Research On Direction Of Arrival Estimation Method Based On Sparse Bayesian Learning And Deep Neural Network

With the rapid development of information and communication technology being applied in various industrial fields,the environment of array signal processing systems has been becoming more and more complex.The limitations of conventional array signal processing methods under the conditions of low signal-to-noise ratio,small number of snapshots,spatial proximity signals and related signals have limited the development of DOA estimation methods.In recent years,the rapid development of compressed sensing and deep learning has been promising to advance the conventional methods.Sparse representation and deep neural network make it possible to achieve high-precision and high-efficiency DOA estimation in many bad conditions.However,the former has problems of grid selection,off-grid gap,performance susceptible to the position of grid points,and balancing resolution and computational efficiency,while the latter has problems of requiring huge training data set and complex network structure to ensure the estimation accuracy.In order to address these problems,this dissertation focuses on DOA estimation methods based on sparse Bayesian learning and deep neural networks.The main contents are as follows:1.A grid-based multi-snapshot sparse Bayesian learning method(MSBL)is first introduced in this dissertation.To solve the off-grid gap problem of MSBL,an improved algorithm IMSBL is proposed.The grid points are treated as parameters and updated iteratively to approximate the DOAs via a dynamic grid update strategy,which can obtain higher estimation accuracy than MSBL.In order to speed up the convergence of the dynamic grid update strategy and solve the problem of susceptibility to the initial position of the grid point,another improved algorithm SMSBL is proposed.The step grid update strategy leads SMSBL to achieve higher estimation accuracy and efficiency than IMSBL,and has greater advantage over MSBL and other sparse representation algorithms.2.For the grid-based sparse Bayesian learning methods,it is difficult to balance the resolution and computing efficiency because of their failure to estimate multiple signals between adjacent grid points.To address this issue,a grid reconfiguration DOA method(GRDOA)is proposed,which contains two processes of coarse estimation and fine estimation.Based on the idea of dynamic grid update and grid fission,the initial uniform grid is adaptively reconfigured into a non-uniform grid,which can successfully estimate the multiple signals between adjacent grid points.GRDOA can achieve high resolution while maintaining high computational efficiency and estimation accuracy.3.Deep Neural Network(DNN)technology is examined.In the unfavorable environment characterized by such as low signal-to-noise ratio,small number of snapshots and spatial adjacent signals,the performance of the traditional methods will deteriorate drastically.Moreover,since these methods contain time-consuming operations(e.g.,feature decomposition and multi-dimensional search),it is not suitable to be applied in the real-time scenarios.To solve these problems,DNN is used to learn the feature of the array output covariance matrix,and the DOA estimation methods DNN?RI and DNN?AP are obtained.They have higher estimation accuracy and computational efficiency than the traditional methods under the unfavorable conditions.To further reduce the estimation error of DNN based methods,a spatial partition DNN based method(DNN?FQ)is proposed.The spatial area is divided equally and the partition of the incoming signal is determined firstly,and then the DOA estimation of each partition is realized independently.Since the partition method fully suppresses the noise interference without increasing network complexity,it further improves the estimation accuracy of the DNN method,and maintains a high computational efficiency,which is suitable for real application scenarios.4.In order to combine the advantages of both sparse representation and deep learning,a sparse representation-based dilated convolutional neural network(DCNN)method(DCNN?S)is proposed,in which the sparse representation of the array output is used as a feature extraction method,and a Dilated Convolutional Neural Network(DCNN)is designed for feature learning.Due to the full use of the sparse prior knowledge of the array output and the dilated convolutional layer,which can expand the receptive field of the network without increasing the its complexity,the learning ability of the network is improved greatly.DCNN?S has better estimation performance than other deep learning and sparse representation based DOA methods.5.To solve the problem that traditional source number estimation methods cannot work in unfavorable environments,a DCNN-based source number estimation method(DCNN?K)is proposed,which can work in bad conditions and has a superior estimation performance.At the same time,in order to fully extract the prior information of the array output to improve the estimation performance of deep learning-based methods,combining subspace theory and deep learning methods,a DOA estimation method(DCNN?SU)is proposed.A subspace spectral is first constructed as the feature representation,and then a dilated convolutional neural network is designed to realize DOA estimation.Because DCNN?SU makes full use of the prior information of the array output signal subspace and noise subspace,it is superior to other subspace methods and deep learning methods in terms of estimation accuracy and operating efficiency,and is insensitive to the estimation results of the number of sources.