Direction Of Arrival Estimation Based On Deep Neural Network

Direction of arrival estimation is a research focus in the field of array signal processing.At present,there are many mature DOA estimation algorithms.These traditional algorithms have excellent performance in ideal conditions.However,because they are all model driven algorithms and rely heavily on the matching degree between the algorithm model and the real array model,the array errors and other real complex environments will lead to the serious performance degradation.As a data-driven algorithm,deep learning is based on the idea that the model can autonomously find the nonlinear mapping relationship between input and output from the training data.It has received extensive attention in the fields of image processing,language translation,driverless technology and so on.In recent years,some scholars has used the method of deep learning to solve the problems in the field of array signal processing,overcoming the difficulties that some traditional algorithms are difficult to deal with.This thesis focuses on two DOA estimation methods based on different network structures,and introduces the details of the algorithm from the aspects of network model design,input data processing,and parameter selection.The actual direction finding performance of the network is studied under the conditions of reduced training samples and unknown array errors.The main contents of this thesis include the following parts:(1)The basic principle of array signal processing is introduced,including array receiving model based on narrowband signal model,traditional DOA estimation method and common array errors.The basic principle of neural network is introduced,including activation function,loss function,back-propagation algorithm,two main network structures,and the evaluation index of the network.(2)A DNN-based partitioned DOA estimation network using the array covariance matrix as input is built.The network consists of a partitioned network and multiple parallel DOA estimation networks.The main function of the partition network is to divide the input signal into different partitions according to the incident angle of the target,and each partition is connected with a DOA estimation network responsible for the task of estimating the angle of the area.The partition network can reduce the angle measurement range of each DOA estimation network,which can make the original incomplete training data set complete and reduce the dependence of the entire network on the training set.The generalization performance of the algorithm in the non-training environment is verified by experiments.Then,the neural network is trained with training data containing unknown errors,and the comparison with the MUSIC algorithm shows that the algorithm has better adaptability to array errors.Finally,we compared the performance differences of different partitioned networks under various training set conditions.It is proved that the number of partitions will lead to different direction finding performance under different training conditions,and the optimal selection strategy of the number of partitions is given.(3)A CNN-based DOA estimation network using the beamforming spectrum as input is built.The algorithm extracts the local features of the input data by convolution check,and makes the network learn the mapping relationship between the low-precision spectrum and the high-precision spectrum.The direction-finding performance of the algorithm is tested under various conditions.The experimental results show that the CNN network using beamforming spectrum as input retains some of the characteristics of the beamforming algorithm,resulting in a certain directionality of direction finding.In the generalization ability test experiment,CNN network shows better generalization ability of target number,but slightly worse generalization ability of angle interval.Finally,a test is carried out under the condition of array error.It is proved that this method also has good robustness against unknown error,but it will lose some generalization performance.(4)The test platform of DOA estimation system is built.The system is mainly composed of hardware and software.The hardware includes receiving antenna,receiver,acquisition card and signal processor;The software part is mainly responsible for estimating the azimuth parameters of the signal using DOA estimation algorithm.In this paper,the hardware and software parts are introduced respectively.The system software interface is described in detail,and the results of the system function realization are displayed based on the simulation signal.