A deep learning-based approach for hybrid nonlinear systems dynamics approximation

An increasingly important class of nonlinear systems includes the nonaffine hybrid systems, in particular those in which the underlying dynamics explicitly depends on a switching signal. When the inherent complexity is treatable and the phenomena governing the system dynamics are known an implicit model can be derived to describe its behaviour over time. When these assumptions are not met the system dynamics can still be approximated by regression-based techniques, provided datasets comprising input/output signals collected from the system are available. One approach relies on intelligent computing-based frameworks, in which artificial neural networks stand out as a class of universal approximation models. This paper, proposes a new approach for capturing nonlinear hybrid system dynamics based on 1D spatio-temporal convolutional neural networks, in which the inputs are represented by regressors and structural configuration parameters. The proposed deep neural network architecture is compared against a shallow multilayer layer perceptron framework, in which each structural configuration is independently approximated. Experimental results point out to the superiority of the 1D spatio-temporal convolutional network.