Experiment Design for the Identification of locally affine multi models

Dipl.-Ing. Matthias Gringard

Since September 2014

German Research Foundation (DFG)

As an alternative to physical modeling, the methods of system identification are used at an increasing rate. The quality of the identified models directly depends on the properties of the used data. In general, identified models cannot be used for extrapolation, therefore technical systems have to be excited in a way that the identification data contains all relevant information about the system dynamics. The test signal design is a sub-problem of the experiment design. The major difference in design strategies is whether the experiment design makes use of the model equations or not. Without using model equations it is possible to design the input signals with respect to external measures such as the co-domain-coverage as well as the energy injected into the system. Frequency domain designs are especially useful for experiment designs for linear systems since the frequency response fully describes linear systems. The model equations can be used in the experiment design to minimize the parameter estimation uncertainty. Since the model parameters enter the model equations of the used multi models in a nonlinear way, the objective function contains these unknown parameters. Therefore in a first step non-optimal process-model-free approaches are investigated, since they don't require a priori knowledge. In a second step, optimal process-model-based approaches are investigated. It is assumed that the initially estimated parameters already are in a neighbourhood of the true parameters. The methods are applied on mechatronic systems.