Autoencoder-Inspired Identification of LTI Systems

Tobias Nagel, Marco Huber

Abstract: Identifying the state space representation of a dynamical system during usage enables a controller to adapt quickly in a changing environment. We propose a new method for identifying linear time-invariant (LTI) systems online based on the measurement of input-output data. Therefore, we implement the calculation of a system response in a machine learning framework and use an autoencoder-related approach to find a neural network which performs a system identification by one single forward pass. This is computationally efficient and can be performed online during usage. We validate the approach by identifying the wear of a robot leg.