Home > Research Groups > Junior Research Group DART > Research > Hybrid Modeling and Observer Design

Hybrid Modeling and Observer Design

In this focus, concepts for hybrid modelling types are researched under control engineering aspects. In many cases, only one subsystem of a plant to be controlled has a very precisely known, usually physical model, while no model exists for the rest of the subsystem. For example, in the case of fatigue strength investigations of passenger car axles, the test bench itself is modelled very accurately, while no models are usually available for the changing axles as test specimens. For test bench control, however, knowledge of the dynamic axle behaviour would be very desirable, as the test bench could then be operated with much higher performance.

For such scenarios we wish to develop a data-efficient method for identifying hybrid models. For this purpose, an existing physical part model is to be complemented by the residual dynamics that are not mapped in the model. These residual dynamics are to be learned e.g. with Bayesian optimisation in interaction with the real overall system. It will be investigated, for example, to what extent prior physical knowledge can support and improve the data-driven learning process in a targeted manner, e.g. in the form of physically based learning functions. Further aspects include ensuring robustness properties of the hybrid model.

Additonally, further work in this focus deals with the automatic parameterisation of observers in combination with machine learning methods as well as with the synthesis of robust hybrid sliding mode observers. For state control, an observer is usually indispensable, because normally not all state variables are measured, but are needed in state control. Again, we will assume that the dynamics of a subsystem have already been modelled well physically, but the rest of the subsystem is not or only rudimentarily known. In order to design an observer for this system, a data-driven approach for the automatic design of the simulator, which is an as accurate as possible (internal) model of the controlled system, and of the corrector, which provides a stabilising feedback of the deviations between the real measured and the estimated measured variables, shall be developed.