Array Signal Processing Based On Machine Learning

Array signal processing is an important branch of modern signal processing,which has important applications in mobile communication,radar,sonar,radio astronomy,electronic medicine and other fields.At present,the classical direction-finding methods are model driven,such as MUSIC,ESPRIT and various improved algorithms.However,in practical applications,the model driven methods always face severe challenges of complex environments such as array imperfections,low Signal to Noise Ratio(SNR),etc.Machine learning is a useful data-driven tool for signal processing and information acquisition,which has been widely concerned in data mining,text understanding,image processing,pattern recognition(face,voice)and many other fields.In recent years,many scholars at home and abroad have begun to explore the application of machine learning methods in the field of array signal processing,in order to solve the problem of poor adaptability to complex environment of classical direction-finding methods.The essence of using machine learning methods for direction-finding is to extract the feature containing the incident angle of the signal from the array observations,and establish a nonlinear mapping relationship between the extracted feature and the signal angle through learning or training.The recent achievements show that machine learning based array direction finding methods have great advantages in adaptability to complex environment.However,at present,the research work on machine learning based direction-finding methods is still in its initial stage.How to furtherly improve the adaptability of such methods to actual environments such as different signal types,various array imperfections,small training samples,etc.,is still the next research direction.This dissertation focuses on the application of machine learning theory in solving the problems of wideband signal direction finding,super-resolution direction finding,direction finding with model mismatch and direction finding with small training sample support.The main contributing work of the dissertation are as follows:(1)In view of the demand of variable carrier frequency signal direction finding and broadband signal direction finding in actual systems,two methods based on SVR(Support Vector Regression)are proposed accordingly,namely VFSVR(variable frequency SVR)and CWSVR(coherent wideband SVR)respectively.VFSVR uses pre-processing and post-processing strategies to complete variable carrier frequency direction finding using the training data with a single frequency.Specifically,the preprocessing strategy uses the array structure characteristics of the uniform linear array to reduce the dimension of the input features without decreasing the direction-finding performance.The post-processing strategy uses the DOA estimation model trained with data of single carrier frequency to complete the angle estimation under different frequencies,making use of the coupling relationship between frequency and angle.Based on the above two points,VFSVR greatly reduces the amount of training data required by the current variable carrier frequency direction finding methods,thus reduces the cost and complexity of training greatly.CWSVR takes the covariance matrixes in multiple frequency bins as the input and the signal angle as the output,then establishes the function mapping relationship between them through SVR.CWSVR avoids the frequency focusing procedure in the conventional subspace based wideband direction-finding methods,therefore,it relaxes the requirements of the array element spacing to ensure unambiguous direction finding,and the corresponding direction-finding performance is superior than the existing subspace based wideband direction-finding methods.(2)In order to satisfy the increasing requirements in direction-finding accuracy and resolution in complex environment,a deep convolutional neural network based direction of arrival estimation method(DCNN-DOA)is proposed by exploiting the powerful representation ability of deep learning to establish the mapping relationship between the spatial spectrum and the array output.DCNN-DOA makes use of the inherent spatial sparsity of the incident signal and the characteristic that the signal-to-noise ratio of the array output covariance vector is higher than that of the original array output under the condition of large snapshots number,and construct a deep convolutional neural network to realize high-precision spatial spectrum reconstruction with array output covariance vector as input and real spatial spectrum as output.DCNN-DOA is superior to the traditional sparsity inducing DOA estimation methods in adaptability to low signal-to-noise ratio and spatial adjacent signals.Moreover,it has two to three orders of magnitude advantages in computational efficiency.Compared with the existing direction-finding methods based on deep learning,DCNN-DOA has exploited the spatial sparsity of the incident signals,so it has more advantages in the DOA estimation performance.(3)The problem of direction-finding for array signals under the condition of model mismatch is studied,and a method based on "deep unfolding" deep network LISTA(learned iterative soft threshold algorithm)is proposed to enhance the adaptability of direction-finding system to various array imperfection.LISTA uses the iterative rules of the iterative regularization method to design the network structure and initialize the parameters,then uses the training data to further optimize the network parameters in the training process.It is a typical "Data-Driven + model prior" learning architecture,which enables LISTA to achieve the same performance as the original iterative regularization algorithm with fewer layers.At the same time,LISTA can directly use the training data to correct the model errors in the learning process under the conditions with obvious array errors,so its error adaptability is better than the conditional model driven methods.(4)In view of the fact that it is relatively difficult to obtain labeled samples in actual direction-finding systems,especially when the systems are disturbed by various array imperfections,this dissertation has explored the application of semi-supervised learning in DOA estimation to improve the learning performance.Semi-supervised learning can automatically exploits unlabeled samples during the training process.Two semi-supervised direction-finding methods are proposed,one is based on Manifold Regularization,which is named MR-SSDOA(Manifold Regularization Based semi supervised DOA)and the other is based on deep network pretraining,which is named Pretrained DCNN.MR-SSDOA method uses a small number of labeled samples and a large number of unmarked data accumulated in the working procedure to gradually modify the angle estimation function by manifold regularization constraints,so as to improve the direction-finding performance in the case of limited labeled data.The pretrained DCNN method constructs an autoencoder to realize supervised pretraining so as to provide a better initial parameter for the deep neural network.After pretraining,the labeled samples are used to fine tune and optimize the network parameters furtherly,so that the deep network can converge to a better solution with a small number of labeled samples,so as to improve the performance of spatial spectrum reconstruction.These semi-supervised DOA estimation methods can make full use of the unlabeled samples in practical scenarios,and alleviate the performance degradation of the supervised learning based methods when labeled data is insufficient.At the same time,pretrained DCNN can maintain the advantage of error adaptability of the machine learning based direction-finding method.