Noise Control Using Neural Networks with Dropout

Abstract

This paper describes the noise control for a moving evaluation point using neural networks by making the best use of its learning ability. Noise control is a technology which is effective on low-frequency noise. Based on the principle of superposition, a primary sound wave can be cancelled at an evaluation point by emitting a secondary opposite sound wave. To obtain good control performance, it is important to precisely identify the characteristics of all the sound paths. One of the most popular algorithms of noise control is filtered-x LMS algorithm. This algorithm can deliver a good result while all the sound paths do not change. However, the control system becomes uncontrollable while the evaluation point is moving. To solve the problem, the characteristics of all the paths are required to be identified at all time. In this paper, we applied neural networks with the learning ability to the noise control system to follow the time-varying paths and verified its control performance by numerical simulations. Then, dropout technique for the networks was also applied. Dropout is a technique that prevents the network from overfitting and that realizes better control performance. We concluded that applying dropout to noise control prevents the system from diverging.