Delayless Generative Fixed-filter Active Noise Control based on Deep Learning and Bayesian Filter

Zhengding Luo, Dongyuan Shi, Woon-Seng Gan

Abstract: The selective fixed-filter active noise control (SFANC) method can select suitable pre-trained control filters to attenuate incoming noises. However, the limited number of pre-trained filters is insufficient to effectively control various forms of noise, especially when the incoming noise differs much from the filter-training noises. To address this limitation and generate more appropriate control filters, a generative fixed-filter active noise control approach based on Bayesian filter (GFANC-Bayes) is proposed in this paper. The GFANC-Bayes method can automatically generate suitable control filters by combining sub control filters. The combination weights of sub control filters are predicted via a one-dimensional convolutional neural network (1D CNN). Based on prior information and predicted information, Bayesian filtering technique is applied to decide the combination weights. By considering the correlation between adjacent noise frames, the Bayesian filter can enhance the accuracy and robustness of predicting combination weights. Simulations on real-world noises indicate that the GFANC-Bayes method achieves superior noise reduction performance than SFANC and a faster response time than FxLMS. Moreover, experiments on different acoustic paths demonstrate its robustness and transferability.