Deep Learning Classification Network For Sound Source Locatization

In recent years,sound source localization technology has been widely used in intelligent robots,emergency rescue,fault judgment and other fields.In order to obtain better positioning accuracy,traditional sound source localization methods have higher requirements on the environment and hardware equipment when solving practical problems.Compared with traditional methods,deep learning converts the positioning method into pattern recognition,which improves the shortcomings of traditional sound source localization methods.This paper studies the deep learning classification network method based on power features,which is used to determine the source direction of sound signals,and proposes a centroid optimization algorithm based on deep learning classification network.By obtaining the phase information of sound signal,the paper studies the convolutional neural network based on the GCC-PATH feature,and uses the centroid optimization algorithm to further process the method.This paper first studies the application of the deep learning classification network in sound source localization.However,since the classification network is essentially a multi-label classification problem with discrete nature,there are inevitable errors when using the classification network to locate the sound source from any direction.In order to improve the sound source localization accuracy of deep learning classification network(DNNC),this paper proposes a centroid optimization algorithm based on deep learning classification network(CO-DNNC).The CO-DNNC method includes three parts: sound signal preprocessing,classification network prediction and centroid optimization.In the sound signal preprocessing,in order to improve the accuracy and robustness of the network,the signal strength is normalized,and random noise is added to the power data.The classification network uses convolutional neural network which is good at processing grid data.According to the predicted probability output by the classification network and the corresponding classification label,the centroid optimization algorithm predicts the direction of sound source.The experimental results show that in the case of low signal-to-noise ratio,the CO-DNNC method characterized by the sound signal power can better improve the accuracy of sound source localization.This paper obtains the phase information of the sound signal according to the time difference of the signal arrival at different detectors,then processes it through the generalized cross-correlation-phase transformation algorithm(GCC-PATH),and feeds the delay characteristics to the convolution classification network.The convolutional neural network(GCNNC)network structure based on GCC-PATH features uses a6-layer CNN network to better fit GCC-PATH features.CO-GCNNC uses centroid optimization for the results of GCNNC network prediction,which is used for more accurate calculation of network prediction results.The experimental results show that the accuracy and error of CO-GCNNC are better than GCNNC,under the same signal-to-noise ratio and the number of classifications.However,in the case of high signal-to-noise ratio,increasing the number of classifications has a better result in improving the accuracy.Compared with CO-DNNC,this way demand for data sets increases and the complexity of the network increases.Through the above research,the accuracy of sound source localization prediction with any azimuth source can be improved for power characteristics and GCC-PATH characteristics.