(a)(f)(b)(g)(c)(d)(e)(h)Fig. 3. Performance of the proposed i<strong>de</strong>ntification metho<strong>do</strong>logy for differentinitial seeds of stator parameters used by the EKF. (a) Generated electricalrotor speed by the vector control scheme, (b) estimated magnitu<strong>de</strong> of rotorflux, (c) estimated state x 3 - scaled inverse of rotor time constant, (d)estimated state x 4 - scaled magnetizing inductance, (e) estimated parameter θ 1- scaled stator transient inductance, (f) estimated parameter θ 2 - scaled statorresistance, (g) simulated/generated and estimated rotor flux q component and(h) generated stator current q component and the simulated one using theonline estimated parameters.One of the main difficulties of any full parameter and fluxestimator is the start-up procedure if no information isavailable beforehand about the electrical parameters. It iseven not possible to <strong>do</strong> it, since we search for too muchinformation as fluxes and parameters, based on just statorsignals and rotor speed and, on the other hand, the mo<strong>de</strong>lsensibility in relation to the stator parameters is low andstrongly <strong>de</strong>pen<strong>de</strong>nt of the dynamic conditions during thestart-up and immediately after. Due to this, a robust solutionis proposed here. We only need to have rough initial values ofthe electrical parameters for scaling the state vector x in (3)and parameter vector θ in (4). Many solutions can be used forthis purpose, some of them being very simple, such as452
classical methods or other rough parameter estimations as in[2], based on nominal characteristics of the induction motor.Within the estimation algorithms themselves, the onlyrequirement with respect to initial values of the electricalparameters is related to the initialization of stator parameters.Since in full estimation the rotor flux components are jointlyestimated with the electrical parameters, the algorithmshardly converge to correct values because of the lack ofconfi<strong>de</strong>nce in estimated flux components at the beginningand, on the other hand, the lack of sensibility of the inductionmotor mo<strong>de</strong>l with respect to stator parameters withoutspecific supply conditions.By this way, the EKF is started without any assumedknowledge of rotor parameters, but with initial seeds forstator parameters. Thus, 1 second after the EKF has beenstarted, when the rotor parameters and flux are supposed tobe close to the right values, the RPEM algorithm(s) is(are)started with rough estimates or even zero values in theparameter vector θ as used in (13). One second later allparameters and flux components are supposed to beconverging to their real values and the algorithms startworking according to the proposed i<strong>de</strong>ntificationmetho<strong>do</strong>logy.The initial values used in the algorithms are nowpresented. For the EKF, the initial state, the state covariancematrix and the system and measurement noise covariancematrices are, respectively:Rs(0) = Rx(0)=P(0)=Rs= diagR = 10ms− 50%;'s'sL (0) = L− 50%[ 0 0 0.1 0.1]diag[ 1e− 5 1e− 5 1e− 4 1e− 4][ 1e− 8 1e− 8 1e− 9 1e− 9]The RPEM algorithms were initialized as follows:θ (0) = 01P (0) = 1e− 3θ1RRs,θ1m,θ1= 1e− 7= 10θ (0) = 0P2θ2RR(0) = 1e− 3s,θ2m,θ2= 1e− 8= 10(12)(13)and they were based on the Kalman filter approach but otherapproaches can be used like forgetting factor (recursive leastsquares), normalized and unnormalized gradient approachesamong other recursive prediction error methods that can befound in [8] and [9].The scaling factors k 1 to k 6 referred in (3) and (4) were,respectively, 1, 1, 0.2, 5, 100 and 0.5.The simulation results shown in fig. 3 were obtained withthe configuration of fig. 2(a) and a square speed reference inthe range -600 to +600rpm, in or<strong>de</strong>r to ensure persistentenough excitation of the stator transient inductance from thetransient conditions, and stator resistance from the low speedzones, where the term R s i s becomes important in the virtualmeasured output of the output equation expressed by (6).In or<strong>de</strong>r to prove the robustness of the proposedmetho<strong>do</strong>logy with respect to initial seeds of stator parametersin the EKF based algorithm, three runs were achieved withdifferent initial seeds: In the first one the initial seeds werema<strong>de</strong> equal to the parameters’ real values and in the secondand third they were ma<strong>de</strong> equal to 50% below and above ofits real values, respectively. The results are presented infigures 3(b) to 3(f). The steady-state errors in all estimatedparameters are less than 2%, and the estimated flux waveformvery closely matches the simulated one as shown in fig. 3(g).Also the simulated stator current matches the “measured” one(generated by the vector control scheme) using the onlineestimated parameters, as can be seen in fig. 3(h).V. EXPERIMENTAL RESULTSExperimental results, shown in fig. 4, proved thefeasibility and applicability of the metho<strong>do</strong>logy <strong>de</strong>scribed inthe previous sections, and the results of a practicalexperiment were recor<strong>de</strong>d to validate the proposed strategy,and confirm the simulation results presented in the abovesection. In the practical experiment the stator voltages andstator currents, expressed in the stator reference frame, havebeen sampled at 5 kHz. Elliptic low-pass pre-filters of fifthor<strong>de</strong>r with a 500Hz cutoff frequency have been used on thesamples. Specific hardware was <strong>de</strong>veloped that has thesignals available in the range of ±10V in both rotor and statorreference frames, by using the AD2S100 analogue processor,but only the stator-referred signals are used for this work. Thestator voltages and currents are acquired in the statorreference frame with a data acquisition system that consistsof the dSPACE <strong>de</strong>velopment system, ACE Kit 1103, base<strong>do</strong>n the DS1103 PPC controller board, the Real Time Interface(RTI) blockset for Simulink as well as experiment software(ControlDesk, MLIB/MTRACE). The dSPACE <strong>de</strong>velopmentplatform was used for the real time i<strong>de</strong>ntification task and notfor control.The rotor position and speed, nee<strong>de</strong>d for reference frameconversions, are obtained via an incremental enco<strong>de</strong>r. For theEKF the stator voltages and currents are converted into therotor reference frame and the estimated rotor flux dqcomponents are converted into the stator reference frame.<strong>Digital</strong> filters, with the same specifications as the analogones, were used for filtering the computed rotor speed fromrotor angle as well as electrical signals after being acquired.Unfortunately, both filters introduce <strong>de</strong>lays. In or<strong>de</strong>r tosynchronize all the signals, the rotor angle and filtered speedhad to be suitably <strong>de</strong>layed. A 2.2KW induction motorcontrolled by an industrial frequency converter from ABB hasbeen used. The motor was loa<strong>de</strong>d by a pow<strong>de</strong>r break whichwas programmed for 12Nm, about half of the nominal torque.The vector control scheme <strong>de</strong>veloped by us in Simulink453