Reinforcement learning algorithm for secondary path identification in active noise control systems

To address the issue of inaccurate secondary path modeling and the challenges of handling nonlinear and time-varying characteristics in traditional active noise control systems, this study proposes a secondary path identification method based on reinforcement learning (RL) and periodic prediction mechanisms. First, the modeling performance of different strategies in complex dynamic systems was analyzed, confirming the advantages of RL methods. Based on this analysis, a policy network with shared self-attention mechanisms and periodic prediction bypass structures was constructed to enhance the model’s ability to capture periodic features and improve the modeling accuracy and training stability. Real-vehicle test results show that this method significantly outperforms traditional algorithms, reducing the average error by 37.2%, increasing the signal-to-noise ratio by 5.1 dB, reducing the relative error by 12.6%, and improving the overall noise reduction by 6.8 dB.