW per actuator based on the influence function<br />
measurements for a 0.9m telescope using a ∅32mm illuminated area on the DM <strong>and</strong> a<br />
Von Karmann spectrum with a Fried parameter r0=0.16m (λ=550nm) <strong>and</strong> an outer scale<br />
L0=100m.<br />
Atmospheric Flattening Total<br />
turbulence<br />
Edge 1.5 23.8 25.3<br />
Illuminated area 1.4 5.5 6.9<br />
Full area average 1.4 14.5 15.9<br />
facesheet at the aperture edge if there is at least one ring of actuators outside the illuminated<br />
aperture. Therefore, an illuminated diameter of 32mm will be considered over which<br />
the Von Karmann spectrum is to be corrected. Actuators outside this area are only used<br />
for flattening <strong>and</strong> for prescribing the boundary conditions. By varying the <strong>control</strong> effort<br />
weighting factor γ <strong>and</strong> assuming a wavelengthλ = 550nm, evaluation of (7.12) yields the<br />
relation between fitting error <strong>and</strong> power dissipation shown in figure 7.19. When a fitting<br />
error of 35nm is taken, which is 2.5% above the best achievable value, the corresponding<br />
average power dissipation of the actuators is listed in table 7.2. A distinction is made<br />
between actuators in the illuminated area <strong>and</strong> those at the edges. Also the two causes for the<br />
dissipation – i.e. the correction of wavefront errors due to atmospheric turbulence <strong>and</strong> the<br />
correction of the initial DM unflatness – are stated separately. The dissipation for actuators<br />
in the illuminated area required for atmospheric wavefront correction is approximately<br />
1.4mW. Although this is less than the 3.2mW estimated in section 6.8, it depends on the<br />
chosen regularization factor γ. Further, the dissipation required for flattening is larger than<br />
7
7<br />
176 7 System modeling <strong>and</strong> characterization<br />
Table 7.3: The expected average power dissipation in mW per actuator based on the influence function<br />
measurements for a 2.3m telescope using a ∅102mm illuminated area on the DM <strong>and</strong> a<br />
Von Karmann spectrum with a Fried parameter r0=0.16m (λ=550nm) <strong>and</strong> an outer scale<br />
L0=100m.<br />
Atmospheric Flattening Total<br />
turbulence<br />
Edge 0.2 0.4 0.6<br />
Illuminated area 0.6 0.2 0.8<br />
Full area average 0.4 0.2 0.6<br />
the functional dissipation, due to significant initial unflatness of the DM, especially at the<br />
edge of the DM.<br />
For the ∅150mm dummy DM a similar table is derived based on an illuminated area<br />
of ∅102mm. This corresponds to the largest diameter, which is fully filled with actuators,<br />
as can be observed in figure 7.23. With the same actuator density as the 5000 actuator DM<br />
proposed for an 8m class telescope, this DM is suitable for a 2.3m wavefront. The forces as<br />
derived in section 7.3.2 are assumed to be indicative of the forces required to compensate<br />
the initial unflatness of the ∅150mm facesheet. Figure 7.20 shows the trade-off between<br />
performance <strong>and</strong> effort for this DM <strong>and</strong> table 7.3 shows the estimated power dissipation for<br />
the same fitting error σfit,d of 35nm as used for the 61 actuator DM. The table shows that<br />
both the power dissipated to correct the initial unflatness as well as the power dissipated to<br />
correct the atmospheric wavefront distortions with a Von Karman spectrum is smaller than<br />
for the 61 actuator DM. However, this only holds true for the regularization factor chosen.<br />
7.4 Dynamic system validation<br />
In section 7.3 the static behavior of the DM system was validated. In this section the dynamic<br />
behavior will be added to the DM facesheet model <strong>and</strong> combined with the actuator<br />
models. The resonance frequencies, mode shapes, modal damping <strong>and</strong> transfer functions<br />
from this model will then be compared to a black-box model, identified from measurement<br />
data, using the PO-Multivariable Output-Error State-sPace (MOESP) subspace identification<br />
algorithm.<br />
7.4.1 Dynamic modeling<br />
There is no known analytic solution available for the biharmonic plate equation for a circular<br />
plate including inertia <strong>and</strong> viscous damping terms subjected to multiple point forces. The<br />
dynamic behavior can be modeled using a Finite Element Model (FEM) approach, but here<br />
it will be done by combining the derived models for the actuators <strong>and</strong> the facesheet <strong>and</strong><br />
extending this with lumped masses <strong>and</strong> dampers.<br />
Let the force equilibrium in (7.2) be extended accordingly to:<br />
Fa −Fρ −Caza −Ba˙za −Maf¨za = 0, (7.15)
7.4 Dynamic system validation 177<br />
where each (i,i) element of the diagonal matrices Ba <strong>and</strong> Maf are the viscous actuator<br />
dampingba <strong>and</strong> the sum of the lumped facesheet mass <strong>and</strong> the moving actuator mass at coordinateρi<br />
respectively. Note that for simplicity it is assumed that the actuator <strong>and</strong> lumped<br />
mass/damper locations coincide, but this can be generalized by using an arbitrary grid <strong>and</strong><br />
appropriately attributing mass, stiffness <strong>and</strong> damping values. The mass distribution chosen<br />
leads to adequate approximations of the mode shapes with spatial frequencies significantly<br />
below the Nyquist frequency of the actuator grid, which corresponds to the lower eigenfrequencies.<br />
These are the most relevant for the achievable correction quality, since they pose<br />
the tightest limit on the achievable <strong>control</strong> b<strong>and</strong>width. Moreover, the stiffness matrix used<br />
corresponds to the solution of the biharmonic plate equation under the assumption of pure<br />
bending. For high spatial frequencies, shear forces become dominant <strong>and</strong> this assumption<br />
loses its validity.<br />
When the result in (7.3) for Fρ is substituted into (7.15) <strong>and</strong> transformed to the Laplace<br />
domain this yields the dynamic system:<br />
Caf +Bas+Mafs 2 za = Fa, (7.16)<br />
where Caf = Km + Ca. The undamped mechanical eigenfrequencies fe,(i) <strong>and</strong> mode<br />
shapesx (i) fori = 1...Na can be obtained by solving the generalized eigenvalue problem:<br />
<br />
Caf −λ (i)Maf x(i) = 0,<br />
wheref e,(i) = λ (i)/2/π.<br />
This procedure has been performed for two different cases. Firstly for a DM with regularly<br />
placed actuators with 6mm pitch in a hexagonal pattern. Equal lumped masses were added<br />
at all actuator grid points <strong>and</strong> all actuator <strong>and</strong> facesheet properties were used as given in<br />
tables 5.1, 6.2 <strong>and</strong> 7.1. The first 100 resonance frequencies for this model are plotted as<br />
circles in figure 7.21 <strong>and</strong> the first twelve resonance modes are shown in figure 7.22. From<br />
figure 7.21 it is clear that the first resonance modes occur in a small frequency b<strong>and</strong> <strong>and</strong><br />
from figure 7.22 it is clear that the modal shapes correspond very well to the Zernike polynomials<br />
[136]. Traditionally, these polynomials are used to describe both the aberrations in<br />
the optical domain as well as the dynamic modes of the wavefront corrector in the mechanical<br />
domain.<br />
However, the resonance frequencies <strong>and</strong> corresponding modal shapes are influenced by the<br />
edge conditions of the reflective facesheet. When considering the 427 actuator DM prototype<br />
that consists of seven hexagonal actuator modules, observe that the gaps between<br />
the hexagons at the outside have no actuators that support the <strong>mirror</strong> facesheet. In the dynamic<br />
model in (7.16) leads to ’zero’ elements on the diagonals of the matrices Ca <strong>and</strong><br />
Ba <strong>and</strong> to lower values of the corresponding diagonal elements in Maf due to the absence<br />
of moving actuator masses. The lack of support stiffness in the edge areas leads to lower<br />
resonance frequencies with local mode shapes. To properly attribute lumped mass fractions<br />
of the facesheet to grid points in edge areas, its mass is distributed based on the Voronoi<br />
diagram of the grid points (figure 7.23). The Voronoi diagram is the dual of the Delaunay<br />
grid triangulation [130, 159] that is frequently used to draw a surface defined at arbitrary<br />
grid of points. The Voronoi diagram creates a polygon area around each grid point in which<br />
all points are closest to that particular grid point. For most grid points this yields a closed<br />
polygon, but for edge points this is open towards the edge. The polygon area determines the<br />
7
7<br />
178 7 System modeling <strong>and</strong> characterization<br />
Frequency [Hz]<br />
1800<br />
1600<br />
1400<br />
1200<br />
1000<br />
800<br />
600<br />
400<br />
200<br />
427 actuators in seven grids of 61<br />
559 actuators in a full hexagonal array<br />
20 40 60 80 100<br />
Mode number [−]<br />
Figure 7.21: The lowest 100 undamped mechanical<br />
resonance frequencies.<br />
Figure 7.22: The lowest 12 undamped mechanical<br />
resonance modes corresponding to the frequencies<br />
plotted as circles in figure 7.21.<br />
lumped fraction of the facesheet mass <strong>and</strong> can for closed polygons be computed using Matlab’sÔÓÐÝÖfunction<br />
[130]. The open edge polygons are first extended with two points<br />
on the circular facesheet edge by which the polygon is closed. The total area is then the sum<br />
of the area of the artificially closed polygon <strong>and</strong> the area between these two points <strong>and</strong> the<br />
circle, which follows from the distance between the two additional edge points. Since all<br />
areas can be calculated analytically, no approximations are made <strong>and</strong> the summed area for<br />
all grid points exactly equals πr2 f . For the 427 actuator DM prototype the result is plotted<br />
in figure 7.23.<br />
The lowest 100 eigenfrequencies for this case are plotted in figure 7.21 as the solid dots.<br />
The first resonance frequencies are lower than in the homogenously supported case <strong>and</strong><br />
the corresponding modal shapes are local bending modes of the unsupported edge areas.<br />
Clearly, the actuator layout of future DMs should be chosen such that the facesheet edges<br />
are uniformly supported.<br />
Figure 7.23: Voronoi diagram for the hexagonally<br />
arranged grid points marked with a small<br />
dot. The actuator grid of the 427 actuator DM<br />
prototype is marked with ”o”.
7.4 Dynamic system validation 179<br />
Magnitude (abs)<br />
Phase (deg)<br />
10 −5<br />
10 −10<br />
180<br />
0<br />
−180<br />
−360<br />
−540<br />
10 1<br />
Central actuator<br />
To first neighbor<br />
To second neighbor<br />
To third neighbor<br />
10 2<br />
Frequency [Hz]<br />
Figure 7.24: Bode plots of the modeled transfer<br />
functions between the PWM voltage of the central<br />
actuator <strong>and</strong> the position of itself <strong>and</strong> three<br />
neighbors.<br />
10 3<br />
Displacement [µm]<br />
2.5<br />
2<br />
1.5<br />
1<br />
0.5<br />
0<br />
Central actuator<br />
To first neighbor<br />
To second neighbor<br />
To third neighbor<br />
−0.5<br />
0 10 20 30 40 50<br />
Time [ms]<br />
Figure 7.25: Step response of the modeled transfer<br />
functions between the PWM voltage of the<br />
central actuator <strong>and</strong> the position of itself <strong>and</strong><br />
three neighbors.<br />
Although the mechanics largely determine the behavior of the DM in terms of resonance<br />
frequencies, the electronics influence the damping, response time, gain, etc. Therefore, the<br />
state-space description for the single actuator behavior in (6.4) on page 136 is extended to<br />
describe the behavior of multiple actuators <strong>and</strong> combined with the mechanical equations of<br />
motion in (7.15). The scalar statesIRl ,Ia,VCl ,za <strong>and</strong> ˙za become the vectorsIRl ,Ia,VCl ,<br />
za <strong>and</strong> ˙za whose ith elements contain the corresponding states of the ith actuator. Similarly,<br />
the scalar parametersKa, Ra, Rl, La, Ll <strong>and</strong>Cl for each actuatoribecome the(i,i)<br />
diagonal elements of the square diagonal matricesKa,Ra,Rl,La,Ll <strong>and</strong>Cl respectively.<br />
Further, instead of the mechanical equation of motion for the single, uncoupled actuator in<br />
(5.24), the dynamic equation in (7.16) will be used, leading to:<br />
⎡ ⎤<br />
˙Ia<br />
⎢ ˙zf ⎥<br />
⎢ ⎥<br />
⎢¨zf<br />
⎥<br />
⎢ ⎥<br />
⎣ ˙Va<br />
⎦<br />
˙IRl<br />
=<br />
⎡<br />
−L<br />
⎢<br />
⎣<br />
−1<br />
a Ra 0 −L−1 a Ka L−1 a 0<br />
0 0 I 0 0<br />
M −1 −1 −1<br />
afKa−M afCaf−MafBa 0 0<br />
−C −1<br />
l 0 0 0 C −1<br />
l<br />
0 0 0 −L −1<br />
l −L−1<br />
l Rl<br />
⎤⎡<br />
⎤<br />
Ia<br />
⎥⎢<br />
zf ⎥<br />
⎥⎢<br />
⎥<br />
⎥⎢<br />
˙zf ⎥<br />
⎥⎢<br />
⎥<br />
⎦⎣Va⎦<br />
IRl<br />
<br />
+<br />
⎡<br />
0<br />
⎢ 0<br />
⎢ 0<br />
⎣ 0<br />
L −1<br />
⎤<br />
⎥<br />
⎦<br />
l<br />
<br />
Afm<br />
Bfm<br />
VPWM<br />
(7.17)<br />
The transfer matrixH(s) between the PWM voltageVPWM <strong>and</strong> the facesheet deflectionzf<br />
can be expressed accordingly as:<br />
H(s) = 0 I 0 0 0 (sI−Afm) −1 B T fm.<br />
The DC-gain of this transfer matrix is equal to (7.6) <strong>and</strong> can be derived by evaluatingH(s)<br />
fors = 0:<br />
H(0) = − 0 I 0 0 0 A −1<br />
fmBTfm = C−1<br />
afKt(Ra +Rl) −1 = Bρ, (7.18)<br />
7
7<br />
180 7 System modeling <strong>and</strong> characterization<br />
whereAfm <strong>and</strong>Bfm are defined in (7.17) <strong>and</strong> the final expression follows from the definition<br />
ofBρ in (7.6).<br />
The model in (7.17) has been generated numerically for a∅150mm reflective facesheet that<br />
is regularly supported by 6mm spaced actuators over its entire area. The actuator, electronics<br />
<strong>and</strong> facesheet properties used are given in tables 5.1, 6.2 <strong>and</strong> 7.1. In figure 7.24 Bode<br />
plots are shown of the entries of the transfer matrix corresponding to the PWM voltage of the<br />
central actuator <strong>and</strong> the position of itself <strong>and</strong> three neighbors at 6, 12 <strong>and</strong> 18mm distance. It<br />
shows that the static DC response to neighboring actuator positions indeed decays rapidly<br />
with the spatial distance, but that the global shapes of the lightly damped, lowest dynamic<br />
modes (figure 7.22) lead to a strong coupling between the actuators at high frequencies. To<br />
illustrate this low damping, figure 7.25 shows the step response of the same actuators due<br />
to a step input at the PWM voltage of the central actuator. In practice the damping will be<br />
higher due to the presence of air above the facesheet, intrinsic damping of the facesheet <strong>and</strong><br />
strut materials, deformation of the glue between the struts <strong>and</strong> the facesheet, etc. To quantify<br />
this effect, the relative damping of the resonant modes of this model will be compared<br />
to the damping derived from a modal analysis in section 7.4.3. The consequences of these<br />
observations for <strong>control</strong> performance will be discussed in section 7.5.<br />
7.4.2 System identification<br />
Modal (or structural) analysis is frequently performed using the Eigensystem Realization<br />
Algorithms (ERAs) [15, 60] <strong>and</strong> its variants [116]. The ERAs are a subspace based identification<br />
methods that estimate a state space by taking the Singular Value Decomposition<br />
(SVD) of a Hankel matrix of the system’s impulse response function. This impulse response<br />
function is either measured directly from impulse excitation or estimated from more generic<br />
input-output data. For open-loop measurement data of Multi-Input Multi-Output (MIMO)<br />
systems the MOESP algorithm [183] <strong>and</strong> its variants [186] are very suitable. For closedloop<br />
identifications the Predictor Based Subspace IDentification (PBSID) identification algorithm<br />
[29, 31] can be used, which has been applied for the identification of a DM with<br />
60 actuators <strong>and</strong> 104 sensors of the MAD (Multi-conjugate <strong>Adaptive</strong>-optics Demonstrator)<br />
system in [30]. However, the DM can here be identified in open-loop, making the<br />
added complexity of the PBSID algorithm superfluous. Besides subspace based identification<br />
algorithms that use state-space parameterizations, other algorithms can be used for<br />
MIMO system identifications with other parameterizations. For instance, in [170] a MIMO<br />
Transfer Function (TF) parametrization is shown to be very efficient in both the number of<br />
parameters <strong>and</strong> the required computational effort, when compared to subspace algorithms.<br />
The identification algorithm used for modal analysis of the DM prototype is chosen based<br />
on several requirements. It must be able to deal with 61 simultaneous inputs <strong>and</strong> at least<br />
as many outputs. Since the sampling frequency of the setup used is limited to 10kHz, the<br />
<strong>dynamics</strong> of the electronics that become dominant above 5kHz will not be well observable<br />
from the measurement data. Although the DM facesheet <strong>dynamics</strong> are of infinite order,<br />
the low frequent resonance modes can be adequately described using a limited number of<br />
lumped masses. This means that the system order required for identification is at least twice<br />
the number of lumped masses, which in this case is in the order of hundreds. It also means<br />
that the model parametrization used must allow the large number of poles <strong>and</strong> zeros to be<br />
independent to properly describe the numerous resonance modes of the DM facesheet. The
7.4 Dynamic system validation 181<br />
poles will be used after the identification step to compute the resonance frequencies, their<br />
relative damping <strong>and</strong> the modal shapes.<br />
Since for identification of a state-space model a high state dimension <strong>and</strong> large numbers of<br />
in- <strong>and</strong> outputs are required, the identification algorithm used must be efficient w.r.t. both<br />
memory <strong>and</strong> computation steps. Moreover, the method must be suitable for the available<br />
measurement setup, which is the same as used previously to identify the behavior of single<br />
actuators <strong>and</strong> depicted in figure 6.14. In this setup, the deflection of the DM facesheet can<br />
be measured only at a single point at a time with a Polytec laservibrometer. The obtained<br />
measurement data is thus expected to be significantly corrupted both by measurement <strong>and</strong><br />
process noise, since the measurements are performed in a noisy environment without vibration<br />
isolation facilities. The identification algorithm must be robust for these types of noise.<br />
On the other h<strong>and</strong>, since all quantization is performed in the <strong>control</strong> PC in an open-loop<br />
setting there is no quantization noise.<br />
Unbiasedness to noise can be achieved using instrumental variable techniques [119, 170,<br />
186] that e.g. exploit the fact that the excitation signal is uncorrelated with measurement<br />
or process noise but correlated with the future system output. Instrumental variables can be<br />
used in both the MOESP algorithm – e.g. PI-MOESP uses past inputs <strong>and</strong> PO-MOESP uses<br />
also past outputs as instrumental variables – <strong>and</strong> the TF method in [170], which uses future<br />
inputs as instrumental variables.<br />
The method in [170] assumes a model in a canonical MIMO TF form of which it is argued<br />
by the authors that – in general – it has a smaller number of unknown coefficients than<br />
state-space parameterizations. The parametrization is based on a denominator polynomial<br />
with scalar coefficientsak <strong>and</strong> a numerator polynomial with matrix coefficientsBk:<br />
n<br />
aky(t−k) =<br />
k=0<br />
n<br />
Bku(t−k)+<br />
k=1<br />
n<br />
akv(t−k),<br />
wherea0 = 1 <strong>and</strong>v(t) ∈ N(0,Cv) is white measurement noise. However, when the number<br />
of in- <strong>and</strong> outputs as well as the degree of any irreducible matrix fraction description<br />
of the system become large, the number of parameters to be identified for this parametrization<br />
becomes larger than that of a corresponding state-space parametrization. Since this is<br />
the case for the system to be identified, the PO-MOESP algorithm will be used for system<br />
identification.<br />
Identification using the MOESP algorithm<br />
The MOESP algorithm is a subspace identification algorithm that uses a QR decomposition<br />
[77] of the input-output data matrix to compress the data <strong>and</strong> thus improve the computational<br />
efficiency. The algorithm <strong>and</strong> its variants are found in literature [184–186] <strong>and</strong> Matlab implementations<br />
are readily available. Since the identification is subject to both measurement<br />
<strong>and</strong> process noise, the PO-MOESP algorithm will be used, which uses past inputs <strong>and</strong> outputs<br />
as instrumental variables to provide unbiased estimates w.r.t. measurement <strong>and</strong> process<br />
noise. Since in literature this variant of the algorithm is generally intended when referring<br />
to MOESP, the prefix PO- will here also be neglected. For the MOESP algorithm, the data<br />
k=0<br />
7
7<br />
182 7 System modeling <strong>and</strong> characterization<br />
generating system is described in innovation form as:<br />
<br />
x(t+1) = Ax(t)+BVPWM(t)+Ke(t)<br />
ˆzf(t) = Cx(t)+DVPWM(t)+e(t)<br />
where e(t) ∈ N(0,Ce) expresses the effect of both process <strong>and</strong> measurement noise. The<br />
Kalman gainKis not of interest here <strong>and</strong> will not be included in the estimation procedure.<br />
Furthermore, the system has at least one sample delay, hence the direct feed-through term<br />
D is assumed to be zero.<br />
The MOESP algorithm estimates a minimal realization of a system by determining the<br />
column space of an estimate of the system’s extended observability matrix. To be able to<br />
determine this column space, the rank of this matrix must be equal to the system order n,<br />
which for an arbitrary observable Linear Time-Invariant (LTI) system requires it to have at<br />
least n block-rows. Consequently, the number s of block-rows of the data Hankel matrix<br />
used in the MOESP algorithm should also be larger or equal to the chosen system order n.<br />
Due to the large number of inputs <strong>and</strong> outputs together with a large number of recorded<br />
samples, for the DM system prototypes this would lead to huge memory requirements<br />
beyond the specifications of the available computers. To relax this requirement, note that<br />
for systems with multiple outputsl the rank of the observability matrix may become equal<br />
to the system order n for less than n block-rows with a minimum set by l ·s > n. This is<br />
not unlikely for the system at h<strong>and</strong>, since the number of outputs is large compared to the<br />
system order when considering lumped masses to be located at actuator positions only. For<br />
the DM system the minimum possible number s of block rows to be used for the MOESP<br />
algorithm is then around 5: the number of block-states in (7.17). However, since in reality<br />
the reflective facesheet forms an infinite order system that was modeled using a finite<br />
number of lumped masses, choosing a suitable numbers of block-rows is not trivial.<br />
After identification, the quality of the obtained system is evaluated by applying the<br />
estimated system realization to a validation data sequence <strong>and</strong> computing the VAF, defined<br />
as: ⎡ <br />
⎤<br />
VAF =<br />
⎢<br />
⎣1−<br />
(ˆzf(t)−˜zf(t)) T (ˆzf(t)−˜zm(t))<br />
<br />
ˆz T f (t)ˆzf(t)<br />
N<br />
t0<br />
N<br />
t0<br />
⎥<br />
⎦·100%,<br />
where ˜zf(t) is obtained by simulating the identified model for the known excitation signal<br />
VPWM(t) <strong>and</strong> ˆzf(t) is the vector of facesheet deflections measured by the laser vibrometer.<br />
The VAF represents the fraction of the signal variance that is accounted for by the model<br />
<strong>and</strong> should be close to 100%.<br />
7.4.3 Modal analysis<br />
The modal analysis of the identified system begins with a modal decomposition of its Amatrix.<br />
This decomposition is such thatΛÂ = MÂ, where the diagonal matrixΛcontains<br />
the (complex) eigenvalues of <strong>and</strong> the columns ofMcontain the (complex) eigenvectors.<br />
The matrix M forms a state transform matrix that diagonalizes the system <strong>and</strong> yields a
7.4 Dynamic system validation 183<br />
state-space description with statex(t) whose state transition matrix is equal toΛ:<br />
<br />
˜x(t+1) = Λ˜x(t)+M −1 Bu(t),<br />
zf(t) = CM˜x(t),<br />
where the influence of process <strong>and</strong> measurement noise is neglected. For the discrete time<br />
system, each eigenvalue λ (i) on the diagonal of Λ is related to the resonance frequency<br />
ω n,(i) <strong>and</strong> relative dampingζ (i) of modei as [62]:<br />
λ (i) = e Ts<br />
<br />
−ζ (i)ωn,(i)±jωn,(i) 1−ζ2 <br />
(i) = |λ (i)|·∠λ (i),<br />
whereTs is the sampling frequency <strong>and</strong><br />
|λ| = e −Tsζ (i)ω n,(i), ∠λ = e ±Tsjω n,(i)<br />
<br />
1−ζ 2<br />
(i) .<br />
Inversely, the resonance frequenciesω n,(i) <strong>and</strong> damping ratio’sζ (i) can be computed as:<br />
ω n,(i) = ln∠λ−ln|λ|<br />
Ts<br />
ln|λ|<br />
<strong>and</strong> ζ (i) = − . (7.19)<br />
ln|λ|−ln∠λ<br />
Further, the vectors of the matrix CM form the facesheet shapes corresponding to the system’s<br />
eigenfrequencies. Since complex poles occur in conjugate pairs (a,b) with the same<br />
modal frequency ωn,a = ωn,b, the matrix’s complex part <strong>and</strong> columns corresponding to<br />
complex conjugate eigenvalues will for the modal analysis be ignored.<br />
It should be noted that the estimated system matrices are corrupted by process noise – e.g.<br />
due to external vibrations – <strong>and</strong> measurement noise in the laser vibrometer <strong>and</strong> the data acquisition<br />
card. Moreover, they are subject to aliasing effects <strong>and</strong> an ill-chosen system order.<br />
This leads to errors in the derived eigenvalues <strong>and</strong> eigenvectors, some of which may be nonstructural<br />
extraneous [60]. Various criteria have been developed to separate the structural<br />
<strong>and</strong> extraneous modes, e.g. using the concept of modal amplitude coherence (MAC) [60] or<br />
modal dispersion analysis (DA) [57]. Besides these criteria that select modes considering<br />
the estimated system matrices, statistical techniques (e.g. bootstrap or Monte-Carlo) can<br />
be used to estimate confidence intervals for the desired modal parameters [102]. However,<br />
for the MIMO system at h<strong>and</strong> the computations involved to derive reliable statistics are<br />
extremely time consuming. Confidence in the identified system behavior will therefore be<br />
derived in a qualitative manner from the application of the PO-MOESP algorithm with variations<br />
in the number of block rows s <strong>and</strong> the obtained VAF values <strong>and</strong> consistency of the<br />
identified system modes. Also the DC-gain of the identified realizations should match the<br />
influence matrix previously identified in section 7.3.2 at the laser-vibrometer measurement<br />
locations.<br />
Results<br />
For the 61 actuator DM prototype, the facesheet response is measured on the 79 points<br />
shown in figure 7.27. These are the 61 actuator locations <strong>and</strong> 18 points on the facesheet<br />
along its edge. The signal acquisition <strong>and</strong> DM setpoint update rate is chosen as high as<br />
possible to minimize the effects of sampling <strong>and</strong> aliasing. The 10kHz sampling frequency<br />
7
7<br />
184 7 System modeling <strong>and</strong> characterization<br />
Table 7.4: VAF values obtained for the PO-MOESP algorithm on the laser-vibrometer measurement<br />
data of the 61 actuator DM prototype for various values of the number of block-rows s.<br />
s [-] 8 9 10 11 12 13 14 15 16 17<br />
n [-] 105 108 110 112 109 108 108 110 113 113<br />
VAF [%] 93.3 93.7 93.9 94.0 94.1 94.1 94.2 94.4 94.4 94.4<br />
used, is close to the upper limit of the serial communication chain but still significantly lower<br />
than the PWM actuator voltage base frequency. The laser vibrometer is pointed at each<br />
grid location for 10 seconds, producing 100.000 measurements. A zero-mean, b<strong>and</strong>limited,<br />
white noise sequence VPWM(t) ∈ N(0,σ 2 e I) is generated with σe = 0.13V <strong>and</strong> t =<br />
0...10s <strong>and</strong> applied on each location. Except for small variations in initial conditions <strong>and</strong><br />
timing (i.e. jitter), the obtained data is equal to the data that would have been obtained when<br />
the response of all points is measured simultaneously.<br />
Before applying the MOESP system identification algorithm, the suitability of using a small<br />
number of block rows is tested by evaluating the model in (7.17) for properties of the system<br />
to be identified. The properties of all 61 actuators are taken equal to the averages of the<br />
identified values (table 6.4) <strong>and</strong> lumped masses are assumed to be located at all points in<br />
the measurement grid (figure 7.27). According to (7.17) this yields a state-space system of<br />
ordern = 2·79+3·61= 341, which consists of twice the number of lumped masses (2·79)<br />
<strong>and</strong> three times the number of actuators (3·61). The system is subsequently discretized to a<br />
10kHz sampling frequency assuming a Zero Order Hold (ZOH) input using Matlab’s<br />
function. The rank of the observability matrix of the resulting system reaches the system<br />
order n for six block-rows with a condition number of 10 8 , which indicates the suitability<br />
of small numberssof block-rows for the MOESP algorithm.<br />
However, the available efficient implementations of the MOESP algorithm do not allow to<br />
chooses smaller than the system ordernsuch that a customized implementation is required.<br />
To reduce memory requirements, batch-wise computation of the economy size R-factor of<br />
the data Hankel matrix is used together with the efficient algorithm described in [185] for<br />
estimating the matrixBfrom the already computed R-factor <strong>and</strong> corresponding SVD.<br />
The obtained measurement data set is split into an identification set of 85.000 samples <strong>and</strong><br />
a validation set of 15.000. The MOESP system identification method is applied to the first<br />
part of this input-output data <strong>and</strong> the VAF value is computed after simulation on the second<br />
part. The number of block rows s is varied between 8 <strong>and</strong> 17 <strong>and</strong> the model order n is<br />
chosen from the singular values of the data Hankel matrices as the break-point of the initial<br />
negative slope visible in figure 7.26, which shows the singular values for two instances of<br />
s. The singular values are normalized to the largest one to clarify that the breakpoint of<br />
the initial slope hardly changes with the number s of block-rows used. The VAF-values<br />
obtained for the PO-MOESP algorithm with various values of s are listed in table 7.4 <strong>and</strong><br />
are generally around 95%. Despite the presence of process <strong>and</strong> measurement noise, the<br />
derived models are consistent with the measurement data.<br />
Figures 7.28 <strong>and</strong> 7.29 show the first 12 resonance frequencies <strong>and</strong> the corresponding<br />
modal shapes derived from the analytic model in (7.17) <strong>and</strong> the black-box model identified<br />
with the MOESP algorithm. For the modal analysis of the analytic model, the average<br />
measured actuator properties listed in table 6.4 are used. When all actuators have equal
7.4 Dynamic system validation 185<br />
properties, the modal shapes show a high degree of symmetry (figure 7.28). In practice the<br />
actuator properties vary, leading to the asymmetric mode shapes in figure 7.29. The lowest<br />
resonance mode of the system lies at ∼725Hz <strong>and</strong> corresponds to a motion of the lowerleft<br />
edge area of the facesheet. This frequency is lower than expected. Since the edges are<br />
supported by a few actuators, the low resonance may be caused by a lower stiffness ca of<br />
Singular value [−]<br />
10 0<br />
10 −1<br />
10 −2<br />
10 −3<br />
s = 8<br />
s = 17<br />
50 100 150 200 250 300 350 400<br />
Singular value index [−]<br />
Figure 7.26: Singular values of the data Hankel<br />
matrix in the PO-MOESP algorithm normalized<br />
to the largest one. The drop around the 100 th<br />
singular value forms an indication of the system<br />
order.<br />
Figure 7.28: The first 12 modal shapes derived<br />
from the analytic model in (7.17) using the average<br />
actuator properties listed in table 6.4 <strong>and</strong><br />
the facesheet properties listed in table 7.1.<br />
66<br />
19<br />
67<br />
18<br />
20<br />
14<br />
65<br />
15<br />
64<br />
10<br />
68 17<br />
16 12<br />
13<br />
8<br />
9<br />
4<br />
5 63<br />
36<br />
69<br />
46<br />
11<br />
6<br />
7<br />
2<br />
3<br />
26<br />
31<br />
62<br />
37<br />
45 1 21 32<br />
70 51 44 41 27 38 79<br />
50 43 22 33<br />
56<br />
71<br />
55<br />
49<br />
48<br />
42<br />
23<br />
28<br />
34<br />
39<br />
78<br />
61<br />
72 60<br />
54<br />
53<br />
47<br />
24<br />
29 40<br />
35 77<br />
59<br />
73<br />
58<br />
74<br />
52<br />
57<br />
30<br />
76<br />
25<br />
75<br />
Figure 7.27: The grid of points at which the<br />
response of the 61 actuator DM prototype was<br />
measured. Points 1...61 correspond to the actuator<br />
locations.<br />
Figure 7.29: The first 12 modal shapes derived<br />
with the MOESP system identification method for<br />
s = 17.<br />
7
7<br />
186 7 System modeling <strong>and</strong> characterization<br />
only a few actuators. The relative damping ζ has been computed for the identified system<br />
modes using (7.19) <strong>and</strong> plotted against the eigenfrequency for the models identified using<br />
MOESP for various values ofsin figure 7.30. Since for most eigenvalues of the system the<br />
corresponding eigenfrequencies, relative damping <strong>and</strong> modal shapes are independent of the<br />
value of s, these estimates are considered to be reliable. However, note from the figure that<br />
this is not the case for all eigenvalues <strong>and</strong> these results have therefore not been included in<br />
figure 7.29.<br />
Figure 7.30 also shows the relative damping for the analytical model based on the average<br />
actuator properties listed in table 6.4. The relative damping for this model is significantly<br />
lower than for the identified black-box models, which means that the damping observed<br />
in the DM system is not entirely due to the actuators. This suggests the presence of other<br />
dissipative processes such as intrinsic material damping in the facesheet material, glued<br />
connections or damping due to the movement of air above the facesheet. The latter is a very<br />
likely explanation, since the facesheet vibrations due to the noise excitations used were<br />
clearly audible.<br />
A Bode plot of the identified modelfor s = 17 is shown in figure 7.31 for the transfer<br />
functions from the actuator voltage at actuator 2 to the displacement at its first, second <strong>and</strong><br />
third neighboring actuators 41, 42 <strong>and</strong> 47. This shows the high peaks in the magnitude<br />
response as a result of the low relative damping. However, where the magnitude of the<br />
peaks does not decrease with frequency in the Bode plot of the analytical model in figure<br />
7.24, they do in the model as identified, which suggests the presence of additional dissipative<br />
processes at high frequencies.<br />
Further, the influence functions have been derived from the model identified by computing<br />
the DC gain matrix of the system according to (7.18) as Ĉ(I − Â)−1ˆ B whose columns<br />
contain the influence functions. The influence function of actuator 2 is shown in figure 7.32<br />
together with the influence functions of the same actuator derived from the analytic model<br />
Relative damping [−]<br />
10 0<br />
10 −1<br />
10 −2<br />
MOESP (s=8)<br />
MOESP (s=12)<br />
MOESP (s=17)<br />
Analytical model<br />
10<br />
500 1000 1500<br />
−3<br />
Frequency [Hz]<br />
Figure 7.30: Relative damping ζ corresponding<br />
to the eigenfrequencies of the models identified<br />
with the MOESP algorithm for various values of<br />
the number of block-rows s <strong>and</strong> of the analytical<br />
model from (7.17).<br />
Magnitude [m/V]<br />
Phase [deg]<br />
10 −6<br />
10 −8<br />
0<br />
−180<br />
−360<br />
10 1<br />
Poked actuator<br />
To first neighbor<br />
To second neighbor<br />
To third neighbor<br />
10 2<br />
Frequency [Hz]<br />
Figure 7.31: Bode plot of the model identified<br />
with the MOESP algorithm for s = 17 between<br />
the comm<strong>and</strong> voltage at actuator 2 <strong>and</strong> its first,<br />
second <strong>and</strong> third neighboring actuators 41, 42<br />
<strong>and</strong> 47 respectively.<br />
10 3
7.4 Dynamic system validation 187<br />
<strong>and</strong> the Wyko measurements of section 7.3.2. The shape <strong>and</strong> magnitude match qualitatively,<br />
but some quantitative error can be observed.<br />
This is partly due to the poor alignment accuracy of the laser vibrometer spot.<br />
Finally, the step response functions derived from the model from the central actuator voltage<br />
setpoint to the displacement of four points on the reflective surface are shown in figure 7.33.<br />
The four points are the location of the actuator itself <strong>and</strong> that of three neighbors with 1, 2<br />
<strong>and</strong> 3 actuator spacings distance. In figure 7.34, the corresponding step response derived<br />
from the analytical model is shown for the same actuators. A comparison with figure 7.33<br />
confirms that in the analytical model of the 61 actuator DM damping is underestimated. A<br />
comparison with figure 7.25 on page 179 is also of interest as the settling time of the 427<br />
actuator DM model is significantly shorter than that of the 61 actuator DM. Although this<br />
suggests that the scale of the system has an influence, this difference is caused by differences<br />
in the system properties used to generate the figures. For instance, for figure 7.25 a damping<br />
constant ba = 0.4mNs/m <strong>and</strong> motor constant ka = 0.19N/A was used, which are higher<br />
than the ba = 0.30mNs/m <strong>and</strong>Ka = 0.11N/A used for figure 7.34. Nevertheless, to better<br />
underst<strong>and</strong> the effect of the number of actuators on the DM system damping, consider the<br />
lowest resonance mode – which has a global modal shape – to form a mass-spring-damper<br />
system. For such a system the relative damping coefficient ζ is related to the damping<br />
coefficientb, massm<strong>and</strong> resonance frequencyωn as2ζωn = b/m, hence:<br />
ζ = b<br />
.<br />
2mωn<br />
As discussed in [174], the first facesheet resonance frequency (ωn in the above equation)<br />
is scale independent because each increase in facesheet mass is matched by a proportional<br />
increase in stiffness. On the other h<strong>and</strong>, this means that the modal massmis proportional to<br />
the number of actuators Na. When considering only viscous actuator damping, the modal<br />
dampingb is also proportional to Na such that ζ should not depend on the number of actuators.<br />
Figure 7.32: The influence function of actuator 2 (figure 7.27) as derived from the analytical model<br />
(left), from the Wyko measurements of section 7.3.2 (middle) <strong>and</strong> from the model identified<br />
using MOESP (right).<br />
7
7<br />
188 7 System modeling <strong>and</strong> characterization<br />
Displacement [µm]<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
Central actuator<br />
First neighbor<br />
Second neighbor<br />
Third neighbor<br />
−0.2<br />
0 10 20 30 40 50<br />
Time [ms]<br />
Figure 7.33: Step response of the central actuator<br />
in the model identified with the MOESP algorithm<br />
for s = 17. The response is shown of four<br />
locations: the central actuator <strong>and</strong> three actuators<br />
at 1, 2 <strong>and</strong> 3 actuator spacings distance.<br />
7.5 Discrete time <strong>control</strong><br />
Displacement [µm]<br />
2<br />
1.5<br />
1<br />
0.5<br />
0<br />
Central actuator<br />
First neighbor<br />
Second neighbor<br />
Third neighbor<br />
−0.5<br />
0 10 20 30 40 50<br />
Time [ms]<br />
Figure 7.34: Step response of the central actuator<br />
of the analytical model in (7.17), using the<br />
average actuator parameters listed in table 6.4.<br />
The response is shown of four locations: the central<br />
actuator <strong>and</strong> three actuators at 1, 2 <strong>and</strong> 3<br />
actuator spacings distance.<br />
To investigate the relevance of the high frequent dynamic behavior of the DM for <strong>control</strong>ler<br />
design, the discrete time behavior of the system is of interest. The discrete time black-box<br />
model identified in the previous section can be used for this purpose, but this is a discrete<br />
time model with a 10kHz sampling frequency. Although the data acquisition devices used<br />
to obtain the measurements did not include anti-aliasing filters, the aliasing effects caused<br />
by sampling will have affected the behavior measured below the 5kHz Nyquist frequency.<br />
However, the ZOH filter on the system input has a limited b<strong>and</strong>width <strong>and</strong> the damping<br />
of DM system resonances increases for high-frequent resonances (figure 7.31), suggesting<br />
that aliasing effects are likely limited. Under this assumption, the identified model can be<br />
resampled to other sampling frequencies without significant error.<br />
It has been assumed throughout this thesis that when the DM system is used in closed-loop<br />
<strong>control</strong> for the purpose of <strong>Adaptive</strong> Optics (AO), the sensor used for feedback is a CCDbased<br />
WaveFront Sensor (WFS). As outlined in the introductory chapter, such a sensor<br />
integrates photons over time <strong>and</strong> thus introduces additional temporal <strong>dynamics</strong>. In section<br />
6.6 these <strong>dynamics</strong> were modeled in the continuous time domain <strong>and</strong> it was shown how<br />
they affect the discrete time system behavior. Discrete time system models for arbitrary<br />
sampling frequencies are here derived by first transforming the identified model back to<br />
continuous time under the assumption of a ZOH input signal. Subsequently, thez-transform<br />
in (6.7) on page 140 is used to obtain a model for the system behavior for a desired sampling<br />
frequency, where the CCD integration time is assumed to be equal to the sampling time<br />
<strong>and</strong> communication delays are neglected. Note that the z-transform procedure described<br />
in section 6.6 can also be used to obtain the corresponding discrete time behavior for the<br />
analytical, continuous time model in (7.17). However, here only the resampling procedure<br />
for the identified model will be considered. The impulse responses of so obtained models for
7.5 Discrete time <strong>control</strong> 189<br />
sampling frequencies of 500Hz <strong>and</strong> 1, 2 <strong>and</strong> 10kHz are plotted in figure 7.35 for the poked<br />
actuator 2 <strong>and</strong> its first, second <strong>and</strong> third neighboring actuators 41, 42 <strong>and</strong> 47. The figure<br />
shows that the settling behavior significantly improves for a reduced sampling frequency.<br />
In fact, for the design value of the sampling frequency of 1kHz the figure suggests that<br />
the DM system may be well modeled using a low order Finite Impulse Response (FIR)<br />
description. Such a model structure has been previously described in AO literature for<br />
different correctors <strong>and</strong> successfully used for <strong>control</strong>ler synthesis [100]. In fact, most AO<br />
literature considers the DM to be a static gain without any temporal <strong>dynamics</strong> other than<br />
Response [µm]<br />
Response [µm]<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
−0.2<br />
−0.4<br />
−0.6<br />
0.2<br />
0<br />
−0.2<br />
−0.4<br />
−0.6<br />
−0.8<br />
−1<br />
f s = 10000Hz<br />
−0.8<br />
0 5 10 15 20 25<br />
Time [ms]<br />
f s = 1000Hz<br />
−1.2<br />
0 5 10 15 20 25<br />
Time [ms]<br />
Response [µm]<br />
Response [µm]<br />
0.4<br />
0.2<br />
0<br />
−0.2<br />
−0.4<br />
−0.6<br />
−0.8<br />
−1<br />
−1.2<br />
0 5 10 15 20 25<br />
Time [ms]<br />
0.2<br />
0<br />
−0.2<br />
−0.4<br />
−0.6<br />
−0.8<br />
−1<br />
f s = 2000Hz<br />
f s = 500Hz<br />
Poked actuator<br />
To first neighbor<br />
To second neighbor<br />
To third neighbor<br />
−1.2<br />
0 5 10 15 20 25<br />
Time [ms]<br />
Figure 7.35: The impulse response function of the MOESP identified model for s = 17 resampled<br />
to different sampling frequencies including CCD-type temporal behavior. The poked<br />
actuator is actuator 2 <strong>and</strong> the response is also plotted for actuators 41, 42 <strong>and</strong> 47.<br />
7
7<br />
190 7 System modeling <strong>and</strong> characterization<br />
delays [191].<br />
Now consider the traditional <strong>control</strong>ler structure for AO as described in section 3.2.2. When<br />
disregarding the spatial <strong>dynamics</strong> introduced by the WFS, this consists of a discrete time<br />
integrator with tunable parameters α <strong>and</strong> β <strong>and</strong> the inverse of the static system response<br />
matrixH(0) as in (7.18):<br />
C(z) = α(I−βz −1 I) −1 H −1 (0).<br />
Since the <strong>control</strong>ler contains the inverse of the static response H(0), it diagonalizes the<br />
loop gain for low frequencies. Up to the first resonance mode, the magnitude response of<br />
the discretized plant is highly constant such that cross-coupling can be neglected. For small<br />
values of β, the parameter α is thus equal to the b<strong>and</strong>width defined as the 0dB crossing<br />
of the loop gain. Using this <strong>control</strong>ler for the identified model resampled to 1kHz <strong>and</strong><br />
augmented with the temporal <strong>dynamics</strong> introduced by a CCD-based WFS, a b<strong>and</strong>width<br />
of approximately 60Hz can be achieved. This is significantly less than the design goal of<br />
200Hz (section 2.2.3) <strong>and</strong> is mainly due to the low damping of the system. The b<strong>and</strong>width<br />
can be increased as suggested in chapter 6 by adding a proportional term to reduce the<br />
phase lag around the b<strong>and</strong>width or by using an optimal <strong>control</strong>ler synthesis approach in<br />
which dynamic models for both the DM system <strong>and</strong> the atmospheric disturbances to be<br />
suppressed are considered.<br />
Unfortunately, due to lack of time, no results of the closed loop performance of the designed<br />
AO system on either a breadboard setup or an actual telescope can be presented here, leaving<br />
the final system verification for future research.<br />
7.5.1 A note on distributed <strong>control</strong><br />
Although throughout this thesis the feasibility of a <strong>modular</strong>ly distributed <strong>control</strong> system was<br />
investigated, the realized 61 actuator DM prototype has too few actuators to demonstrate the<br />
feasibility <strong>and</strong> benefit of such a <strong>control</strong>ler structure.<br />
7.6 Conclusions<br />
The assembly of two DMs prototypes is shown: a ∅50mm DM with 61 actuators <strong>and</strong> a<br />
∅150mm DM with 427 actuators. In the first prototype a single actuator grid is used,<br />
whereas for the second prototype <strong>modular</strong>ity is shown by the assembly of seven identical<br />
grids on a common base. The seven actuator grids, with accompanying dedicated driver<br />
boards, are attached to a single, continuous, facesheet.<br />
The actuator model derived in the chapter 6 is extended with a linear model of the continuous<br />
facesheet, based on an analytic solution of the biharmonic plate equation <strong>and</strong> point<br />
forces. Lumped masses are added to obtain the dynamic behavior. Both static <strong>and</strong> dynamic<br />
performance is validated on the∅50mm DM using measurements. Scaled Hadamard matrices<br />
for the actuator voltage comm<strong>and</strong> vectors are used to measure the 61 influence functions<br />
in front of a Wyko interferometer. This approach minimizes the variance estimation errors<br />
of the influence matrix due to measurement noise.<br />
The measured actuator coupling for the central actuator is 52%, which is close to the value
7.6 Conclusions 191<br />
obtained from the static DM system model. Nevertheless, variation in accuracy was observed<br />
between actuators, which is attributed mainly to unknown variation in the mechanical<br />
stiffness, motor constant <strong>and</strong> electrical resistance of the actuators.<br />
The measured shape <strong>and</strong> amplitude of the influence functions agree with the prediction with<br />
the static model (figures 7.13 <strong>and</strong> 7.14). This includes the increased static gain (m/V) of<br />
the actuators at the edge of the DM due to the reduced facesheet stiffness from the <strong>mirror</strong>’s<br />
free-edge boundary condition. The variation in the static gain is observed <strong>and</strong> attributed to<br />
variation in motor constant, actuator stiffness <strong>and</strong> electrical resistance <strong>and</strong> driver circuits.<br />
The influence matrix derived from the measurements is used to shape the <strong>mirror</strong> facesheet<br />
into the first 28 Zernike modes, which includes the piston term that represents the best flat<br />
<strong>mirror</strong>. The interferometrically measured shapes are compared to the perfect Zernike modes<br />
<strong>and</strong> to the Zernike modes as made with the limited number of regularly spaced actuators in<br />
the actuator grid. The total RMS error is ≈25nm for all modes, whereas the inevitable fitting<br />
error varies between 0 <strong>and</strong> 23nm depending on the mode.<br />
The power dissipation in each actuator of the ∅50mm <strong>mirror</strong> to correct the Von Karman<br />
turbulence spectrum (D/r0 = 5.4, L0 = 100m) is estimated. Actuators outside the illuminated<br />
area are distinguished from those inside this area. Furthermore, the estimated power<br />
dissipation is split into turbulence correction <strong>and</strong> <strong>mirror</strong> flattening. For the turbulence correction,<br />
1.5mW for the outer <strong>and</strong> 1.4mW for the inner actuators, is dissipated. For static<br />
flattening these values are 23.8mW <strong>and</strong> 5.5mW respectively.<br />
The 427 actuator DM prototype was broken during placement of its protective cover, therefor<br />
it was not possible to measure its unflatness, static or dynamic behavior. A dummy<br />
<strong>mirror</strong> (figure 7.16) with similar facesheet specifications, which was used to test the assembly<br />
procedure of the 427 connection struts, was interferometrically measured instead. This<br />
<strong>mirror</strong> showed≈4.5µm PTV unflatness. This <strong>mirror</strong> was fictitiously flattened, with the use<br />
of the modeled influence matrix <strong>and</strong> requires 0.17V RMS actuator voltage, which corresponds<br />
to 5% of the available output voltage range <strong>and</strong> a power dissipation of 0.2mW per<br />
actuator.<br />
The predicted dynamic behavior of the DM is validated by measurements. A laser vibrometer<br />
is used to measure the displacement of the <strong>mirror</strong> facesheet, while the actuators are<br />
driven by zero-mean, b<strong>and</strong>limited, white noise voltage sequence. Using the MOESP system<br />
identification algorithm, high-order black-box models are identified with VAF values<br />
around 95%. This identified model is compared with the derived analytical model. The<br />
latter uses the average actuator properties as measured in chapter 6. The first resonance frequency<br />
identified is 725Hz, <strong>and</strong> lower than the 974Hz expected from the analytical model.<br />
This is attributed to the variations in actuator properties, such as actuator stiffness.<br />
The relative modal damping of the model identified is an order of magnitude higher than the<br />
damping in the analytical model, where only actuator damping is considered. The difference<br />
is attributed to the presence air damping <strong>and</strong> damping in the glue used for the connection<br />
struts. To evaluate the behavior of the realized DM system for discrete-time closed-loop<br />
<strong>control</strong>, the identified black-box model is resampled to the foreseen AO system sampling<br />
frequency of 1kHz, while considering the temporal <strong>dynamics</strong> of a CCD-based WFS. When<br />
neglecting spatial <strong>dynamics</strong> of the WFS <strong>and</strong> considering a traditional integrator-type <strong>control</strong><br />
law a b<strong>and</strong>width of approximately 60Hz can be achieved. This is lower than the design<br />
goal of 200Hz <strong>and</strong> is the result of the low damping of the system’s resonances. To achieve<br />
a higher b<strong>and</strong>width, either a more complex <strong>control</strong>ler structure should be considered or a<br />
7
7<br />
192 7 System modeling <strong>and</strong> characterization<br />
more suitable sampling frequency should be chosen for which the system <strong>dynamics</strong> become<br />
the most desirable. Nevertheless, for future DM designs damping of the system resonances<br />
should be considered an explicit design requirement.
ÔØÖØ<br />
ÓÒÐÙ×ÓÒ×ÒÖÓÑÑÒØÓÒ×<br />
In this chapter the conclusions over all chapters will be summarized <strong>and</strong> recommendations<br />
for future research will be given.<br />
193
8<br />
194 8 Conclusions <strong>and</strong> recommendations<br />
8.1 Conclusions<br />
Over the last half a century <strong>Adaptive</strong> Optics (AO) has provided a solution for the wavefront<br />
distortions introduced by the earth’s atmosphere that limit the resolution of ground based<br />
optical telescopes. To increase this resolution, the size of telescopes must increase together<br />
with the correction quality of their AO systems. This thesis presents the design, realization<br />
<strong>and</strong> testing of a new Deformable Mirror (DM) system concept. The main design drivers of<br />
this concept are low cost, low power dissipation, low hysteresis <strong>and</strong> drift <strong>and</strong> high linearity.<br />
The developed system consists of the DM including the required driver electronics <strong>and</strong><br />
<strong>control</strong> system, but excludes the WaveFront Sensor (WFS) that is assumed to be of the<br />
Shack-Hartmann type. This thesis considers the specific application of the DM system for<br />
an 8m class telescope <strong>and</strong> discusses extendibility of its design towards larger telescopes.<br />
Detailed design requirements for the DM system are derived based on general assumptions<br />
on the properties of atmospheric wavefront disturbances modeled by Kolmogorov<br />
statistics <strong>and</strong> the Taylor hypothesis (frozen flow) for representative conditions (r0 = 0.16m<br />
for λ = 550nm <strong>and</strong> fG = 25Hz). The derivation is driven by a desired optical quality in<br />
terms of a Strehl ratio of 0.85, which is linked to a certain residual wavefront variance. This<br />
residual error is budgeted over the main sources of error that are identified as the fitting error<br />
of the DM due to the limited number of actuators <strong>and</strong> the temporal error due to the limited<br />
<strong>control</strong> b<strong>and</strong>width. A trade-off based on foreseen challenges on both fields of competence<br />
leads to a <strong>control</strong> b<strong>and</strong>width of 200Hz with 5000 actuators having at least 5.6µm stroke <strong>and</strong><br />
0.36µm inter-actuator stroke.<br />
The foreseen design consists of a single reflective, <strong>deformable</strong> facesheet supported via struts<br />
by hexagonally arranged actuators with a pitch of 6mm on a∅500mm DM. The hexagonal<br />
actuator layout is chosen since this gives the highest actuator areal density. The actuators<br />
have a low mechanical stiffness that is only high enough to provide a first mechanical<br />
resonance frequency of approximately 1kHz. This choice not only reduces the required<br />
complexity of a <strong>control</strong> system <strong>and</strong> keeps power dissipation to a minimum, but also limits<br />
the detrimental effect of a failed actuator on the DM shape. The actuators are chosen of<br />
the magnetic reluctance type because of their high efficiency, low driving voltage <strong>and</strong> low<br />
moving mass. Moveover, this type of actuators shows little hysteresis or drift. They are built<br />
as modules of 61 on a stiff base-plate that also serves as a carrier for magnetic flux. These<br />
modules are designed in layers that extend over many actuators to reduce the number of<br />
parts <strong>and</strong> the complexity of assembly <strong>and</strong> improve the uniformity of the actuator properties.<br />
Multiple actuator modules are supported by a stiff backing structure. Extendibility of the<br />
design concept towards larger telescopes requiring more Degrees Of Freedoms (DOFs) is<br />
achieved through <strong>modular</strong>ity of the design in terms of mechanics <strong>and</strong> electronics.<br />
To allow for <strong>modular</strong>ity of the <strong>control</strong> system hardware, a distributed <strong>control</strong>ler design<br />
is required. A spatially distributed architecture is proposed in which each DM actuator is<br />
matched to a <strong>control</strong>ler node that can only communicate with a small number of nearby<br />
nodes <strong>and</strong> receives only a local subset of the available WFS measurements. To allow this<br />
architecture to be extendible towards DMs with more DOFs, both the amount of information<br />
exchanged per sample with neighboring nodes as well as the size of the local subset of<br />
the WFS measurement may not vary with the number of DOF.<br />
The computational cost of existing <strong>control</strong> methods is investigated together with their suit-
8.1 Conclusions 195<br />
ability for implementation on the <strong>modular</strong>ly distributed hardware architecture. This shows<br />
that even for the most efficient methods available the wavefront reconstruction step required<br />
for Shack-Hartmann type WFSs involves a number of computations that is more than linearly<br />
proportional to the number of DOF of the WFS. This renders the wavefront reconstruction<br />
step to be the most critical part in the design of a distributed <strong>control</strong>ler for which<br />
new algorithms need to be developed. Several concepts for such algorithms are explored,<br />
starting from a st<strong>and</strong>ard Steepest Descent (SD) solver whose computations are shown to<br />
have a distributed structure. Simulations showed that by extending this solver with an Least<br />
Mean Squares (LMS) based wavefront prediction mechanism, the required number of computations<br />
does not increase with the number of DOFs of the WFS. The distributed dynamic<br />
wavefront reconstructor is also generalized towards a network of output-interconnected <strong>control</strong>lers<br />
with an Auto-Regressive Moving Average (ARMA) structure. By assuming this<br />
parametrization it is implicitly assumed that the structure of the disturbance generating system<br />
in innovations form is also a network of output interconnected ARMA systems. Performance<br />
loss is thus expected if this is not the case in practice.<br />
Using a two-stage algorithm the unknown <strong>control</strong>ler parameters are identified. Two methods<br />
are presented to enforce stability of the resulting closed-loop. However, both identification<br />
steps must be solved off-line <strong>and</strong> the first is a centralized operation that involves the measurement<br />
data of all sensors.<br />
Results presented on measurement data obtained from an optical breadboard <strong>and</strong> on synthetic<br />
data show the performance to depend on the chosen communication radius, but even<br />
for very small radii it is found to exceed that of a r<strong>and</strong>om walk baseline strategy. However,<br />
the performance shows a decays w.r.t. the number of DOF of the WFS, which is in contrast<br />
with the SD/LMS based reconstructor <strong>and</strong> is neither the case for the structured Finite<br />
Impulse Response (FIR) approximation obtained as an intermediate result at the first identification<br />
step. Future research is required to analyze such scaling properties in more detail,<br />
preferably using wavefront disturbance measurements from an actual large telescope.<br />
The second part of this thesis discusses the electromagnetic actuators, the electronics<br />
<strong>and</strong> the system modeling <strong>and</strong> validation. The actuators consist of a closed magnetic circuit<br />
in which a Permanent Magnet (PM) provides static magnetic force on a ferromagnetic core<br />
that is suspended in a membrane. This attraction force is influenced by a current through a<br />
coil, which is situated around the PM to provide movement of the core. With the direction<br />
of the current the attractive force of the PM is either increased or decreased, allowing<br />
movement in both directions. The actuators are free from mechanical hysteresis, friction<br />
<strong>and</strong> play <strong>and</strong> therefore have a high positioning resolution with a high reproducibility. The<br />
stiffness of the actuator is determined by both the membrane suspension <strong>and</strong> the magnetic<br />
circuit. There exists a large design freedom for both.<br />
To drive a desired current through the actuator coils the computed comm<strong>and</strong> value of the<br />
<strong>control</strong> system must be communicated to driver electronics. In chapter 6 these two parts of<br />
the required electronics are discussed. Since the inductance-over-resistance time-constant<br />
of the actuator is short (75µs), voltage <strong>control</strong> is chosen over current <strong>control</strong>. The motor<br />
constant, stiffness <strong>and</strong> resistance of the actuator circuit will vary from actuator to actuator<br />
<strong>and</strong> vary with temperature to cause slow gain variations. Current <strong>control</strong> would compensate<br />
variations in the resistance, but still leave variations of the motor constant <strong>and</strong> stiffness for<br />
the AO <strong>control</strong> system. Pulse Width Modulation (PWM) based voltage drivers are chosen<br />
because of their high efficiency <strong>and</strong> suitability to be implemented in large numbers with<br />
8
8<br />
196 8 Conclusions <strong>and</strong> recommendations<br />
only a few electronic components. A drawback of this choice is that suppression of the<br />
PWM ripple requires a second order low-pass filter per actuator. The driver electronics for<br />
61 actuators are located on a single, multi-layer Printed Circuit Board (PCB) <strong>and</strong> consist<br />
of Field Programmable Gate Arrays (FPGAs) to generate the PWM signals, Field Effect<br />
Transistors (FETs) for the H-bridge switches <strong>and</strong> coil/capacitor pairs for the 2 nd order<br />
low-pass filters. A serial communication system is chosen based on the Low Voltage Differential<br />
Signalling (LVDS) st<strong>and</strong>ard for its low power consumption (15mW/transceiver), high<br />
b<strong>and</strong>width (up to 655Mb/s) <strong>and</strong> consequently low latency, low communication overhead<br />
<strong>and</strong> extensive possibilities for customization. A flat-cable connects up to 32 electronics<br />
modules to a custom designed communications bridge, which translates ethernet packages<br />
into LVDS packages <strong>and</strong> vice versa. The ethernet side of the communications bridge is<br />
connected to the <strong>control</strong> computer at a speed of 100Mbit/s <strong>and</strong> uses the User Datagram<br />
Protocol (UDP) protocol to minimize overhead <strong>and</strong> latency.<br />
Two DMs prototypes were successfully assembled: a ∅50mm DM with 61 actuators<br />
<strong>and</strong> a ∅150mm DM with 427 actuators. In the first prototype a single actuator grid is used,<br />
whereas for the second prototype <strong>modular</strong>ity is shown by the assembly of seven identical<br />
grids on a common base. All actuators from the seven grids are attached to a single,<br />
continuous, facesheet.<br />
A nonlinear mathematical model of the actuator is derived that describes both its static <strong>and</strong><br />
dynamic behavior based on equations from the magnetic, mechanic <strong>and</strong> electric domains.<br />
This model is linearized to obtain expressions for general actuator properties such as motor<br />
constant, inductance, stiffness <strong>and</strong> resonance frequency. Frequency response function<br />
measurements are performed on each actuator using a general purpose current source <strong>and</strong><br />
a laser vibrometer, showing all actuators to be functional. From these measurements the<br />
motor constant, actuator stiffness <strong>and</strong> resonance frequency are identified. On average, these<br />
properties deviate slightly from the modeled values, but their statistical spread is small,<br />
stressing the reproducibility of the manufacturing <strong>and</strong> assembly process. Moreover, the<br />
average actuator stiffness <strong>and</strong> resonance frequency of 471N/m <strong>and</strong> 1.83kHz respectively<br />
are close to their design values of 500N/m <strong>and</strong> 1885Hz. The fact that the measured average<br />
motor constant of 0.12N/A is lower than the modeled value of 0.17N/A, can be partly<br />
attributed to leakage fluxes.<br />
The frequency response measurements are repeated using the custom built communication<br />
<strong>and</strong> driver electronics. The expected change in behavior is modeled by extending the<br />
actuator model with a pure delay for the communication electronics <strong>and</strong> a voltage source<br />
with an analog 2 nd order low-pass filter for the driver electronics. Measurement show the<br />
communication latency to be well represented by τc = 89.7·10 −6 +39·10 −6 Nm, where<br />
Nm is the number of the actuator grid. From the frequency response measurements the<br />
actuator properties are again identified, yielding <strong>and</strong> average a stiffness of 473±46N/m,<br />
a motor constant of 0.11±0.02N/A, a damping of 0.30±0.11mNs/m, an inductance of<br />
3.0±0.2mH<strong>and</strong> a resonance frequency of 1.83±91Hz. These properties show some<br />
variation between actuators, but this cannot be attributed to the location of the actuator in<br />
the grid.<br />
The time domain response of an actuator to a 4Hz sine voltage over the full stroke shows<br />
hysteresis to be negligible <strong>and</strong> static nonlinearities in the response of the actuator to remain<br />
below 5% for the intended±10µm stroke. Measurements also showed that in the expected
8.1 Conclusions 197<br />
operating range, the total power dissipation is dominated by indirect losses in the FPGAs<br />
to generate the PWM signals. Solutions are sought in alternative FPGA implementations,<br />
yielding a reduction of 40% in the master FPGA <strong>and</strong> 29% in the slave FPGAs.<br />
The actuator model is extended with a linear model of the continuous facesheet, based<br />
on an analytic solution of the biharmonic plate equation in presence of point forces. The<br />
static performance is validated on the ∅50mm DM using interferometric measurements.<br />
Scaled Hadamard matrices for the actuator voltage comm<strong>and</strong> vectors are used to measure<br />
the 61 influence functions in front of a Wyko interferometer. This approach minimizes<br />
the variance of estimation errors in the influence matrix due to measurement noise. The<br />
measured shape <strong>and</strong> amplitude of the influence functions agrees with the prediction of the<br />
static, linear model. This includes the increased static gain (m/V) of the actuators at the<br />
edge of the DM due to the reduced facesheet stiffness from the <strong>mirror</strong>’s free-edge boundary<br />
condition. The variation in the static gain is observed <strong>and</strong> attributed to variation in motor<br />
constant, actuator stiffness <strong>and</strong> electrical resistance <strong>and</strong> driver circuits. The measured<br />
actuator coupling of the central actuators of 52% is close to the modeled value.<br />
The influence matrix derived from the measurements is used to shape the <strong>mirror</strong> facesheet<br />
into the first 28 Zernike modes, which includes the piston term that represents the best<br />
flat <strong>mirror</strong>. The interferometrically measured shapes are compared to the perfect Zernike<br />
modes <strong>and</strong> to the best-fit Zernike modes created by the static DM model. The total Root<br />
Mean Square (RMS) error is ≈25nm for all modes, whereas the inevitable fitting error<br />
varies between 0 <strong>and</strong> 23nm depending on the mode.<br />
The static DM model is extended with lumped masses for the facesheet to obtain insight<br />
in the dynamic behavior of the DM system. This behavior is validated by measurements<br />
in which a laser vibrometer measures the displacement of each of a number of points on<br />
the <strong>mirror</strong> facesheet, while the actuators are driven by a zero-mean, b<strong>and</strong>limited, white<br />
noise voltage sequence. Using the Multivariable Output-Error State-sPace (MOESP)<br />
system identification algorithm, high-order black-box models are identified with Variance<br />
Accounted For (VAF) values around 95%. The so obtained models are compared with the<br />
analytical model in which the measured average actuator properties as substituted. The first<br />
identified resonance frequency of 725Hz is lower than the 974Hz expected from the model,<br />
which is attributed to the variations in actuator properties, such as actuator stiffness. The<br />
relative modal damping of the identified model is an order of magnitude higher than the<br />
damping in the analytical model, where only the identified actuator damping is considered.<br />
The difference is attributed to the presence of air damping <strong>and</strong> damping in the glue used for<br />
the connection struts.<br />
The power dissipation in each actuator of the∅50mm <strong>mirror</strong> to correct a Von Karmann<br />
turbulence spectrum with D/r0 = 5.4 <strong>and</strong> an outer scale of 100m is estimated. Approximately<br />
1.5mW is dissipated for the purpose of turbulence correction by actuators near the<br />
edge <strong>and</strong> 1.4mW by the inner actuators. These values are 23.8mW <strong>and</strong> 5.5mW respectively<br />
to perform static flattening of the DM.<br />
8
8<br />
198 8 Conclusions <strong>and</strong> recommendations<br />
8.2 Recommendations<br />
In this thesis the feasibility of a <strong>modular</strong>ly distributed <strong>control</strong>ler architecture is investigated.<br />
The main focus lies with the <strong>control</strong>ler synthesis <strong>and</strong> how to implement it, but as the WFS<br />
is excluded from this research project the efficient transmission of the measurements to the<br />
distributed <strong>control</strong>ler nodes has received no attention <strong>and</strong> requires further research.<br />
A distributed adaptive reconstruction <strong>and</strong> prediction algorithm is presented that requires<br />
further research into several of its properties. Firstly the proposed cost function does<br />
not weight the unseen modes of the WFS such that the update law allows them to grow<br />
arbitrarily large. Incorrect values for the unseen ’waffle’ mode of the Fried geometry<br />
will compromise performance <strong>and</strong> incorrect values for the ’piston’ mode will require an<br />
overly large stroke of the corrector. Weighting of these modes in the distributed setting<br />
is not possible without affecting the overall AO performance. A strategy (e.g. through<br />
weighting) must therefore be sought that constrains the unseen modes with the least effect<br />
on performance.<br />
Further, the stability <strong>and</strong> performance of the algorithm has to be determined on actual<br />
measurement data from a large telescope <strong>and</strong> also realistic DM models have to be incorporated<br />
into the algorithm, which currently assumes this to be the identity operator. When<br />
the influence of the DM actuators is local, the latter may be approached by estimating<br />
the coefficients of a second distributed filter by minimization of a quadratic cost function.<br />
Similar to the separation principle in centralized <strong>control</strong> design, a distributed <strong>control</strong>ler is<br />
then obtained that consists of two parts: a reconstructor/predictor <strong>and</strong> a regulator.<br />
Finally, the effect of sampling frequency on the performance trade-off has not been<br />
considered. Sampling frequency largely determines the Signal to Noise Ratio (SNR)<br />
of the WFS measurements. It is conceivable that in a distributed setting the sampling<br />
frequency that gives optimal performance in terms of correction quality is different than in<br />
the centralized setting, which makes it an interesting tuning parameter that deserves future<br />
attention.<br />
An entirely different approach to efficient <strong>control</strong> may be to investigate the relation between<br />
the number of sensors <strong>and</strong> the required number of states for a atmospheric disturbance<br />
model identified from measurement data to maintain a sufficient level of accuracy. It is not<br />
unconceivable that the accuracy of models identified using black box system identification<br />
techniques (e.g. subspace identification) does not (significantly) deteriorate as the number<br />
of sensors in the AO increases while keeping the model order fixed. This still allows the<br />
number of model coefficients <strong>and</strong> the corresponding computational load to increase linearly<br />
with the number of sensors.<br />
To further improve the properties of the actuator considered in chapter 5 it is recommended<br />
to lower the reluctance of the radial air gap. This can be realized by a smaller gap<br />
width or by a larger gap area. A sensitivity analysis of the actuator model showed this to<br />
strongly affect the motor constant <strong>and</strong> actuator stiffness. A factor two reduction of this reluctance<br />
will increase the motor constant to 0.37N/A <strong>and</strong> increase the actuator stiffness to<br />
750N/m. When neglecting the effects of actuator stiffness, the factor four increase of the<br />
motor constant leads to a factor 16 reduction in power dissipation. A higher motor constant<br />
also leads to increased electronic damping of the mechanical resonance frequencies of the
8.2 Recommendations 199<br />
DM system such that they become less limiting for the closed-loop performance without<br />
adding complexity to the actuator or electronics design.<br />
8
200<br />
ÔÔÒ×
ÔÔÒÜ<br />
ËÀÖØÑÒÒ×ÔÓØÔÓ×ØÓÒÒ<br />
The size <strong>and</strong> alignment of the Hartmann array relative to the actuator grid is very<br />
important for the quality of measurements from a Shack Hartmann sensor (SHS) for use in<br />
closed loop <strong>control</strong>. The sensor should be well able to observe the Deformable Mirror (DM)<br />
shape. In fact, as noted by [65, 66], the Hartmann grid affects both the <strong>control</strong>lability as<br />
well as the observability of the system <strong>and</strong> thus its closed loop performance.<br />
For a distributed <strong>control</strong> setting it is important that all local <strong>control</strong>lers have access to<br />
measurements containing relevant information. When optical, spatial aliasing filters are<br />
used, the spatial resolution of the sensor does not need to exceed that of the DM actuators.<br />
This allows the sensor spots to form a regular pattern together with the DM actuator grid<br />
<strong>and</strong> only leaves the position of the spots relative to the actuators to be chosen arbitrarily. It<br />
will here be shown that this alignment makes a big difference for the observability of the<br />
DM shape.<br />
Let the deformationds as measured by a SHS in terms of spatial gradients of the <strong>mirror</strong><br />
surface be expressed in terms of the actuator comm<strong>and</strong> vectoruas:<br />
ds = Bu,<br />
where B is the DM influence matrix <strong>and</strong> temporal <strong>dynamics</strong> are neglected. The conditioning<br />
of this matrix B determines the observability <strong>and</strong> <strong>control</strong>lability of the DM system <strong>and</strong><br />
therefore the achievable suppression of the wavefront disturbance [65]. However, since the<br />
number of slope measurements is often larger than the number of actuators the matrix B<br />
is often rectangular. This yields zero-valued singular values, corresponding to modes in<br />
sensor space that cannot be the result of actuator actions <strong>and</strong> will therefore be neglected.<br />
To determine the best spot location, the condition number of the matrix B has been evaluated<br />
for various locations of the sensor spots relative to the actuator positions. The matrix<br />
was calculated using equation (7.7) on page 164 with the DM parameters from table 7.1<br />
using a hexagonally arranged actuator grid with 547 actuators over a circular aperture. The<br />
comm<strong>and</strong>-to-slope influence function matrix was obtained from four comm<strong>and</strong>-to-phase<br />
matrices based on slightly displaced (1/60 th actuator spacing) instanced of the sensor grid.<br />
The two grids displaced in x direction provided the slopes in x direction, whereas the two<br />
displaced in y direction provided the slopes in y direction. To reduce the effect of DM<br />
edges, the outer three rings of actuators <strong>and</strong> sensor spots were modeled, but not considered<br />
in the computation of the condition number.<br />
The relation between the sensor-to-actuator grid alignment <strong>and</strong> the condition number of B<br />
is plotted in figure A.1. This shows that the best (lowest) condition numbers are obtained<br />
201
202 Appendices<br />
Figure A.1: Condition number of the matrix B<br />
as a function of the spot displacement in X <strong>and</strong> Y<br />
directions w.r.t. the actuator grid (boxes).<br />
Condition number [−]<br />
10 4<br />
10 3<br />
10 2<br />
10 1<br />
500 1000 1500 2000<br />
Actuator stiffness [N/m]<br />
Figure A.2: The condition number of the matrix<br />
B versus the actuator stiffness ca for a spot displacement<br />
in y direction of 2.6mm.<br />
when the sensor spots not not lie on the actuator grid lines. The best location shows a<br />
∼70 times better conditioning than the worst location, when the sensor spots are exactly<br />
aligned with the actuators. In that case, the influence of an actuator is not observed at<br />
all by the corresponding sensor spot, since the spatial gradients remain practically zero.<br />
The influence of an actuator is only observed at the location of its neighbors, where the<br />
gradients are already much smaller.<br />
Besides the grid alignment, the mechanical stiffness of the DM actuators influences<br />
the condition number. This significantly affects the width <strong>and</strong> amplitude of the influence<br />
functions <strong>and</strong> thus the magnitude of the slope measurements to a unit actuator comm<strong>and</strong>.<br />
This relation has been evaluated for a number of different actuator stiffness values <strong>and</strong> is<br />
shown in figure A.2. The condition number first improves for higher actuator stiffnesses,<br />
but a minimum is obtained around 1650N/m after which it increases again.
ÔÔÒÜ<br />
ÇÒÙ×ÒÐÓÐÔ××ÚØÝØÓÒÓÖ<br />
ÐÓÐ×ØÐØÝ<br />
The concept passivity can be used in the context of stability of interconnected systems<br />
[8]. Passivity can be posed as a condition to guarantee stability <strong>and</strong> it can be shown that the<br />
interconnection of two passive systems is again passive [133]. Various definitions of passivity<br />
exist that can be further divided in several sub-classes [110, 133]. A passive system is a<br />
special type of dissipative system. Let passivity here be defined using quadratic Lyapunov<br />
functions on the system state, which is in literature referred to as internal passivity [110].<br />
To show stability <strong>and</strong> interconnection stability of passive systems, consider two discrete<br />
time Single-Input Single-Output (SISO) systems S1(z) <strong>and</strong> S2(z) that are given in state-<br />
space form as:<br />
Si(z) :<br />
<br />
xi(t+1) = Aixi(t)+Biui(t)<br />
yi(t) = Cixi(t).<br />
(B.1)<br />
Let these systems be called strictly passive if there exist Lyapunov functions Vi(xi(t)) =<br />
x T i (t)Pixi(t) <strong>and</strong> matrices Pi ≻ 0 such that for the supply function si(yi(t),ui(t)) =<br />
2yi(t)ui(t) the following holds for any inputui(t) ∈ R at any time instantt:<br />
Vi(xi(t+1))−Vi(xi(t)) < si(yi(t),ui(t)). (B.2)<br />
Asymptotic stability requires that for the autonomous system (ui(t) = 0) <strong>and</strong> all t that<br />
Vi(xi(t+1))−Vi(xi(t)) < 0, which is implied by the above definition of strict passivity.<br />
However, note that this is a sufficient condition for stability – not a necessary condition –<br />
since the Lyapunov functions are here restricted to quadratic functions.<br />
Now consider the negative feedback interconnection of the systemsS1(z) <strong>and</strong>S2(z) (figure<br />
B.1). To show that the negative feedback interconnection is also strictly passive (<strong>and</strong> thus<br />
stable), consider the c<strong>and</strong>idate Lyapunov functionV(x1(t),x2(t)) = V1(x1(t))+V2(x2(t))<br />
<strong>and</strong> supply function s(y1(t),u(t)) = 2y1(t)u1(t). By summation of the passivity conditions<br />
in (B.2) for both systems <strong>and</strong> substitution ofu1(t) = u(t)−y2(t) <strong>and</strong> the definitions<br />
ofV(x1(t),x2(t)) <strong>and</strong>s(y1(t),u(t)), the following is obtained:<br />
V(x1(t+1),x2(t+1))−V(x1(t),x2(t)) < s(y1(t),u(t)).<br />
This means that the interconnected system is again strictly passive. Note that a similar proof<br />
can be given for the fact that a series connection of two passive systems is again passive.<br />
This result can be used to derive conditions for stability of the interconnected network of<br />
203
204 Appendices<br />
Figure B.1: Feedback interconnection<br />
of the local <strong>control</strong>lers<br />
Ri(z) <strong>and</strong> Rm(z)<br />
from chapter 4.<br />
ˆφm<br />
sl<br />
ˆφk<br />
-1<br />
+<br />
ûm<br />
ˆφmi<br />
Rm<br />
Rim(z)<br />
Rmi(z)<br />
<strong>control</strong>lers Ri(z) in chapter 4. To properly state these conditions, first use linearity of the<br />
local <strong>control</strong>lers to express ˆ φi(t) as the sum of filtered inputs:<br />
ˆφi(t) = <br />
Rim(z) ˆ φm(t)+ <br />
m∈Ci<br />
m∈Mi<br />
ˆφim<br />
Ri<br />
ûi<br />
R (s)<br />
im (z)sm(t).<br />
The systems R (s)<br />
im (z) <strong>and</strong> Rim(z) thus describe the contributions from sm(t) <strong>and</strong> ˆ φm(t)<br />
onto the output ˆ φi(t) respectively. A sufficient condition for stability of the interconnected<br />
network is that allRim(z) fori = 1...Nn <strong>and</strong>m ∈ Ci are passive <strong>and</strong> all interconnections<br />
are negative feedback interconnections. This is illustrated in figure B.1, showing the output<br />
feedback interconnection of two local <strong>control</strong>lers.<br />
However, observe that the local <strong>control</strong>ler parts Rim(z) are implicitly defined in (4.4)<br />
of chapter 4 <strong>and</strong> have no direct feed-through terms. This implies that the feedback interconnected<br />
subsystems have a full sample delay that – as will now be shown – makes passivity<br />
unfeasible. This can be observed by making the passivity condition in (B.2) explicit for any<br />
subsystem by substituting the system equations from (B.1), yielding:<br />
(Aixi +Biui) T Pi(Aixi +Biui)−x T i Pixi < 2x T i CT i yi,<br />
where for brevity the dependence on time t has been omitted. This can be rewritten into a<br />
matrix form:<br />
T <br />
T<br />
xi Ai PiAi −Pi AT i PiBi<br />
T <br />
T<br />
xi xi 0 Ci xi<br />
<<br />
.<br />
Ci 0<br />
yi<br />
B T i PiAi B T i PiBi<br />
Since this inequality must hold for all input <strong>and</strong> state trajectoriesyi(t) ∈ R <strong>and</strong>xi(t) ∈ Rn ,<br />
the following matrix inequality must hold:<br />
<br />
T Ai PiAi −Pi AT i PiBi −C T <br />
i ≺ 0.<br />
yi<br />
B T i PiAi −Ci B T i PiBi<br />
The two Schur complements for this inequality are:<br />
<br />
T Ai PiAi −Pi ≺ 0,<br />
BT i PiBi − BT i PiAi<br />
<br />
T −1<br />
T −Ci Ai PiAi −Pi Ai PiBi −C T <br />
i ≺ 0,<br />
yi<br />
+<br />
sj<br />
ˆφq<br />
ˆφi<br />
yi
Appendix B On using local passivity to enforce global stability 205<br />
<br />
B T i PiBi ≺ 0,<br />
A T i PiAi −Pi − A T i PiBi −C T i<br />
<br />
T −1 <br />
T Bi PiBi Bi PiAi −Ci ≺ 0.<br />
The second complement is not feasible by the fact that Pi ≻ 0 <strong>and</strong> thus B T i PiBi must<br />
be positive (semi-)definite. To see that the first complement is not feasible, first note that<br />
A T i PiAi −Pi can be negative definite. However, in that case the matricesB T i PiBi <strong>and</strong> the<br />
product−X T (A T i PiAi−Pi) −1 X for any matrixX will be positive semi-definite. As their<br />
sum will also be positive definite, this is in contradiction with the second part of the first<br />
Schur complement. Therefore, this concept of passivity cannot be used to enforce stability<br />
of the interconnected network of discrete time <strong>control</strong>lersRi(z) in chapter 4.
ÔÔÒÜ<br />
ÓÙÖÖ×Ö×ÓÈÙÐ×ÏØ<br />
ÅÓÙÐØÓÒ ÈÏÅ ×ÒÐ<br />
A PWM signal is a periodic signal with period TPWM that switches between a high<br />
value <strong>and</strong> a low value depending on a duty cyclerPWM ∈ {0,1}. For the BD modulation<br />
principle used in section 6.2.2, the high value is equal to the clamp voltageVcc <strong>and</strong> the low<br />
value equal to zero. Let the PWM signaly(t) be defined as a function of time t as:<br />
<br />
Vcc, for −TPWMrPWM/2+kTPWM ≤ t ≤ TPWMrPWM/2+kTPWM,<br />
y(t) =<br />
0, otherwise,<br />
(C.1)<br />
wherek ∈ Z.<br />
According to Fourier theory, any periodic signal can be written as a sum of sines <strong>and</strong> cosines.<br />
Since this signal is symmetric int = 0, it can be expressed in cosines only as:<br />
y(t) = a0 +<br />
∞<br />
ancos(2πnfPWMt), (C.2)<br />
n=1<br />
where fPWM = 1/TPWM. The Fourier coefficients an for n ≥ 1 can be determined by<br />
integrating the product ofy(t) with a single cosine over one full period. Let this integralIn<br />
be defined as:<br />
In =<br />
=<br />
TPWM/2 <br />
−TPWM/2<br />
TPWM/2 <br />
−TPMW/2<br />
y(t)cos(2πfPWMnt)dt,<br />
<br />
a0 +<br />
∞<br />
<br />
amcos(2πmfPWMt) cos(2πfPWMnt)dt,<br />
m=1<br />
where the second step follows from substitution of (C.2). This can be further simplified<br />
to In = anTPWM/2 for n ≥ 1 <strong>and</strong> to I0 = a0TPWM for n = 0 using the following<br />
goniometric identities:<br />
TPWM/2 <br />
−TPWM/2<br />
⎧<br />
⎪⎨ 0 forn = m,<br />
cos(2πfmt)cos(2πfnt) = TPWM forn = m = 0,<br />
⎪⎩<br />
TPWM/2 forn = m ≥ 1.<br />
206
Appendix C Fourier series of a PWM signal 207<br />
Conversely, the coefficientsan forn ≥ 1 can be calculated as:<br />
an = 2In<br />
TPWM<br />
=<br />
=<br />
2<br />
TPWM/2 <br />
TPWM<br />
−TPWM/2<br />
2<br />
τ/2<br />
TPWM<br />
−τ/2<br />
= 2Vccsin(2πfPWMnt)<br />
2πfPWMnTPWM<br />
= Vcc<br />
= Vcc<br />
y(t)cos(2πfPWMnt)dt,<br />
Vcccos(2πfPWMnt)dt, (C.3)<br />
τ/2<br />
,<br />
−τ/2<br />
sin(πfPWMnτ)<br />
,<br />
πn<br />
sin(πnrPWM)<br />
. (C.4)<br />
πn<br />
where τ = TPWMrPWM . In (C.3), the signal y(t) from (C.1) is substituted <strong>and</strong> the remaining<br />
steps use TPWMfPWM = 1 <strong>and</strong> other common algebra identities to simplify the<br />
result.<br />
The derived result is not valid forn = 0, buta0 can be obtained directly asa0 = I0/TPWM,<br />
which reduces to:<br />
TPWM/2 <br />
1<br />
a0 =<br />
TPWM<br />
−TPWM/2<br />
y(t)dt = Vcct<br />
τ/2 = VccrPWM.<br />
TPWM −τ/2<br />
Summarizing, the PWM signaly(t) as defined in (C.1) can be expressed as an infinite series<br />
of cosines by substituting thisa0 <strong>and</strong>an from (C.4) into (C.2), yielding:<br />
<br />
∞<br />
<br />
sin(πnrPWM)<br />
y(t) = Vcc rPWM + cos(2πnfPWMt) .<br />
πn<br />
n=1
ÔÔÒÜ<br />
ÌÄÎËÔÖÓØÓÓÐ<br />
The electronics modules communicate with the communication bridge over an Low<br />
Voltage Differential Signalling (LVDS) connection. This communication method uses a<br />
current source to transmit information instead of a voltage source, which makes it much<br />
less sensitive to the cable length or resistance. The sign of the current defines the binary<br />
high <strong>and</strong> low values.<br />
The LVDS connection does not use a clock signal to synchronize the communication. Each<br />
message is preceded by 18 pause bits on which a message pointer can be synchronized <strong>and</strong><br />
one start <strong>and</strong> one stop bit on which the 16-bit data words can be synchronized. A start bit<br />
is high <strong>and</strong> the pause <strong>and</strong> stop bits are low. No parity bits are used.<br />
Each message consists of four parts: 18 pause bits, a header, the data <strong>and</strong> a checksum. The<br />
header consists of one 16-bit word of which the lower eight bits are formed by the module<br />
identifier <strong>and</strong> the upper eight by the message identifier:<br />
byte index number of bits description<br />
0 8 module identifier<br />
1 8 message identifier<br />
2 depends on message type comm<strong>and</strong> data<br />
? 16 checksum<br />
The sum of all data words – including the checksum itself – equals zero, which makes<br />
the checksum the 2’s complement of the sum of the preceding words. The data structure<br />
depends on the message identifier, which can be one of the following.<br />
Burst write (1)<br />
A burst-write message contains setpoints for all 61 actuators on the specified module. The<br />
message identifier for this message is 1 <strong>and</strong> the comm<strong>and</strong> data field is defined as:<br />
byte index number of bits description<br />
0 16 setpoint for actuator 1<br />
2 16 setpoint for actuator 2<br />
.<br />
.<br />
.<br />
120 16 setpoint for actuator 61<br />
208
Appendix D The LVDS protocol 209<br />
Register write (2)<br />
A register write message can be used to modify a specific register on one of the modules.<br />
The message identifier for this message is 2 <strong>and</strong> the comm<strong>and</strong> data field is defined as:<br />
Register read (3)<br />
byte index number of bits description<br />
0 16 register number<br />
2 16 new register value<br />
A register read message can be used to read back a specific register on one of the modules.<br />
The message identifier for this message is 3 <strong>and</strong> the comm<strong>and</strong> data field consists only of a<br />
16-bit register number.<br />
The module will respond with a message with the format of a register write <strong>and</strong> the module<br />
<strong>and</strong> message identifiers set to the values used in the requesting register read message.<br />
Registers<br />
Each module contains the following registers:<br />
index description<br />
0x00. . . 0x1e PWM setpoints for actuators 1. . . 31<br />
0x1f not used<br />
0x20. . . 0x21 Enable PWM A bits for actuators 1. . . 31<br />
0x22. . . 0x23 Enable PWM B bits for actuators 1. . . 31<br />
0x24. . . 0x25 Enable PWM C bits for actuators 1. . . 31<br />
0x26. . . 0x27 Coil integrity <strong>control</strong> bits for actuators 1. . . 31 (read-only)<br />
0x28 Global settings register for actuators 1. . . 31<br />
0x29. . . 0x3f not used<br />
0x40. . . 0x5d PWM setpoints for actuators 32. . . 61<br />
0x5e. . . 0x5f not used<br />
0x60. . . 0x61 Enable PWM A bits for actuators 32. . . 61<br />
0x62. . . 0x63 Enable PWM B bits for actuators 32. . . 61<br />
0x64. . . 0x65 Enable PWM C bits for actuators 32. . . 61<br />
0x66. . . 0x67 Coil integrity <strong>control</strong> bits for actuators 32. . . 61 (read-only)<br />
0x68 Global settings register for actuators 32. . . 61<br />
0x69. . . 0x7f not used<br />
0x80. . . 0xff reserved<br />
Global settings registers<br />
Each slave Field Programmable Gate Array (FPGA) has a global, 16-bit settings register<br />
with which its global behavior can be <strong>control</strong>led. Its bits have the following meaning:
210 Appendices<br />
Bit index Description<br />
0. . . 3 Dead-time in 8ns steps (default = ’0111’)<br />
4. . . 6 Not used<br />
7 Set global PWM B to high (used for coil integrity check)<br />
8 Global PWM C enable<br />
9 Global PWM B enable<br />
10 Global PWM A enable<br />
11 Testmode (used for coil integrity check)<br />
12:15 Revision number (= ’0001’)<br />
Coil integrity check<br />
The FPGA slaves can be put into test-mode to perform the so-called coil integrity check.<br />
This can be used to determine whether the actuator coil conducts electricity or not.<br />
First, the fine PWM C signal that is directly provided by the FPGA is set to high (figure 6.4)<br />
<strong>and</strong> the course PWM signals are disabled. If the actuator coil does not conduct (e.g. because<br />
of a broken wire), this will charge capacitorCl (figure 6.3) <strong>and</strong> build up a capacitor voltage.<br />
Otherwise, the actuator coil will prevent this build-up <strong>and</strong> the voltage will quickly drop to<br />
zero when PWM C is disabled. This behavior can be checked by reversing the directionality<br />
of the FPGA pin corresponding to the PWM C signal <strong>and</strong> using it to measure the capacitor<br />
voltage. If the voltage is zero, the actuator coil is fine <strong>and</strong> if it’s high, it does not conduct<br />
electricity.<br />
These operations are <strong>control</strong>led via specific bits of the global settings registers. The coil<br />
integrity check procedure should be as follows:<br />
1. Enable bits 7 <strong>and</strong> 11 of the global settings registers. This enables the test mode for<br />
which the course PWM signals are disabled <strong>and</strong> the fine PWM signal is permanently<br />
high.<br />
2. Wait at least four times the time constantRaCl of the system to allow charging of the<br />
capacitor.<br />
3. Set bits 7 of the global settings registers to zero, causing the FPGA pins corresponding<br />
to PWM’s C to go into tri-state <strong>and</strong> allowing them to measure the capacitor voltages.<br />
4. Disable the test-mode by setting bits 11 of the global settings registers to zero. The<br />
results of the voltage measurements can now be found in the coil integrity <strong>control</strong> registers.<br />
These should be zero for conducting actuator coils <strong>and</strong> one for malfunctioning<br />
ones.
ÔÔÒÜ<br />
ÌÍÈÔÖÓØÓÓÐ<br />
User Datagram Protocol (UDP) is a very lean communication protocol, since no feedback<br />
is given whether the message is properly received. Each message is a st<strong>and</strong>-alone<br />
message <strong>and</strong> not part of a stream such as a Transmission Control Protocol (TCP)/Internet<br />
Protocol (IP) message. Besides the required IP header, the message contains a small UDP<br />
header that contains the source <strong>and</strong> destination ports, the message length <strong>and</strong> a checksum:<br />
byte index number of bits description<br />
0 16 source port<br />
2 16 destination port<br />
4 16 message length<br />
6 16 checksum<br />
8 ?? UDP data<br />
The checksum value is ignored by the LVDS communications bridge to improve latency.<br />
The data of the UDP message comes after the checksum field <strong>and</strong> has a substructure that<br />
consists of the following parts: a module identifier (8 bits), a message identifier (8 bits) <strong>and</strong><br />
the comm<strong>and</strong> data:<br />
byte index number of bits description<br />
0 8 module identifier<br />
1 8 message identifier<br />
2 ?? comm<strong>and</strong> data<br />
The module identifier byte specifies the module for which the message is intended <strong>and</strong> the<br />
message identifier specifies the type of message that follows. The comm<strong>and</strong> data specification<br />
depends on the type of message used. The message type can be one of the following.<br />
Burst write<br />
A burst-write message contains setpoints for all 61 actuators on the specified module. The<br />
message identifier for this message is 1 <strong>and</strong> the comm<strong>and</strong> data field is defined as:<br />
211
212 Appendices<br />
byte index number of bits description<br />
0 16 setpoint for actuator 1<br />
2 16 setpoint for actuator 2<br />
.<br />
.<br />
.<br />
.<br />
.<br />
.<br />
120 16 setpoint for actuator 61<br />
122 16 Reserved<br />
124 16 Sequence number (for diagnostics)<br />
Together with the module <strong>and</strong> message identifier bytes, a burst write message is thus exactly<br />
128 bytes in length. It is possible to include up to eight of such burst messages in one single<br />
UDP message. Since the UDP-to-LVDS communication bridge has a processing time of<br />
85µs per message, this significantly reduces the latency in case of many modules.<br />
The communications bridge will recognize a combined UDP message by its size <strong>and</strong> splits<br />
it into multiple LVDS messages before transmitting them sequentially over the LVDS connection.<br />
The limit of eight burst messages per UDP message is due to the protocol limit of<br />
the UDP message size.<br />
Register write<br />
A register write message can be used to modify a specific register on one of the modules.<br />
The message identifier for this message is 2 <strong>and</strong> the comm<strong>and</strong> data field is defined as:<br />
Register read<br />
byte index number of bits description<br />
0 16 register number<br />
2 16 new register value<br />
A register read message can be used to read back a specific register on one of the modules.<br />
The message identifier for this message is 3 <strong>and</strong> the comm<strong>and</strong> data field consists only of a<br />
16-bit register number.<br />
The module will respond over the LVDS connection in the form specified in appendix D.<br />
This is wrapped into the st<strong>and</strong>ard UDP form by the communication bridge. The module <strong>and</strong><br />
message identifiers will be identical to those bytes of the request message <strong>and</strong> the comm<strong>and</strong><br />
data field is identical to that of a register write message.<br />
Diagnostic messages<br />
Finally, the communications bridge counts the number of burst messages that it has properly<br />
received since start-up. This 16-bit number can be read back by reading register 255 of<br />
module 255. This counter can also be set by writing to this same register.
ÔÔÒÜ<br />
ËÔØÐÚÖØÓÒÓØÙØÓÖ<br />
ÔÖÓÔÖØ×<br />
The two figures below illustrate that there is no obvious correlation between the values<br />
of the resonance frequency <strong>and</strong> motor constant of the actuators measured <strong>and</strong> their location<br />
in the grid.<br />
Figure F.1: The value of the resonance frequencies,<br />
represented proportionally by the size of the<br />
dots for the modules 1...7 from left to right <strong>and</strong><br />
then top to bottom.<br />
213<br />
Figure F.2: The value for the motor constants,<br />
represented proportionally by the size of the dots<br />
for the modules 1...7 from left to right <strong>and</strong> then<br />
top to bottom.
ÔÔÒÜ<br />
ÉÙÒØÞØÓÒ<br />
In chapter 6 a choice was made to use a 16 bit PWM voltage signal generator to drive the<br />
actuators, leading to sufficiently small quantization errors. At that point, the consideration<br />
was based on general design parameters of the system, which is here performed in more<br />
detail to validate the choice.<br />
The Root Mean Square (RMS) wavefront error σquant due to quantization errors has been<br />
specified in chapter 2 as at most 5nm. Due to quantization, the actual comm<strong>and</strong> vector<br />
˜VPWM will be the sum of the intended actuator voltage VPWM <strong>and</strong> quantization noise n.<br />
Since wavefront correction is only possible for frequencies below the <strong>control</strong> b<strong>and</strong>width,<br />
which is generally below the first system resonance, the static DM system model from (7.7)<br />
on page 164 with influence matrixBf will here be used:<br />
ˆzf = Bf ˜ VPWM = Bf(VPWM +n),<br />
where ˆzf denotes the actual facesheet deflection in contrast to the facesheet deflection for<br />
the intended comm<strong>and</strong> VPWM. According to the design requirements, the variance σ 2 quant<br />
of the quantization noisenshould be such thatσquant = 〈BfnF〉 ≤ 5nm, where<br />
σ 2 quant<br />
1<br />
= Tr<br />
Na<br />
T<br />
Bf nn B T f . (G.1)<br />
To derive the covariance matrix nnT , let the elementsni of n be uncorrelated stochastic<br />
values with a square probability density functionPn(ni):<br />
<br />
q<br />
Pn(ni) =<br />
−1 for |ni| ≤ q/2,<br />
0 otherwise,<br />
where q is the comm<strong>and</strong> quantization step size in Volt. Each diagonal element of nn T<br />
can thus be expressed as [132]:<br />
2<br />
ni =<br />
∞<br />
−∞<br />
n 2 iPn(ni)dni =<br />
q/2<br />
−q/2<br />
n 2 i<br />
q dni = q2<br />
12<br />
hence nn T = q<br />
12 I. Using (G.1) the q for which σquant is smaller than 5nm can now be<br />
214
Appendix G Quantization 215<br />
derived as:<br />
1<br />
Na<br />
2 q<br />
Tr<br />
12 BfB T <br />
f ≤ (5·10 −9 ) 2 ,<br />
q2 12 Tr(BfB T f ) ≤ (5·10−9 ) 2 Na,<br />
q ≤<br />
<br />
12(5·10 −9 ) 2Na Tr(BfBT f )<br />
.<br />
Using the influence matrix identified in section 7.3.2 for the 61 actuator DM system this<br />
yields q ≈ 190µV, which is only slightly more than the 100µV accuracy provided by the<br />
realized 16 bit PWM voltage source with a supply voltage of 3.3V. The quantization value<br />
of the driver electronics was thus properly chosen.<br />
Finally, note that incorporation of noise shaping techniques [78, 158, 162] into the <strong>control</strong>ler<br />
design may enlarge the required quantization step. Such techniques could push the<br />
quantization effects to (high) temporal frequencies for which the wavefront disturbance has<br />
a low magnitude, so reducing the effect on the optical performance.
ÐÓÖÔÝ<br />
[1] www.goodfellow.com.<br />
[2] C. A. Aerts. Optimizing <strong>and</strong> implementing leaf springs for the reluctance motor.<br />
Technical Report DCT 2005.94, Technische Universiteit Eindhoven, 2005.<br />
[3] S. T. Alex<strong>and</strong>er. <strong>Adaptive</strong> signal processing. Texts <strong>and</strong> monographs in computer<br />
science. Springer-Verlag, New York, 1986.<br />
[4] P. Alriksson <strong>and</strong> A. Rantzer. Distributed Kalman filtering using weighted averaging.<br />
In Proceedings of the 17 th International Symposium on Mathematical Theory of<br />
Networks <strong>and</strong> Systems, Kyoto, Japan, july 2006.<br />
[5] P. Alriksson <strong>and</strong> A. Rantzer. Experimental evaluation of a distributed Kalman filter<br />
algorithm. In Proceedings of the 46 th IEEE Conference on Decision <strong>and</strong> Control,<br />
New Orleans, LA, December 2007.<br />
[6] T. Andersen, Mette-Owner-Petersen, <strong>and</strong> H. Riewaldt. An integrated simulation<br />
model of the euro50. In T. Andersen, editor, Proceedings of SPIE: Integrated Modeling<br />
of Telescopes, volume 4757, pages 84–92, July 2002.<br />
[7] V. Apollonov, V. Borodin, A. Brynskikh, G. Vdovin, <strong>and</strong> S. Murav. Cooled adaptive<br />
<strong>mirror</strong> with magnetostrictive spring-type actuators. Soviet Journal of Quantum<br />
Electronics, 20(11):1403–1406, 1990.<br />
[8] M. Arcak <strong>and</strong> E. D. Sontag. A passivity-based stability criterion for a class of interconnected<br />
systems <strong>and</strong> applications to biochemical reactor networks. In Proceedings<br />
of the 46 th IEEE Conference on Decision <strong>and</strong> Control, pages 4477–4482, New Orleans,<br />
LA, USA, December 2007.<br />
[9] R. Arsenault, R. Biasi, D. Gallieni, A. Riccardi, P. Lazzarini, N. Hubin, E. Fedrigo,<br />
R. Donaldson, S. Oberti, S. Stroebele, R. Conzelmann, <strong>and</strong> M. Duchateau. A <strong>deformable</strong><br />
secondary <strong>mirror</strong> for the VLT. In Proceedings of SPIE, volume 6272, 2006.<br />
[10] F. Assémat, E. Gendron, <strong>and</strong> F. Hammer. The FALCON concept: multi-object adaptive<br />
optics <strong>and</strong> atmospheric tomography for integral field spectroscopy Ű principles<br />
<strong>and</strong> performance on an 8-m telescope. Monthly Notices of the Royal Astronomical<br />
Society, 376(1):287–312, March 2007.<br />
[11] N. Assimakis. Optimal distributed kalman filter. In Nonlinear analysis, volume 47,<br />
pages 5367–5378. Elsevier Science Ltd., 2001.<br />
[12] H. Babcock. The possibility of compensating astronomical seeing. Publications of<br />
the Astronomical Society of the Pacific, 65(386):229–236, October 1953.<br />
[13] B. Bamieh, F.Paganini, <strong>and</strong> M. A. Dahleh. Optimal <strong>control</strong> of distributed arrays with<br />
spatial invariance. Lecture Notes in Control <strong>and</strong> Information Sciences, 245:329–343,<br />
1999.<br />
[14] B. Bamieh, F.Paganini, <strong>and</strong> M. A. Dahleh. Distributed <strong>control</strong> of spatially invariant<br />
systems. IEEE Transactions on Automatic Control, 47:1091–1107, July 2002.<br />
216
Bibliography 217<br />
[15] F. Bazán. Eigensystem realization algorithm (ERA): reformulation <strong>and</strong> system pole<br />
perturbation analysis. Journal of Sound <strong>and</strong> Vibration, 274:433–444, 2004.<br />
[16] C. Béchet, M. Tallon, <strong>and</strong> E. Thiébaut. FRIM: minimum-variance reconstructor with<br />
a FRactal Iterative Method. In B. L. Ellerbroek <strong>and</strong> D. C. B. Calia, editors, Proceedings<br />
of SPIE: Astronomical Telescopes <strong>and</strong> Instrumentation, volume 6272, May<br />
2006.<br />
[17] A. Bejan. Heat transfer. John Wiley <strong>and</strong> Sons Ltd, 1990.<br />
[18] R. Biasi, M. Andrighettoni, D. Veronese, V. Biliotti, L. Fini, A. Riccardi, P. Mantegazza,<br />
<strong>and</strong> D. Gallieni. LBT adaptive secondary electronics. In P. L. Wizinowich<br />
<strong>and</strong> D. Bonaccini, editors, Proceedings of SPIE: Astronomical Telescopes <strong>and</strong> Instrumentation,<br />
volume 4839, pages 772–782, Waikoloa, Hawaii, USA, August 2002.<br />
[19] K. Birch <strong>and</strong> M. Downs. Correction to the updated Edlen equation for the refractive<br />
index of air. Metrologia, 31(4):315–316, 1994.<br />
[20] M. Born <strong>and</strong> E. Wolf. Principles of Optics: Electromagnetic Theory of Propagation,<br />
Interference, <strong>and</strong> Diffraction of Light. Cambridge University Press,7 th edition, 1999.<br />
[21] A. H. Bouchez, R. G. Dekany, J. R. Angione, C. Barane, K. Bui, R. S. Burruss,<br />
J. R. Crepp, E. E. Croner, J. L. Cromer, S. R. Guiwits, D. D. Hale, J. R. Henning,<br />
D. Palmer, J. E. Roberts, M. Troy, T. N. Truong, <strong>and</strong> J. Zolkower. Status of the<br />
PALM-3000 high-order adaptive optics system. In N. Hubin, C. E. Max, <strong>and</strong> P. L.<br />
Wizinowich, editors, Proceedings of SPIE: <strong>Adaptive</strong> Optics Systems, volume 7015,<br />
2008.<br />
[22] J. Brenner <strong>and</strong> L. Cummings. The hadamard maximum determinant problem. American<br />
Mathematics Montly, 79(6):626–630, 1972.<br />
[23] J. W. Brewer. Kronecker Products <strong>and</strong> Matrix Calculus in System Theory. IEEE<br />
Transactions on Circuits <strong>and</strong> Systems, 25(9):772–781, September 1978.<br />
[24] G. Brusa, A. Riccardi, V. Biliotti, C. D. Vecchio, P. Salinari, P. Stefanini, P. Mantegazza,<br />
R. Biasi, M. Andrighettoni, C. Franchini, <strong>and</strong> D. Gallieni. <strong>Adaptive</strong> secondary<br />
<strong>mirror</strong> for the 6.5m conversion of the Multiple Mirror Telescope: first laboratory<br />
testing results. In R. K. Tyson <strong>and</strong> R. Q. Fugate, editors, Proceedings of SPIE:<br />
<strong>Adaptive</strong> Optics Systems <strong>and</strong> Technology, volume 3762, September 1999.<br />
[25] G. Brusa-Zappellini, A. Riccardi, S. D. Ragl<strong>and</strong>, S. Esposito, C. D. Vecchio, L. Fini,<br />
P. Stefanini, V. Biliotti, P. Ranfagni, P. Salinari, D. Gallieni, R. Biasi, P. Mantegazza,<br />
G. Sciocco, G. Noviello, <strong>and</strong> S. Invernizzi. <strong>Adaptive</strong> secondary P30 prototype: laboratory<br />
results. In D. Bonaccini <strong>and</strong> R. K. Tyson, editors, Proceedings of SPIE: <strong>Adaptive</strong><br />
Optical System Technologies, volume 3353, pages 764–775, September 1998.<br />
[26] D. Buscher. A Thous<strong>and</strong> <strong>and</strong> One Nights of Seeing on Mt. Wilson. In J. B. Breckinridge,<br />
editor, Proceedings of SPIE: Amplitude <strong>and</strong> Intensity Spatial Interferometry<br />
II, volume 2200, pages 260–271, 1994.<br />
[27] Y. Carmon <strong>and</strong> E. N. Ribak. Phase retrieval by demodulation of a Hartmann-Shack<br />
sensor. Optics Communications, 215:285–288, January 2003.<br />
[28] R. Changhai <strong>and</strong> S. Lining. Hysteresis <strong>and</strong> creep compensation for piezoelectric<br />
actuator in open-loop operation. Sensors <strong>and</strong> Actuators A, 122:124–130, 2005.<br />
[29] A. Chiuso. The role of vector autoregressive modeling in predictor-based subspace<br />
identification. Automatica, 43:1034–1048, 2007.
218 Bibliography<br />
[30] A. Chiuso, R. Muradore, <strong>and</strong> E. Marchetti. Dynamic calibration of adaptive optics<br />
systems: a system identification approach. In Proceedings of the Conference on<br />
Decision <strong>and</strong> Control, 2008.<br />
[31] A. Chiuso <strong>and</strong> G. Picci. Consistency analysis of some closed loop subspace identification<br />
methods. Automatica, 41(3):377–391, 2005.<br />
[32] P. Ciddor. Refractive index of air: new equations for the visible <strong>and</strong> near infrared.<br />
Applied Optics, 35:1566–1573, 1996.<br />
[33] J.-M. Conan, G. Rousset, <strong>and</strong> P.-Y. Madec. Wavefront temporal spectra in highresolution<br />
imaging through turbulence. Journal of the Optical Society of America A,<br />
12:1559–1570, July 1995.<br />
[34] R. Conan, R. Avila, <strong>and</strong> L. Sanchez. Wavefront outer scale <strong>and</strong> seeing measurements<br />
at San Pedro Martir Observatory. Astronomy <strong>and</strong> Astrophysics, 396:723–730, 2002.<br />
[35] D. T. Corporation. SigLab Users Guide, 1994.<br />
[36] A. Cox, editor. AllenŠs Astrophysical Quantities. AIP Press., 2000.<br />
[37] L. Cuellar, P. Johnson, <strong>and</strong> D. S<strong>and</strong>lar. Performance tests of a 1500 degree-offreedom<br />
adaptive optics system for atmspheric compensation. In Proceedings of<br />
SPIE: Active <strong>and</strong> adaptive optical systems, volume 1542, pages 468–476, 1991.<br />
[38] O. Cugat, S. Basrour, C. Divoux, P. Mounaix, <strong>and</strong> G. Reyne. Deformable magnetic<br />
<strong>mirror</strong> for adaptive optics: technological aspects. Sensors <strong>and</strong> Actuators A: Physical,<br />
89:1–9, 2001.<br />
[39] O. Cugat, P. Mounaix, S. Basrour, C. Divoux, <strong>and</strong> G. Reyne. Deformable magnetic<br />
<strong>mirror</strong> for adaptive optics: first results. In The13 th Annual International Conference<br />
on Micro Electro Mechanical Systems, 2000.<br />
[40] J. Dainty, D. Hennings, <strong>and</strong> K. O’Donnel. Space-time correlation of stellar speckle<br />
patterns. Journal of the Optical Society of America, 71:490–492, April 1981.<br />
[41] R. D’Andrea <strong>and</strong> G. E. Dullerud. Distributed Control Design for Spatially Interconnected<br />
Systems. IEEE Transactions on Automatic Control, 48(9):1478–1495,<br />
September 2003.<br />
[42] W. de Bruijn. Low power <strong>deformable</strong> <strong>mirror</strong> actuator <strong>control</strong>ler. Master’s thesis,<br />
Technische Universiteit Eindhoven (TU/e), Den Dolech 2, 5600MB, Eindhoven, The<br />
Netherl<strong>and</strong>s, 2009.<br />
[43] P. Dierickx <strong>and</strong> R. Gilmozzi. Progress of the OWL 100-m Telescope Conceptual<br />
Design. In T. A. Sebring <strong>and</strong> T. Andersen, editors, Proceedings of SPIE: Telescope<br />
Structures, Enclosures, Controls, Assembly/Integration/Validation, <strong>and</strong> Commissioning,<br />
volume 4004, pages 290–299, Munich, March 2000.<br />
[44] N. Doelman, R. Fraanje, I. Houtzager, <strong>and</strong> M. Verhaegen. <strong>Adaptive</strong> <strong>and</strong> real-time<br />
optimal <strong>control</strong> for adaptive optics systems. European Journal of Control, 15:480–<br />
502, 2009.<br />
[45] N. Doelman, K. Hinnen, F. Stoffelen, <strong>and</strong> M. Verhaegen. Optimal <strong>control</strong> strategy to<br />
reduce the temporal wavefront error in AO systems. In D. B. Calia, B. L. Ellerbroek,<br />
<strong>and</strong> R. Ragazzoni, editors, Proceedings of SPIE: Advancements in <strong>Adaptive</strong> Optics,<br />
volume 5490, pages 1426–1437, June 2004.<br />
[46] M. Ealey. Deformable <strong>mirror</strong>s at Litton/Itek: A historical persepective. In D. Vukobratovich,<br />
editor, Proceedings of SPIE: Precision Engineering <strong>and</strong> Optomechanics,<br />
volume 1167, November 1989.
Bibliography 219<br />
[47] M. Ealey <strong>and</strong> J. Washeba. Continous facesheet low voltage <strong>deformable</strong> <strong>mirror</strong>s. Optical<br />
Engineering, 29:1191–1198, October 1990.<br />
[48] F. Eaton <strong>and</strong> G. Nastrom. Preliminary estimates of the vertical profiles of inner <strong>and</strong><br />
outer scales from White S<strong>and</strong>s Missile Range. Radio Science, 33:895–903, 1998.<br />
[49] Edlén. The refractive index of Air. Metroligia, 2:71–80, 1966.<br />
[50] S. E. Egner. Multi-Conjugate <strong>Adaptive</strong> Optics for LINC-NIRVANA. PhD thesis, Max<br />
Planck Institute for Astronomy, University of Heidelberg, 2006.<br />
[51] R. Ellenbroek <strong>and</strong> R. Hamelinck. Adaptief deformeerbare spiegel voor telescopen.<br />
Precisietechnologie jaarboek, 16:112–118, 2009.<br />
[52] R. Ellenbroek, M. Verhaegen, R. Hamelinck, N. Doelman, M. Steinbuch, <strong>and</strong><br />
N. Rosielle. Distributed <strong>control</strong> in <strong>Adaptive</strong> Optics - Deformable <strong>mirror</strong> <strong>and</strong> turbulence<br />
modeling. In B. L. Ellerbroek <strong>and</strong> D. C. B. Calia, editors, Proceedings of<br />
SPIE: Astronomical telescopes <strong>and</strong> instrumentation - Advances in <strong>Adaptive</strong> Optics,<br />
volume 6272, May 2006.<br />
[53] B. Ellerbroek. Efficient computation of minimum-variance wave-front reconstructors<br />
with sparse matrix techniques. Journal of the Optical Society of America A, 19:1803–<br />
1816, September 2002.<br />
[54] B. Ellerbroek <strong>and</strong> C. Vogel. Simulations of closed-loop wavefront reconstruction for<br />
multiconjugate adaptive optics on giant telescopes. In Proc. SPIE 5169-23, <strong>Adaptive</strong><br />
Optics System Technologies II, pages 206–217, 2003.<br />
[55] B. L. Ellerbroek. First-order performance evaluation of adaptive-optics systems for<br />
atmospheric-turbulence compensation in extended-field-of-view astronomical telescopes.<br />
In Journal of the Optical Society of America A, volume 11, pages 783–805,<br />
Februari 1994.<br />
[56] B. L. Ellerbroek <strong>and</strong> T. Rhoadarmer. Real time adaptive optimization of wavefront<br />
reconstruction algorithms for closed loop adaptive optical systems. In D. Bonaccini<br />
<strong>and</strong> R. K. Tyson, editors, Proceedings of SPIE: <strong>Adaptive</strong> Optical System Technologies,<br />
volume 3353, pages 1174–1185, March 1998.<br />
[57] S. Fassois. MIMO LMS-ARMAX identification of vibrating structures - part I: the<br />
method. Mechanical systems <strong>and</strong> signal processing, 15(4):723–735, 2001.<br />
[58] E. Fedrigo <strong>and</strong> R. Donaldson. Architecture of the MAD realtime computer. In P. L.<br />
Wizinowich <strong>and</strong> D. Bonaccini, editors, Proceedings of SPIE: <strong>Adaptive</strong> Optical System<br />
Technologies II, volume 4839, pages 600–611, February 2003.<br />
[59] A. E. Fitzgerald, C. K. Jr., <strong>and</strong> S. D. Umans. Electric Machinery. McGraw-Hill, 6 t h<br />
edition, 2003.<br />
[60] A. Florakis <strong>and</strong> S. Fassois. MIMO LMS-ARMAX identification of vibrating structures<br />
- part II: a critical assessment. Mechanical systems <strong>and</strong> signal processing, 15,<br />
2001.<br />
[61] R. Fraanje <strong>and</strong> M. Verhaegen. A spatial canonical approach to multidimensional<br />
state-space identification for distributed parameter systems. In Proceedings of the<br />
fourth international workshop on multidimensional systems, 2005.<br />
[62] G. F. Franklin, J. Powell, <strong>and</strong> A. Emami-Naeini. Feedback <strong>control</strong> of dynamic systems.<br />
Addison-Wesley, 3rd edition, 1994.<br />
[63] D. Fried. Statistics of a Geometric Representation of Wavefront Distortion. Journal<br />
of the Optical Society of America, 55:1427–1435, November 1965.
220 Bibliography<br />
[64] D. Fried. Time delay induced mean square error in adaptive optics. Journal of the<br />
Optical Society of America A, 7(7):1224–1225, 1990.<br />
[65] M. E. Furber <strong>and</strong> D. Jordan. Optimal design of wavefront sensors for adaptive optical<br />
systems: part 1, <strong>control</strong>lability <strong>and</strong> observability analysis. Optical Engineering,<br />
36:1843–1855, July 1997.<br />
[66] M. E. Furber <strong>and</strong> D. Jordan. Optimal design of wavefront sensors for adaptive optical<br />
systems: part 2, optimization of sensor subaperture locations. Optical Engineering,<br />
36:1856–1871, July 1997.<br />
[67] A. Fuschetto. Three-actuator <strong>deformable</strong> water-cooled <strong>mirror</strong>. Journal of Optical<br />
Engineering, 20(2):310–315, March/April 1981.<br />
[68] D. Gallieni, E. Anaclerio, P. Lazzarini, A. Ripamonti, R. Spairani, C. DelVecchio,<br />
P. Salinari, A. Riccardi, P. Stefanini, <strong>and</strong> R. Biasi. LBT adaptive secondary units final<br />
design <strong>and</strong> construction. In P. L. Wizinowich <strong>and</strong> D. Bonaccini, editors, Proceedings<br />
of SPIE: <strong>Adaptive</strong> Optical System Technologies II, volume 4839, pages 765–771,<br />
2003.<br />
[69] A. Gelb. Applied optimal estimation. The MIT press, 15th edition, 1999.<br />
[70] D. Geng, S. J. Goodsell, A. G. Basden, N. A. Dipper, R. M. Myers, <strong>and</strong> C. D. Saunter.<br />
FPGA cluster for high-performance AO real-time <strong>control</strong> system. In B. L. Ellerbroek<br />
<strong>and</strong> D. C. B. Calia, editors, Proceedings of the SPIE: Astronomical telescopes <strong>and</strong><br />
instrumentation - Advances in <strong>Adaptive</strong> Optics, volume 6272, 2006.<br />
[71] S. Gerschgorin. Über die Abgrenzung der Eigenwerte einer Matrix. Izv. Akad. Nauk.<br />
SSSR Ser. Mat., 1:749–754, 1931.<br />
[72] L. Gilles. Order-N sparse minimum-variance open-loop reconstructor for extreme<br />
adaptiveoptics. Optics Letters, 28(20):1927–1929, October 2003.<br />
[73] L. Gilles, B. Ellerbroek, <strong>and</strong> C. Vogel. Preconditioned conjugate-gradient wave-front<br />
reconstructors for multiconjugate adaptive optics. Applied Optics, 42:5233–5250,<br />
2003.<br />
[74] L. Gilles, C. R. Vogel, <strong>and</strong> B. Ellerbroek. Multigrid predonditioned conjugategradient<br />
method for large-scale wave-frontreconstruction. Journal of the Optical<br />
Society of America A, 19:1817–1822, September 2002.<br />
[75] R. Gilmozzi <strong>and</strong> J. Spyromilio. The European Extremely Large Telescope. The<br />
Messenger (Telescopes <strong>and</strong> Instrumentation) 127, European Southern Observatory,<br />
March 2007.<br />
[76] A. Glecker, D. Markason, <strong>and</strong> G. Ames. PAMELA; Control of a segmented <strong>mirror</strong><br />
via wavefront tilt <strong>and</strong> segment piston sensing. In M. A. Ealey, editor, Proceedings<br />
op SPIE: Active <strong>and</strong> <strong>Adaptive</strong> Optical Components, volume 1543, pages 176–189,<br />
January 1991.<br />
[77] G. H. Golub <strong>and</strong> C. F. V. Loan. Matrix computations. Johns Hopkins University<br />
Press, Baltimore, MD, USA, 3rd edition, 1996.<br />
[78] G. Goodwin, D. Quevedo, <strong>and</strong> D. McGrath. Moving-horizon optimal quantizer for<br />
audio signals. Journal of the Audio Engineering Society, 51(3):138–149, 2003.<br />
[79] D. Gorinevsky, S. Boyd, <strong>and</strong> G. Stein. Design of Low-b<strong>and</strong>width Spatially Distributed<br />
Feedback. IEEE Transactions on Automatic Control, June 2005.<br />
[80] D. Greenwood. B<strong>and</strong>width specification for adaptive optics systems. Journal of the<br />
Optical Society of America, 67:390, 1977.
Bibliography 221<br />
[81] R. Hamelinck. Ontwerp van een adaptieve deformeerbare spiegel voor het corrigeren<br />
van atmosferische turbulentie. Master’s thesis, Technische Universiteit Eindhoven,<br />
Den Dolech 2, 5600MB, Eindhoven, The Netherl<strong>and</strong>s, 2003.<br />
[82] R. Hamelinck, R. Ellenbroek, N. Rosielle, M. Steinbuch, M. Verhaegen, <strong>and</strong> N. Doelman.<br />
Validation of a new adaptive <strong>deformable</strong> <strong>mirror</strong> concept. In N. Hubin, C. E.<br />
Max, <strong>and</strong> P. L. Wizinowich, editors, Proceedings of SPIE: Astronomical telescopes<br />
<strong>and</strong> instrumentation, volume 7015, Marseille, France, June 2008.<br />
[83] R. Hamelinck, N. Rosielle, <strong>and</strong> N. Doelman. Large adaptive <strong>deformable</strong> membrane<br />
<strong>mirror</strong> with high actuator density: design <strong>and</strong> first actuator tests. In Proceedings of<br />
the5 th Workshop on <strong>Adaptive</strong> Optics for Industry <strong>and</strong> Medicine, August 2005.<br />
[84] R. Hamelinck, N. Rosielle, <strong>and</strong> N. Doelman. Large adaptive <strong>deformable</strong> <strong>mirror</strong>:<br />
design <strong>and</strong> first prototypes. In M. T. Gruneisen, J. D. Gonglewski, <strong>and</strong> M. K. Giles,<br />
editors, Proceedings of SPIE: Advanced Wavefront Control: Methods, Devices, <strong>and</strong><br />
Applications III, volume 5894, San Diego, August 2005.<br />
[85] R. Hamelinck, N. Rosielle, J. Kappelhof, B. Snijders, <strong>and</strong> M. Steinbuch. Large adaptive<br />
<strong>deformable</strong> membrane <strong>mirror</strong> with high actuator density. In D. B. Calia, B. L.<br />
Ellerbroek, <strong>and</strong> R. Ragazzoni, editors, Proceedings of the SPIE: Astronomical Telescopes<br />
<strong>and</strong> Instrumentation, volume 5490, pages 1482–1492, Glasgow, UK, June<br />
2004.<br />
[86] R. Hamelinck, N. Rosielle, <strong>and</strong> M. Steinbuch. Modular adaptive <strong>deformable</strong> <strong>mirror</strong><br />
technology based on electromagnetic actuators. In 8 th International Workshop on<br />
<strong>Adaptive</strong> Optics for Industry <strong>and</strong> Medicine, Shatura, Russia, 2009.<br />
[87] R. Hamelinck, N. Rosielle, M. Steinbuch, <strong>and</strong> N. Doelman. Electromagnetic dm<br />
technology meets future ao dem<strong>and</strong>s. In AO4ELT, Paris, France, 2009.<br />
[88] R. Hamelinck, N. Rosielle, M. Steinbuch, N. Doelman, R. Ellenbroek, <strong>and</strong> M. Verhaegen.<br />
Actuator grid design for an adaptive <strong>deformable</strong> <strong>mirror</strong>. In Proceedings of<br />
the 7 th EUSPEN International Conference, Bremen, Germany, May 2007.<br />
[89] R. Hamelinck, N. Rosielle, M. Steinbuch, R. Ellenbroek, M. Verhaegen, <strong>and</strong> N. Doelman.<br />
Actuator tests for a large <strong>deformable</strong> membrane <strong>mirror</strong>. In B. L. Ellerbroek <strong>and</strong><br />
D. C. B. Calia, editors, Proceedings of SPIE: Astronomical telescopes <strong>and</strong> instrumentation<br />
- Advances in <strong>Adaptive</strong> Optics, volume 6272, May 2006.<br />
[90] R. Hamelinck, N. Rosielle, M. Steinbuch, R. Ellenbroek, M. Verhaegen, <strong>and</strong> N. Doelman.<br />
Deformable membrane <strong>mirror</strong> with high actuator density <strong>and</strong> distributed <strong>control</strong>.<br />
In 6 th International Workshop on <strong>Adaptive</strong> Optics for Industry <strong>and</strong> Medicine,<br />
Galway, Irel<strong>and</strong>, June 2007.<br />
[91] R. Hamelinck, N. Rosielle, M. Steinbuch, R. Ellenbroek, M. Verhaegen, <strong>and</strong> N. Doelman.<br />
Test results of an adaptive <strong>deformable</strong> <strong>mirror</strong> for future large telescopes. In<br />
Proceedings of the 8 th EUSPEN International Conference, 2008.<br />
[92] P. J. Hampton, R. Conan, O. Keskin, C. Bradley, <strong>and</strong> P. Agathoklis. Selfcharacterization<br />
of linear <strong>and</strong> nonlinear adaptive optics systems. Applied Optics,<br />
47(2):126–134, Januari 2008.<br />
[93] H<strong>and</strong>shake Solutions. Haste-Programming Language Manual, 2007.<br />
[94] D. S. Hansen. Digital <strong>control</strong>ler design for alised sampled-data systems. PhD thesis,<br />
Brigham-Young university, department of mechanical engineering, August 2000.
222 Bibliography<br />
[95] C. M. Harding, R. A. Johnston, <strong>and</strong> R. G. Lane. Fast simulation of a Kolmogorov<br />
phase screen. Applied Optics, 38(11):2161–2170, April 1999.<br />
[96] J. Hardy. Analog data processor. US Patent 3921080, November 1975.<br />
[97] J. Hardy. Instrumental limitations in adaptive optics for astronomy. In Proceedings<br />
of SPIE: Active telescope systems, volume 1114, pages 2–13, November 1989.<br />
[98] J. Hardy. <strong>Adaptive</strong> optics for Astronomical Telescopes. Oxford University Press,<br />
New York, 1998.<br />
[99] T. Henningsson <strong>and</strong> A. Rantzer. Scalable Distributed Kalman Filtering for Mass-<br />
Spring Systems. In Proceedings of the 46 th IEEE Conference on Decision <strong>and</strong> Control,<br />
New Orleans, LA, December 2007.<br />
[100] K. J. Hinnen. Data-Driven Optimal Control for <strong>Adaptive</strong> Optics. Phd dissertation,<br />
Delft University of Technology, January 2007. ISBN 978-90-9021188-6.<br />
[101] P. Hinz, M. Kenworthy, D. Miller, V. Vaitheeswaran, G. Brusa, <strong>and</strong> G. Zappellini.<br />
Riding the hub: characterization of the MMT adaptive secondary performance. In<br />
B. L. Ellerbroek <strong>and</strong> D. C. B. Calia, editors, Proceedings of SPIE: Astronomical<br />
telescopes <strong>and</strong> instrumentation, volume 6272, 2006.<br />
[102] J. Hois <strong>and</strong> S. Fassois. Stochastic vector identification <strong>and</strong> uncertain modal parameter<br />
estimation for a smart composite beam. In Proceedings of the 3 rd international<br />
conference smart materials structures systems, Acireale, Italy, June 2008.<br />
[103] N. Hubin. Technical specification for the Conceptual Design, Prototyping, Preliminary<br />
Design of the M4 adaptive unit for the E-ELT. Technical Report E-ESO-SPE-<br />
106-0037 1.0, European Southern Observatory, 2007.<br />
[104] B. Hulburd <strong>and</strong> D. S<strong>and</strong>ler. Segmented <strong>mirror</strong>s for atmospheric compensation. Optical<br />
Engineering, 29(10):1186, 1990.<br />
[105] R. Irwan. Wavefront estimation in astronomical imaging. PhD thesis, Department of<br />
electrical engineering, University of Canterbury, March 1999.<br />
[106] R. Irwan <strong>and</strong> R. G. Lane. Analysis of optimal centroid estimation applied to Shack-<br />
Hartmann sensing. Applied Optics, 38:6737–6743, November 1999.<br />
[107] S. Jiang, P. G. Voulgaris, L. E. Holloway, <strong>and</strong> L. A. Thompson. Distributed Control of<br />
Large Segmented Telescopes. In Proceedings of 2006 American Control Conference,<br />
Minneapolis, Minnesota, USA, June 2006.<br />
[108] F. Jones. The refractivity of air. Journal of Research of the National Bureau of<br />
St<strong>and</strong>ards, 86:27–32, 1981.<br />
[109] M. Kasper, E. Fedrigo, D. P. Looze, H. Bonnet, L. Ivanescu, <strong>and</strong> S. Oberti. Fast<br />
calibration of high-order adaptive optics systems. Journal of the Optical Society of<br />
America A, 2:1004–1008, 2004.<br />
[110] A. G. Kelkar <strong>and</strong> S. M. Joshi. Dissipativity <strong>and</strong> passivity, chapter 3. Springer Berlin<br />
/ Heidelberg, 1996.<br />
[111] H. Kharaghani <strong>and</strong> B. Tayfeh-Rezaie. A hadamard matrix of order 428. Journal of<br />
Combinatorial Designs, 13:435–440, 2005.<br />
[112] A. Kolmogorov. Dissipation of energy in a locally isotropic turbulence, volume 3<br />
of Turbulence, Classic Papers on Statistical Theory. Wiley-Interscience, New York,<br />
October 1961.
Bibliography 223<br />
[113] M. Langlois, R. Angel, M. Lloyd-Hart, F. W. G. Love, <strong>and</strong> A. Naumov. High Order<br />
Reconstructor-Free <strong>Adaptive</strong> Optics for 6-8 meter class Telescopes. In ESO Conference<br />
& Workshop Proceedings, Beyond Conventional <strong>Adaptive</strong> Optics, volume 58,<br />
pages 113–120, Venice, Italy, 2001.<br />
[114] W. Langlois. Isothermal squeeze films. Quarterly applied mathematics, XX:131–<br />
150, 1962.<br />
[115] H. Lev-Ari. Efficient solution of linear matrix equations with application to multistatic<br />
antenna array processing. Communications in information <strong>and</strong> systems,<br />
5(1):123–130, 2005.<br />
[116] J.-S. Lew, J.-N. Juang, <strong>and</strong> R. Longman. Comparison of several system identification<br />
methods for flexible structures. Journal of Sound <strong>and</strong> Vibration, 167(3):461–480,<br />
November 1993.<br />
[117] Y.-T. Liu <strong>and</strong> S. Gibson. <strong>Adaptive</strong> Optics with <strong>Adaptive</strong> Filtering <strong>and</strong> Control. In<br />
Proceedings of the American Control Conference, pages 3176–3179, Boston, Massachusetts,<br />
USA, June/July 2004.<br />
[118] P. M. Livingston. Proposed method for inner scale measurements in a turbulent atmosphere.<br />
Applied Optics, 11(3):684–687, 1972.<br />
[119] L. Ljung. System identification - theory for the user. Prentice Hall, inc., 1987.<br />
[120] M. Lloyd-Hart, F. Wildi, H. Martin, P. McGuire, M. Kenworthy, R. Johnson, B. Fitz-<br />
Patrick, G. Angeli, S. Miller, <strong>and</strong> R. Angel. <strong>Adaptive</strong> optics for the 6.5m MMT. In<br />
P. L. Wizinowich, editor, Proceedings of SPIE: <strong>Adaptive</strong> Optical Systems Technology,<br />
volume 4007, pages 167–174, 2000.<br />
[121] M. Loktev. Modal wavefront correctors based on nematic liquid crystals. PhD thesis,<br />
Delft University of Technology, 2005.<br />
[122] M. Loktev, D. W. D. L. Monteiro, <strong>and</strong> G. Vdovin. Comparison study of the performance<br />
of piston, thin plate <strong>and</strong> membrane <strong>mirror</strong>s for correction of turbulenceinduced<br />
phase distortions. Optics Communications, 192:91–99, May 2001.<br />
[123] D. P. Looze. Minimum variance <strong>control</strong> structure for adaptive optics systems. In<br />
Journal of the Optical Society of America A, 2005.<br />
[124] D. G. MacMartin. Local, hierarchic, <strong>and</strong> iterative reconstructors for adaptive optics.<br />
Journal of the Optical Society of America, 20:1084–1093, June 2003.<br />
[125] E. M<strong>and</strong>ers. Design of a device to measure the out-of-plane, nonlinear stiffness of<br />
circular membranes. Technical report, Technische Universiteit Eindhoven, 2004.<br />
[126] E. Marchetti, R. Brast, B. Delabre, et al. On-sky testing of the multi-conjugate<br />
adaptive optics demonstrator. The Messenger 129, European Southern Observatory,<br />
September 2007.<br />
[127] F. Martin, R. Conan, A. Tokovinin, A. Ziad, H. Trinquet, J. Borgnino, A. Agabi,<br />
<strong>and</strong> M. Sarazin. Optical parameters relevant for High Angular Resolution at Paranal<br />
from GSM instrument <strong>and</strong> surface layer contribution. Astronomy <strong>and</strong> Astrophysics,<br />
144:39–44, May 2000.<br />
[128] P. Massioni, T. Keviczky, <strong>and</strong> M. Verhaegen. New Approaches to Distributed Control<br />
of Satellite Formation Flying. In Proceedings of the 3 rd International Symposium on<br />
Formation Flying, Missions <strong>and</strong> Technologies, ESA/ESTEC, Noordwijk, The Nederl<strong>and</strong>s,<br />
April 2008.
224 Bibliography<br />
[129] P. Massioni <strong>and</strong> M. Verhaegen. Distributed Control for Identical Dynamically Coupled<br />
Systems: a Decomposition Approach. Submitted paper.<br />
[130] T. Mathworks. Scientific computing software, 2009. http://www.mathworks.com.<br />
[131] Mentor Graphics Corporation. ModelSim User’s Manual, 2007.<br />
[132] D. C. Montgomery <strong>and</strong> G. C. Runger. Applied statistics <strong>and</strong> probability for engineers.<br />
John Wiley <strong>and</strong> Sons, Inc., New York,2 nd edition, 1999.<br />
[133] E. M. Navarro-López. Several dissipativity <strong>and</strong> passivity implications in the linear<br />
discrete-time setting. Mathematical Problems in Engineering, pages 599–616, 2005.<br />
[134] R. Negenborn. Multi-Agent Model Predictive Control with Applications to Power<br />
Networks. PhD thesis, Delft University of Technology, December 2007. ISBN 978-<br />
90-5584-093-9.<br />
[135] M. Nicolle, T. Fusco, G. Rousset, <strong>and</strong> V. Michau. Improvement of Shack-<br />
Hartmann wave-front sensor measurement for extreme adaptive optics. Optics Letters,<br />
29(23):2743–2745, 2004.<br />
[136] R. Noll. Zernike polynomials <strong>and</strong> atmospheric turbulence. Journal of the Optical<br />
Society of America, 66:207–211, March 1976.<br />
[137] R. Olfati-saber <strong>and</strong> R. M. Murray. Distributed cooperative <strong>control</strong> of multiple vehicle<br />
formations using structural potential functions. In IFAC World Congress, 2002.<br />
[138] R. Olifati-Saber. Distributed Kalman Filter with Embedded Consensus Filter. In Proceedings<br />
of the 44 th IEEE Conference on Decision <strong>and</strong> Control 2005, pages 8179–<br />
8184. Dartmouth College, December 2005.<br />
[139] J. Owens. Optical refractive index of air: dependence on pressure, temperature <strong>and</strong><br />
composition. Applied Optics, 6:51–59, 1967.<br />
[140] S. Oya, A. Bouvier, O. Guyon, M. Watanabe, Y. Hayano, H. Takami, M. Iye, M. Hattori,<br />
Y. Saito, M. Itoh, S. Colley, M. Dinkins, M. Eldred, <strong>and</strong> T. Golota. Performance<br />
of <strong>deformable</strong> <strong>mirror</strong> for Subaru LGSAO system. In B. L. Ellerbroek <strong>and</strong> D. C. B.<br />
Calia, editors, Proceedings of SPIE: Astronomical telescopes <strong>and</strong> instrumentation -<br />
Advances in <strong>Adaptive</strong> Optics, volume 6272, 2006.<br />
[141] E. Peck <strong>and</strong> K. Reeder. Dispersion of air. Journal of the Optical Society of America,<br />
62:958–962, 1972.<br />
[142] L. A. Poyneer. Correlation wave-front sensing algorithms for Shack-Hartmann-based<br />
<strong>Adaptive</strong> Optics using a point source. In Lawrence Livermore National Laboratory<br />
(LLNL), number UCRL-JC-152975, May 2003.<br />
[143] L. A. Poyneer, D. T. Gavel, <strong>and</strong> J. M. Brase. Fast wave-front reconstruction in large<br />
adaptive optics systems with use of the Fourier transform. Journal of the Optical<br />
Society of America A, 19(10):2100–2111, October 2002.<br />
[144] R. B. Quirino <strong>and</strong> C. P. Bottura. An approach for distributed Kalman filtering. In<br />
SBA Controle & Automação, volume 12, 2001.<br />
[145] R. Ragazzoni. Pupil plane wavefront sensing with an oscillating prism. Journal of<br />
Modern Optics, 43(2):289–293, 1996.<br />
[146] R. Ragazzoni, E. Marchetti, <strong>and</strong> F. Rigaut. Modal tomography for <strong>Adaptive</strong> Optics.<br />
Astronomy <strong>and</strong> Astrophysics, 342:53–56, 1999.<br />
[147] A. Rantzer. Using Game Theory for Distributed Control Engineering. In Proceedings<br />
of Games 2008, 3rd World Congress of the Game Theory Society, July 2008.
Bibliography 225<br />
[148] F. Roddier. Curvature sensing <strong>and</strong> compensation: a new concept in adaptive optics.<br />
Applied Optics, 27(7):1223–1225, 1988.<br />
[149] F. Roddier. <strong>Adaptive</strong> optics in astronomy. Cambridge University Press, 1999.<br />
[150] B. L. Roux <strong>and</strong> J.-M. Conan. Optimal <strong>control</strong> law for classical <strong>and</strong> multiconjugate<br />
adaptive optics. Journal of the Optical Society of America A, 21(7):1261–1276, July<br />
2004.<br />
[151] T. Ruppel, M. Lloyd-Hart, D. Zanotti, <strong>and</strong> O. Sawodny. Modal Trajectory Generation<br />
for <strong>Adaptive</strong> Secondary Mirrors in Astronomical <strong>Adaptive</strong> Optics. IEEE International<br />
Conference on Automation Science <strong>and</strong> Engineering, pages 430–435, 2007.<br />
[152] Y. Saad. Iterative Methods for Sparse Linear Systems. PWS, 1996.<br />
[153] D. Saint-Jacques <strong>and</strong> J. E. Baldwin. Taylor’s hypothesis: good for nuts. In P. J.<br />
Lena <strong>and</strong> A. Quirrenbach, editors, Proceedings of SPIE: Interferometry in Optical<br />
Astronomy, volume 4006, pages 951–962, 2000.<br />
[154] T. Sakai <strong>and</strong> H. Kawamoto. Improvement in Time-Dependent Displacement Degradation<br />
of Piezoelectrc Ceramics by Manganese/Indium Addition 1427. Journal of<br />
the American Ceramic Society, 83:1423–1427, 2000.<br />
[155] P. Salinari, C. D. Vecchio, <strong>and</strong> V. Biliotti. A study of an adaptive secondary <strong>mirror</strong>. In<br />
F. Merkle, editor, ICO-16 Satellite conference on Active <strong>and</strong> <strong>Adaptive</strong> Optics, 1993.<br />
[156] V. Samarkin, A. Aleks<strong>and</strong>rov, <strong>and</strong> V. Dubikovsky. Water-cooled wavefront correctors.<br />
Advanced Optoelectronics <strong>and</strong> Lasers, 1:119, 2005.<br />
[157] J. Sauvage, T. Fusco, G. Rousset, <strong>and</strong> C. Petit. Calibration <strong>and</strong> precompensation<br />
of noncommon path aberrations for extreme adaptive optics. Journal of the Optical<br />
Society of America A, 24:2334–2346, 2007.<br />
[158] R. Schreier <strong>and</strong> G. Temes. Delta-Sigma data converters: theory, design <strong>and</strong> simulation.<br />
IEEE Press, 1997.<br />
[159] T. N. Science <strong>and</strong> T. R. C. for Computation <strong>and</strong> Visualization of Geometric Structures.<br />
Qhull software, 2009. http://www.qhull.org.<br />
[160] R. V. Shack <strong>and</strong> B. C. Platt. Production <strong>and</strong> use of a lenticular hartmann screen.<br />
Journal of the Optical Society of America, 61:656, 1971.<br />
[161] L. S. Shieh <strong>and</strong> Y. T. Tsay. Transformations of a class of multivariable <strong>control</strong><br />
systems to block companion forms. IEEE Transactions on Automatic Control,<br />
27(1):199–203, February 1982.<br />
[162] E. Silva, G. Goodwin, D. Quevedo, <strong>and</strong> M. Derpich. Optimal Noise Shaping for<br />
Networked Control Systems. In Proceedings of the European Control Conference,<br />
Kos, Greece, 2007.<br />
[163] J. R. Sims, A. N. Durney, <strong>and</strong> C. C. Smith. Design of Mechatronic Systems with<br />
Aliased Plant Modes. IEEE/ASME transactions on mechatronics, 3:144–149, 1998.<br />
[164] J. Sinquin, J. Lurçon, <strong>and</strong> C. Guillemard. Deformable <strong>mirror</strong>s technologies for astronomy<br />
at CILAS. In N. Hubin, C. E. Max, <strong>and</strong> P. L. Wizinowich, editors, Proceedings<br />
of SPIE: <strong>Adaptive</strong> Optics Systems, volume 7015, 2008.<br />
[165] H. Song, R. Fraanje, M. Verhaegen, <strong>and</strong> G. Vdovin. Hysteresis compensation for<br />
piezo <strong>deformable</strong> <strong>mirror</strong>. In Proceedings of the6 th International Workshop on <strong>Adaptive</strong><br />
Optics for Industry <strong>and</strong> Medicine, 2007.
226 Bibliography<br />
[166] D. P. Spanos, R. Olfati-Saber, <strong>and</strong> R. M. Murray. Approximate distributed kalman<br />
filtering in sensor networks with quantifiable performance. In Proceedings of the4 th<br />
international symposium on Informationprocessing in sensor networks. IEEE Press,<br />
2005.<br />
[167] T. Stalcup <strong>and</strong> K. Powell. Image motion correction using accelerometers at the MMT<br />
observatory,. In Proceedings of SPIE, volume 7018, 2008.<br />
[168] G. Stein <strong>and</strong> D. Gorinevsky. Design of surface shape <strong>control</strong> for large twodimensional<br />
arrays. IEEE Transactions on Control Systems Technology, 13, 2005.<br />
[169] G. Stewart, D. Gorinevsky, <strong>and</strong> G. Dumont. Feedback <strong>control</strong>ler design for a spatially<br />
distributed system: the paper machine problem. IEEE Transactions on Control<br />
Systems Technology, 11:612–628, September 2003.<br />
[170] P. Stoica <strong>and</strong> M. Jansson. MIMO system identification: state-space <strong>and</strong> subspace approximations<br />
versus transfer function <strong>and</strong> instrumental variables. IEEE Transactions<br />
on Signal Processing, 48:3087–3099, 2000.<br />
[171] K. Szeto, S. Roberts, S. Sun, L. Stepp, J. Nelson, M. Gedig, N. Loewen, <strong>and</strong><br />
D. Tsang. TMT telescope structure system - design <strong>and</strong> development. In L. M.<br />
Stepp, editor, Proceedings of SPIE: Ground-based <strong>and</strong> airborne telescopes, volume<br />
6267, May 2006.<br />
[172] H. Takami, S. Colley, M. Dinkins, M. Eldred, O. Guyon, T. Golot, M. Hattori,<br />
Y. Hayano, M. Ito, M. Iye, S. Oya, Y. Saito, <strong>and</strong> M. Watanabe. Status of Subaru<br />
Laser Guide Star AO System. In B. L. Ellerbroek <strong>and</strong> D. C. B. Calia, editors,<br />
Proceedings of SPIE: Astronomical telescopes <strong>and</strong> instrumentation - Advances in<br />
<strong>Adaptive</strong> Optics, volume 6272, 2006.<br />
[173] V. Tatarski. Wavefront Propagation in a Turbulent Medium. McGraw-Hill, 1961.<br />
[174] R.F.M.M. Hamelinck. <strong>Adaptive</strong> <strong>deformable</strong> <strong>mirror</strong> based on electromagnetic actuators.<br />
PhD thesis, Technische Universiteit Eindhoven, 2010. ISBN 978-90-386-2278-<br />
1.<br />
[175] S. Thomas. Optimized centroid computing in a Shack-Hartmann sensor. In D. B.<br />
Calia, B. L. Ellerbroek, <strong>and</strong> R. Ragazzoni, editors, Proceedings of SPIE: Advancements<br />
in <strong>Adaptive</strong> Optics., volume 5490, pages 1238–1246, 2004.<br />
[176] M. Tillerson, L. Breger, <strong>and</strong> J. P. How. Distributed Coordination <strong>and</strong> Control of<br />
Formation Flying Spacecraft. In Proceedings of the American Control Conference,<br />
2003.<br />
[177] S. Timoshenko <strong>and</strong> S. Woinowsky-Krieger. Theory of plates <strong>and</strong> shells. McGraw-<br />
Hill, Auckl<strong>and</strong>, second edition edition, 1989.<br />
[178] A. A. Tokovinin, B. Gregory, H. E. Schwarz, V. Terebizh, <strong>and</strong> S. Thomas. A visiblelight<br />
AO system for the 4.2-m SOAR telescope. In P. L. Wizinowich <strong>and</strong> D. Bonaccini,<br />
editors, Proceedings of SPIE: <strong>Adaptive</strong> Optical System Technologies II, volume<br />
4839, pages 673–680, 2003.<br />
[179] R. K. Tyson. Principles of adaptive optics. Academic Press, 1998.<br />
[180] G. Vdovin <strong>and</strong> P. Sarro. Flexible <strong>mirror</strong> micro-machined in silicon. Applied Optics,<br />
34(16):2968–2972, 1995.<br />
[181] C. D. Vecchio. Aluminium reference plate, heat sink, <strong>and</strong> actuator design for an<br />
adaptive secondary <strong>mirror</strong>. In D. Bonaccini <strong>and</strong> R. K. Tyson, editors, Proceedings of<br />
SPIE: <strong>Adaptive</strong> Optical System Technologies, volume 3353, pages 839–849, 1998.
Bibliography 227<br />
[182] C. D. Vecchio, W. Gallieni, P. Salinari, <strong>and</strong> P. Gray. Preliminary mechanical design<br />
of an adaptive secondary unit for the MMT-conversion telescope. In M. Cullum,<br />
editor, Proceedings of ESO, volume 54, page 243, 1996.<br />
[183] M. Verhaegen. Subspace model identification part 3. analysis of the ordinary outputerror<br />
state-space model identification algorithm. International Journal of Control,<br />
56:555–586, 1993.<br />
[184] M. Verhaegen. Identification of the deterministic part of state-space models in innovations<br />
form from input-output data. Automatica, 30(1):61–74, 1994.<br />
[185] M. Verhaegen <strong>and</strong> P. Dewilde. Subspace model identification part I - The outputerror<br />
state-space model identification class of algorithms. International Journal of<br />
Control, 56(5):1187–1210, 1992.<br />
[186] M. Verhaegen <strong>and</strong> V. Verdult. Filtering <strong>and</strong> system identification. Cambridge university<br />
press, New York, 1st edition, 2007.<br />
[187] C. Vogel. Sparse matrix methods for wavefront reconstruction revisited. In B. L.<br />
Ellerbroek <strong>and</strong> R. Ragazzoni, editors, Proceedings of SPIE: Advancements in <strong>Adaptive</strong><br />
Optics, volume 5490, pages 1327–1335, 2004.<br />
[188] C. R. Vogel <strong>and</strong> Q. Yang. Fast optimal wavefront reconstruction for multi-conjugate<br />
adaptive optics using the fourier domain preconditioned conjugate gradient algorithm.<br />
Optics Express, 14:7487–7498, 2006.<br />
[189] C. R. Vogel <strong>and</strong> Q. Yang. Multigrid algorithm for least-squares wavefront reconstruction.<br />
Applied Optics, 45(4):705–715, 2006.<br />
[190] M. Vorontsov, G. Izakson, A. Kudryashov, G. Kosheleva, S. Nazarkin, Y. F. Suslov,<br />
<strong>and</strong> V. I. Shmalgauzen. <strong>Adaptive</strong> cooled <strong>mirror</strong> for the resonator of an industrial<br />
laser. Soviet Journal of Quantum Electronics, 15(7):888, 1985.<br />
[191] D. M. Wiberg, C. E. Max, <strong>and</strong> D. T. Gavel. Geometric view of adaptive optics <strong>control</strong>.<br />
Journal of the Optical Society of America A, 22(5):870–880, May 2005.<br />
[192] W. J. Wild, E. J. Kibblewhite, <strong>and</strong> R. Vuilleumier. Sparse matrix wave-front estimators<br />
for adaptive-optics systems for largeground-based telescopes. Optics Letters,<br />
20(9):955–957, May 1995.<br />
[193] F. Wildi, G. Brusa, A. Riccardi, R. Allen, M. Lloyd-Hart, D. Miller, B. Martin, R. Biasi,<br />
<strong>and</strong> D. Gallieni. Progress of the MMT adaptive optics program. In R. K. Tyson,<br />
D. Bonaccini, <strong>and</strong> M. C. Roggemann, editors, Proceedings of SPIE, volume 4494,<br />
pages 11–18, 2001.<br />
[194] P. Wizinowich. <strong>Adaptive</strong> optics at the keck oservatory. IEEE Instrumentation <strong>and</strong><br />
measurement magazine, 2005.<br />
[195] Xilinx Inc. Xilinx ISE 9.1i Software Manuals <strong>and</strong> Help.<br />
[196] Q. Yang, C. R. Vogel, <strong>and</strong> B. L. Ellerbroek. Fourier domain preconditioned conjugate<br />
gradient algorithm for atmospheric tomography. Applied Optics, 45(21):5281–5293,<br />
2006.<br />
[197] D. Young. Iterative solutions for large linear systems. Academic Press, New York,<br />
1971.<br />
[198] A. Zadrozny, M. Chang, D. Buscher, R. Myers, A. Doel, C. Dunlop, R. Sharples, <strong>and</strong><br />
R. Arnold. First Atmospheric Compensation With a Linearised High Order <strong>Adaptive</strong><br />
Mirror -ELECTRA. In E. D. Bonaccini, editor, ESO Conference Workshop Proceedings,<br />
number 56, pages 459–468. ESO, 1999.
228 Bibliography<br />
[199] K. Zhou <strong>and</strong> J. C. Doyle. Essentials of robust <strong>control</strong>. Prentice Hall, 1998.<br />
[200] Y. Zou, P. R. Pagilla, <strong>and</strong> R. T. Ratcliff. Distributed Formation Flight Control Using<br />
Constraint Forces. Journal of guidance, <strong>control</strong>, <strong>and</strong> <strong>dynamics</strong>, 32(1), Januari 2009.
ÙÖÖÙÐÙÑÚØ<br />
Rogier Ellenbroek was born on Januari 28th 1979 in Heerlen (the Netherl<strong>and</strong>s). He studied<br />
Mechanical Engineering at the Technische Universiteit Eindhoven from 1997 to 2003, where<br />
he completed his master’s degree with a project in collaboration with Philips Centre for<br />
Applied Technologies (CFT) titled time-frequency adaptive iterative learning <strong>control</strong> with<br />
application to a wafer stage. After graduating, he worked for a year at the department of<br />
mechanical engineering as a visiting scientist, a.o. assisting in a PhD project on optimizing<br />
vehicle fuel economy through intelligent battery charging.<br />
In 2004 he started this PhD project on adaptive optics at the Delft University of Technology.<br />
As of January 2009 he is employed by Mapper Lithography in Delft, where he currently<br />
works as a senior systems engineer.<br />
229