DEEP RECURRENT NEURAL NETWORK ALGORITHM FOR ACTIVE SOUND QUALITY CONTROL OF WIPER-WINDSHIELD FRICTION NOISE

The effectiveness of traditional Adaptive Noise Equalizer (ANE) algorithm and its extension algorithms for Active Sound Quality Control (ASQC) systems is unsatisfied in engineering application, especially for the nonlinear problems. In this paper, a nonlinear active sound quality control algorithm based on Deep Recurrent Neural Network (DRNN) is proposed for the wiper-windshield friction noise. A DRNN model based on Long Short-Term Memory (LSTM) neutral network is constructed to perform nonlinear mapping on the input signal through setting an objective function. The trained output secondary signal is counteracted with the expected signal to obtain the minimum error signal. Thus, the linear filter in traditional algorithms is replaced by the proposed DRNN model. Setting the wiper-windshield friction noise of an actual vehicles as the input signal for the DRNN algorithm, simulation analysis was conducted. The results were compared with the input original signal in terms of control effectiveness in both time-domain and frequency-domain. Meanwhile, the psychoacoustic attribute metrics such as loudness, roughness, and sharpness are calculated for the simulating output signals. The results demonstrate that the proposed DRNN active sound quality control algorithm has a good control effect on the wiper-windshield friction noise, especially for the frequency range of 0-500Hz, which is 67.39%, 62.58%, and 56.38% lower than the original noise in terms of loudness, roughness, and sharpness, respectively. The amplitude of the noise is simultaneously reduced. Therefore, the proposed algorithm has significant advantages in improving the vehicle interior sound quality by controlling the wiper-windshield friction noise.