Active Noise Control (ANC) technology is of great value in the field of noise mitigation. Recently, traditional linear adaptive control methods, represented by the FxLMS algorithm, are structurally simple and computationally efficient but often suffer from performance degradation or even failure in practical applications due to nonlinear system factors. For this reason, neural network-based ANC methods have attracted significant research interest for their strong nonlinear processing capabilities and have gradually emerged as a focal point for addressing nonlinear ANC problems. This paper systematically reviews the research progress of neural networks in the field of nonlinear ANC, focusing on two key dimensions: network architecture and training methods. In terms of architecture design, existing studies primarily enhance performance through topology optimization, improvements to functional link artificial neural networks, and innovative hidden layer designs. Advancements in training methods focus on the optimization of loss functions, innovation in weight update algorithms, and the introduction of other training strategies. In the future, neural network-based ANC algorithms will continue to deepen, with potential development paths including the integration of advanced network architectures such as Generative Adversarial Networks (GANs), optimization of utility functions, pruning of hidden layers, improvement in loss function design, and the adoption of more efficient training strategies. These efforts will further improve algorithm performance and ultimately provide robust support for achieving more precise and efficient active noise control.
Liu et al. (Wed,) studied this question.