Chapter 5 Robust Performance Tailoring with Tuning - SSL - MIT
Chapter 5 Robust Performance Tailoring with Tuning - SSL - MIT
Chapter 5 Robust Performance Tailoring with Tuning - SSL - MIT
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
3.1 Uncertainty<br />
Historically, structural models are used along <strong>with</strong> test data to help engineers better<br />
understand the physical behavior of the hardware. Model predictions are compared<br />
to test data, and the data is used to provide deeper understanding of the underlying<br />
physics and validate the model for use in trade analyses. However, as space systems<br />
become more complex and more difficult to test on the ground, test data is harder<br />
to obtain and it is necessary to rely solely on models and simulations to provide<br />
predictions of future system behavior <strong>with</strong>out the benefit of validation data. This<br />
task is a much more demanding one and requires a high level of confidence in the<br />
models. As a result, much attention is focused in the area of prediction accuracy and<br />
model uncertainty.<br />
The sources of inaccuracy in model predictions are grouped into three main cate-<br />
gories: parametric errors, discretization errors and model structure errors [27, 93, 12].<br />
Parametric errors refer to inaccuracies in the values of model parameters. For exam-<br />
ple, a finite element model may consist of beam elements that have certain material<br />
properties, such as Young’s Modulus. Although the value of Young’s Modulus is<br />
published for a wide range of materials it is likely that components made from the<br />
same material have slightly different Young’s Modulus values. A value of Young’s<br />
modulus in a finite element model that is slightly different from the Young’s modulus<br />
of the physical component is an example of a parametric error. Discretization errors<br />
exist because finite element models are composed of discrete elements while physical<br />
parts are continuous. These errors can be reduced by using a high-fidelity mesh,<br />
but not eliminated. Finally, model structure errors are global modeling omissions or<br />
mistakes. This category includes any physical system behavior that is not captured<br />
in the model. Examples of these types of errors are unmodelled nonlinearities and<br />
improper choice of element types.<br />
Parametric errors are most often considered in stochastic analysis because they are<br />
the easiest to model and the hardest to reduce. Discretization error can be reduced<br />
by refining the finite element model mesh, and model structure errors are reduced<br />
72