Comparison of neural network architectures for feedforward active control of nonlinear systems

Abstract

In recent decades, advances in digital technologies have allowed for the development of increasingly complex active control solutions for both noise and vibration, which have been utilised in a wide range of applications from automotive to maritime. Such control systems commonly use linear digital filters that cannot fully capture the dynamics of the nonlinear systems under control. To overcome such issues, it has previously been demonstrated that replacing the conventional linear controllers with Neural Networks (NNs) can improve control performance in the presence of nonlinearities in both the system plant and primary path. However, a key design decision in the implementation of such controllers is the choice of an appropriate NN architecture for the specific control application. In this paper, a method of training NN controllers for Single-Input-Single-Output (SISO) acoustic noise control is presented. Multilayer Perceptron (MLP), Elman Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) architectures are implemented as controllers in the presence of nonlinearities in the primary path. An analysis is presented of the relative control performance of the different architectures, and the computational cost in operation and network training is discussed.

More information

Identifiers

ORCID for Jordan Cheer: orcid.org/0000-0002-0552-5506

Catalogue record

Export record