Research On DOA Estimation Of Large Arrays Based On Machine Learning And Random Matrix Theory

Direction of arrival(DOA)estimation is an important branch of array signal processing and has wide applications in wireless communication,radar and other fields.In recent years,the DOA estimation of large array system has become an important development trend in the field of array signal processing.With great potential in improving beam resolution and system capacity,it has gradually received much attention from scholars all over the world.The classic method of DOA estimation is the subspace algorithm.However,in the case of large arrays,the number of array elements is up to tens or even hundreds,which is close to the number of samples.Thus,the subspace method faces the following problems that need to be solved:(1)The sample covariance matrix of the received signals will be distorted and cannot replace its statistical covariance matrix,resulting in the degradation of the estimation performance of the traditional subspace-based DOA estimation algorithm;(2)The introduction of large array the traditional will bring a huge amount of calculations to the subspace-based methods,and it even cannot run;(3)The traditional subspace-based DOA estimation algorithm is poorly adapted to the actual application environment,especially in the case of low signal-to-noise ratio,its estimation performance is seriously degraded.Machine learning can realize big data analysis and processing,and is less affected by noise and its calculation is fast,which can provide a powerful tool for large array DOA estimation.However,in the case of large arrays,the number and dimensions of input data during machine learning training and testing will also increase,affecting the classification performance,resulting in a reduction in estimation performance.This thesis uses the asymptotic spectrum theory in the large-dimensional random matrix theory to modify machine learning,so that the machine learning method can process large array signals more accurately and efficiently,thereby increasing the robustness of estimation.Therefore,basedon machine learning and random matrix theory,this thesis conducts an in-depth study on the problem of DOA estimation for large arrays.The main work done is as follows:First,a DOA estimation method based on multi-class support vector machine(SVM)is implemented.Simulation results prove that under low signal-to-noise ratio,the DOA estimation method based on multi-class SVM has better performance than the traditional algorithm.Furthermore,in the case of large arrays,by using the asymptotic spectrum theory and kernel spectral clustering theory in random matrix theory to modify the kernel function of the least square SVM,a DOA estimation method based on improved kernel least squares support vector machine(k-LSSVM)is proposed.The simulation experiments show that the estimation error of the k-LSSVM method is smaller than the traditional method and has higher estimation accuracy.Finally,for the problem of long training time and slow learning rate of the k-LSSVM method,a large array DOA estimation algorithm based on extreme learning machine(ELM)is studied.The traditional extreme learning machine requires the activation function to be infinitely differentiable,but under the condition of large arrays,the traditional extreme learning opportunities produce problems such as slow training rate.According to random matrix theory,this paper proves that ELM with ReLU function as the activation function has asymptotic convergence in the case of large arrays.Therefore,a ReLU-ELM-based large array DOA estimation method is proposed in this paper.The simulation results show that under low signal-to-noise ratio,ReLU-ELM method has better estimation performance than traditional algorithms.Compared with the k-LSSVM-based method,the ReLU-ELM method for large array DOA estimation can greatly reduce training and testing time and improve learning efficiency under the premise of ensuring similar estimation performance.