Research On The Direction Of Arrival Estimation Method Of Array Signal Based On Machine Learning

Direction of Arrival estimation based on microphone array is an important branch in the field of array signal processing.It has a wide range of applications in military and civilian fields,such as military radar positioning,mobile phones,and video conferencing systems.In recent years,due to the excellent performance of machine learning algorithms and deep learning in various fields,the application of machine learning and deep learning to sound source localization has become the mainstream direction of research.Compared with traditional sound source localization algorithms,sound source localization methods based on these algorithms have better robustness and accuracy.For such methods,this paper has done in-depth research in the following aspects:1.In order to improve the degree of freedom of DOA estimation of microphone array signals,an under-determined wideband DOA estimation algorithm based on sparse bayesian learning(SBL)in the unknown noise field is proposed.The algorithm uses the coprime array to provide the characteristics of the virtual array element and increase the freedom of the array.The SBL algorithm does not require prior information about the number of source signals.Through fixed-point update,the sparse signal reconstruction achieves the effect of global convergence,which further enhances the adaptability of DOA estimation in the case of low signal-to-noise ratio,and improves the stability and accuracy of DOA estimation.2.Aiming at the problem that the feature data is not easy to distinguish due to reverberation and noise interference,and the positioning performance is degraded,this chapter proposes a DOA estimation method of long short-term memory(LSTM)neural network based on linear discriminant analysis(LDA)feature transformation.Firstly,calculate the covariance matrix of the reverberation signal received by the microphone,then use the covariance matrix to calculate the linear discriminant analysis projection matrix W,and convert it into the projection data set Z,use the projection data set Z to train the LSTM neural network to obtain the DOA estimation of the signal.3.Aiming at the problems of poor stability and low accuracy in DOA estimation of traditional neural networks,and increased errors in the case of low signal-to-noise ratio and too many or too few samples,a DOA estimation method based on support vector machine(SVM)locally weighted LSTM(LWLSTM)neural network is proposed.Firstly,the wideband signal is received by the microphone array and preprocessed,and then the upper triangular matrix of the processed signal covariance matrix is converted into the input sequence as the input training network of the LWLSTM neural network,and the DOA estimation result of the subband signal is obtained.Finally,the SVM model is trained through the output result of the subband signal,and the DOA estimation result of the subband signal is fused to obtain the DOA estimation result of the wideband signal.4.Analyze the DOA estimation algorithm proposed in Chapters 3,Chapters 4,and Chapters 5 and implement the speech positioning system.The experimental results show that the DOA estimation algorithm proposed in this paper has high positioning accuracy and stronger robustness,which can meet the needs of actual positioning.Finally,summarize the work done in this paper,and make an outlook for the existing shortcomings.