Research On Adaptive Filtering Algorithm For Indoor Noise Reduction Based On Machine Learning

Most of the daily life of urban residents is spent indoors,so the indoor environment quality has a significant impact on people’s physical health.In order to solve the problem of indoor noise pollution,noise control technology has developed rapidly.Active noise control systems based on adaptive filter algorithm have been widely studied due to the applicability of adaptive filter algorithm in linear prediction of noise signals.However,the noise reduction effect of the adaptive filter algorithm is not ideal since it only references current residual signals and partial noise signal data when updating filter weights.In addition,there is a distance between reference microphone and speaker in indoor scenes.When the filter order is too small or the sampling frequency is too high,it is difficult to effectively reduce noise signal through the adaptive filter algorithm due to the existence of time delay error.This thesis focuses on the following key technical problems to improve the performance of active noise control systems based on adaptive filter algorithm.Firstly,this thesis proposes a method combining random forest regression model to improve noise reduction effect.Using the data collected and calculated by the original algorithm,the regression model predicts the residual signal at the next moment,and corrects the speaker output signal accordingly to improve the noise reduction effect of the active noise control algorithm.Under ideal conditions,the improved algorithm improves the noise reduction effect of airport noise by 16.4 d B compared with Filtered-x Least Normalized Mean Square(Fx NLMS)algorithm.After introducing a weight function to enhance its anti-interference ability,the improved algorithm can still perform noise reduction normally under white noise interference.Secondly,based on the noise control algorithm proposed in this thesis,a delay compensation module is constructed with K-means clustering model to enhance the noise reduction effect of low-order adaptive filters on distant noise.After making delay compensation on the noise reduction system,the K-means clustering model is used to classify the collected data according to its characteristics and predict whether the delay compensation is effective.Simulation results show that under the condition that the increment step of delay compensation is 32 sampling times,the classification accuracy of delay compensation effectiveness can reach 100% through the four-point synergy,and the median of effective delay compensation points is taken as the output value and fed back to the active noise control algorithm.This can ensure the effectiveness of delay compensation and enhance the ability of low-order active noise control systems to reduce distant noise.Finally,an active noise control system based on traditional adaptive filter algorithm is built in a high-rise indoor area,and the actual street noise is actively reduced by this system to analyze the noise reduction effect of the traditional adaptive filter algorithm.Machine learning models are trained with experimental data,and offline experiments are conducted to verify the improved algorithm and compare the noise reduction performance of the two algorithms.Experimental results show that the improved algorithm improves the average noise reduction effect by 14.9 d B compared with Fx NLMS algorithm.In addition,the offline verification experiments of the delay compensation module have also been carried out to verify the classification performance of the K-means clustering model on delay compensation effect under multi-point collaboration.