Time domain neural networks for path decoupling in active noise control

Decentralized active noise control (ANC) systems face two fundamental challenges: (1) performance degradation due to crosstalk interference between multichannel secondary sources and error microphones, and (2) additional modeling errors introduced by reference signal prefiltering in filtered-x (Fx) algorithms. This study first presents an analysis showing how these factors impair control performance in Fx-type algorithms. To address these limitations, we propose a decoupling networks LMS (DecNet-LMS) algorithm that combines fixed-value neural networks with adaptive filtering strategies. The solution features: (i) online LMS-based modeling of primary paths and (ii) a light neural network (DecNet) for simultaneous secondary path inversion and crosstalk decoupling. The proposed algorithm demonstrates superior performance across various acoustic conditions in simulations using measured transfer functions. Comparative results show significant improvements over conventional adaptive filters and existing neural network-based approaches with fixed coefficients.