An adaptive deep neural network for active road noise control

Active road noise control (ARNC) systems have been widely used for low-frequency noise control in vehicle cabins. To address nonlinear distortions in ARNC systems, prior work introduced a fully causal deep neural network (DNN)-based ANC framework [WaveNet-Volterra Neural Network (WaveNet-VNN)], which achieves superior performance over the ideal Wiener solution and conventional adaptive algorithms under rigorous and fair comparisons. However, real-world road noise exhibits significant distribution shifts caused by varying road conditions, traffic density, and vehicle speed, limiting the generalization of DNN-based ANC methods in practical ARNC systems, especially with scarce training data. This paper presents an adaptive neural network that integrates an online adaptation mechanism while preserving the high performance of causal neural networks. Specifically, lightweight Adapter modules are inserted into several layers of the pre-trained WaveNet-VNN model to serve as fine-tuning components during online adaptation, increasing model parameters, and computational cost by only 5%. Experimental results on a 42 × 2 × 2 ARNC system demonstrate that the proposed approach outperforms the ideal Wiener solution under cross-day distribution shifts and consistently delivers superior performance across different vehicle speeds. It also achieves convergence speed comparable to state-of-the-art traditional algorithms while demonstrating strong robustness across various scenarios, providing a feasible and effective solution for real-world DNN-based ARNC.