Physics based machine learning for mechanical system identification

Model parameter and system identification is a key aspect in many mechanical engineering applications. The availability of accurately updated models with respect to experimental data is key for a range of methods from model based control to model based condition monitoring. However, currently most work in this aspect focuses on either performing expensive experiments to update certain parameters in physics based models, for example in modal identification for parameter updating, or by using simple engineering models, or single layer neural networks, which are trained on a limited set of measurements. Recently, a novel approach has been proposed which combines reduced order models with operational measurements to estimate parameters for a physics based model in a convenient low-cost framework. However, this approach was only a first step for linear structural dynamics and still required a rather fixed model structure, which does not always extend well to industrial applications. On the other side of the spectrum, pure data-driven machine learning methods have received a lot of attention over the past decade due to the development of very efficient evaluation and training algorithms. Unfortunately many of the developed general purpose methods do not translate well to dynamic mechanical engineering applications due to the relative lack of data to reliably train the (e.g. recurrent neural network) models. Moreover, most of the existing developments do not focus on a full system state reconstruction, but only a limited feature extraction. For mechanical engineering problems, the full state of the system is of interest, for example, to evaluate the stress distribution. However, recently approaches have been introduced to incorporate physical model information in the machine learning model setup, which enables to train a model with considerably less data and to reconstruct the full model state. The scope of this research projects is to extend these physical model incorporation frameworks to specific linear and nonlinear mechanical system models. This will allow to obtain a faster, more complex and more reliable identification of high-fidelity mechanical system models and will hence enable a range of new developments in design, condition monitoring and control.