A Multi-Channel Decoupling Method for Feedforward Active Noise Control Using Neural Networks

In traditional multi-channel active noise control (ANC) systems, crosstalk between secondary sources and error microphones significantly degrades control accuracy as well as system stability. The non-linearity in acoustic devices further exacerbates the situation. To solve these problems, a multi-channel decoupling method combining neural networks and adaptive strategies is proposed for the feedforward multi-channel ANC system. The method consists of both fixed-parameter and adaptive parts. The fixed-parameter filter decouples and models the inverse of secondary paths using deep neural networks (DNNs), which makes it possible to use the more efficient decentralized control strategy without considering the crosstalk. Moreover, instead of using frequency domain block processing that usually causes latency, the NN is implemented in time domain, ensuring the causality of the control system. Simulation results show that the proposed method outperforms traditional decentralized control algorithms for both linear and nonlinear acoustic systems.