Research On DOA Estimation Algorithm Based On Machine Learning

DOA estimation is an integral part of array signal processing.At present,traditional DOA estimation algorithms,such as the classic MUSIC and ESPRIT algorithms,need to perform complex matrix operations.The MUSIC algorithm needs to perform spectral peak search and the ESPRIT algorithm also requires Estimating the multiples of small eigenvalues results in a longer DOA estimation time.So it is difficult to meet actual needs.On the other hand,the signal propagation and array receiving environment is very complicated in practice.There are various defects in the array system,which makes it impossible to accurately model it.The method of machine learning is based on data.Scenes and various defects are artificially assumed and simplified,and complex models can be reconstructed based on the training data set.This article focuses on applying machine learning algorithms to DOA estimation.The specific research content is as follows:(1)The background significance and research status of DOA estimation are analyzed,the traditional MUSIC and ESPRIT algorithms are studied in depth,and the experimental simulation analysis is carried out.(2)The current mainstream machine learning DOA estimation algorithms-support vector machine and neural network are studied.In view of the large errors of these two algorithms in approaching-90 ° and 90 ° arrival angles with a small number of samples,the XGBoost algorithm based on gradient boosting tree and ensemble learning is proposed to be used in DOA estimation.The simulation analysis of each algorithm in different situations is shown.Experiments show that when the signal source is one,the average mean square error of the XGBoost algorithm is less than that of the MUSIC algorithm.It is 0.008853 degrees less than the MUSIC algorithm,and 0.044989 degrees less than the ESPRIT algorithm.The average estimated time is 26.440107 s less than the MUSIC algorithm,and 0.000438 s less than the ESPRIT algorithm.(3)In order to solve the DOA estimation problem of multiple signal sources,this paper proposes a joint DOA estimation model of support vector classification and XGBoost regression.Support vector classification estimates the number of signal sources.For different numbers of signal sources,different trained models are used to estimate the angle of arrival of each signal source.Experiments show that when the number of signal sources is up to three,the classification accuracy of the support vector classification model is the highest,which can reach 99.7%.At the same time,the average estimation time of the single sample by the joint estimation model is 0.001514 s,which can better meet the needs of realtime performance.(4)The performance of the machine learning algorithm is very dependent on the quality of the data set.Considering that the angle of arrival is a continuous value in practice,the angle of arrival is randomly generated from-90 ° to 90 °,and the data set is constructed and simulated.The results show that the XGBoost algorithm's averages mean square error has increased by 0.009858 degrees,and the performance has decreased.But it is still better than support vector regression and neural network algorithms.Further,different numbers of training sets are created and different algorithms are simulated.Experiments show that the performance of the XGBoost algorithm is more stable under different data volumes,and the neural network approaches the performance of the XGBoost algorithm when the data set is 60,000.