Dynamic neural network switching for active control of nonlinear systems.

Abstract

Feedforward active noise and vibration control systems have been developed for many applications, but are generally designed using linear digital filters, most typically implementing the filtered reference least mean squares algorithm. When the system under control exhibits nonlinearities, linear controllers cannot fully capture the system dynamics to maximize performance. Previous work has shown that neural network (NN) based controllers can improve control performance in the presence of nonlinearities. However, inferring the outputs of NN controllers can be computationally expensive, limiting their practicality, particularly when control is required across a range of nonlinear behaviors. In this paper, a control strategy is proposed where performance is maintained across a nonlinear range of operation by dynamically switching between a set of smaller, and therefore more efficient, NNs that are individually trained over specific ranges of the nonlinear system behavior. It is demonstrated via both simulations of a system with a simple nonlinear stiffness in the primary path and offline simulations using a physical nonlinear dynamical system in the primary path, that the performance of the proposed switching approach offers a control performance advantage compared to both a larger generalized individual NN controller and a functional link artificial neural network based controller.