Generalised performance of neural network controllers for feedforward active noise control of nonlinear systems

Advances in digital technologies have allowed for the development of complex active noise and vibration control solutions that have been utilised in a wide range of applications. Such control systems are commonly designed using linear filters, which cannot fully capture the dynamics of nonlinear systems. To overcome such issues, it has been shown that replacing linear controllers with Neural Networks (NNs) can improve control performance in the presence of nonlinearities. Many real systems are subject to non-stationary disturbances where the magnitude of the system excitation time dependent. However, within the literature, the performance of single NN controllers across different excitation levels has not been thoroughly explored. In this paper, a method of training Multilayer Perceptrons (MLPs) for single-input-single-output (SISO) feedforward acoustic noise control is presented. In a simple time-discrete simulation, the performance of the trained NNs is investigated for different excitation levels. The effects of the properties of the training data and NN controller on generalised performance are explored. It is demonstrated that the generalised control performance of the MLP controllers falls as the range of magnitudes included in the training data is increased, and that this performance can be recovered by increasing the number of hidden nodes within the controller.