Autoencoder Neural Networks for LPV Embedding of Nonlinear Systems

Abstract

In this paper, the problem of automated generation of linear parameter-varying (LPV) state-space models is addressed. A deep neural network (DNN) is developed to embed the dynamical behavior of a nonlinear (NL) system into an LPV model with predefined number of scheduling variables which are the NL functions of the states. Leveraging the Autoencoder (AE) neural networks (NN) and using the input-output plant data, a scheduling NL mapping is defined. The developed LPV model depends affinely on the scheduling variables. Since the proposed method to derive LPV model is based on input-output plant data, the explicit NL equations of the plant are not required. The upper and lower bounds on the scheduling variables can be computed by solving convex optimization problems. The effectiveness of the proposed method is evaluated on a benchmark example.