The Research Of Microphone Array DOA Estimation Based On Convolutional Neural Network

The application of DOA estimation based on microphone arrays has been widespread.However,the current DOA estimation algorithms generally have some defects,such as the performance degradation in the environment with strong noise and reverberation,the high computational complexity of the estimation algorithm etc.The estimation algorithms based on deep learning can work well in extreme environments,and have strong robustness with array imperfection.We researched how to use deep learning to improve DOA performance under conditions with noise,reverberation and array imperfection.Our main tasks as follows:1.Research a multi-narrowband sound source DOA estimation algorithm based on CNN firstly.DOA estimation range is divided into corresponding number of sub-intervals by several CNN-based automatic encoders.Then obtain the DOA estimation result of the spatial spectrum structure with CNN-based multi-mask classifier.The algorithm has strong robustness and accuracy in multi-source DOA estimation under low SNR.2.Use ResNet to estimate the DOA of the microphone array signal under conditions with noise and reverberation.It used the cost function based on Euclidean distance to get the regression model of DOA estimation.We tested the ResNet-based DOA estimation algorithm with the array imperfection.The robustness of the ResNet-based DOA estimation is significantly stronger than other algorithms.3.A data compression structure with RBM-based automatic encoder is researched.It is trained with pure-noisy data set to improve DOA performance in environment with strong noise.The RBM-based automatic encoder can accelerate the neural network training process and ensure the robustness to noise at the same time.4.Test the accuracy and stability of three algorithms above in anechoic chamber.