Research On Deep Learning Based Direction-of-arrival Estimation With High Performance

Direction-of-arrival（DOA）estimation is an important branch of modern array signal processing technology,and its extensive application background has attracted the atten-tion of a large number of researchers.With the increasingly complex development of the electromagnetic environment,the pulse of radiation sources in space is becoming denser,and the types and quantities of signals are more unpredictable.Traditional model-based DOA estimation algorithms have gradually been unable to adapt to higher estimation re-quirements.At the same time,the rapid development of electronic devices and signal processing algorithms has also brought new challenges to DOA estimation technology.The current processing system urgently needs a DOA estimation method with reliable performance,strong anti-noise ability,and high computational efficiency.However,the traditional model-driven DOA estimation algorithm lacks effective modeling of array er-rors and other information,and its estimation performance will be greatly degraded or even fail in a complex electromagnetic environment.Relying on its powerful functions,deep learning technology has long been applied to the field of signal processing by many scholars,among which deep neural networks have made major breakthroughs in pattern recognition,image processing,and semantic segmentation.In recent years,deep learning technology has also been combined with array signal processing technology and has given birth to some new intelligent estimation methods.Since the DOA estimation technology based on deep learning is data-driven,this type of method can get rid of the prior knowledge in the array model and directly use the hidden joint probability distribution function between the network input and output to effectively predict the DOAs of signals.Relying on supervised network optimization,subspace decomposition estimation methods,and sparse reconstruction theory,this thesis uses deep neural network as an excellent tool to mine the depth characteristics of the received signals of the sensor array system and develop new DOA estimation methods with high performance.The com-plex mapping relationship between signal and spatial spectrum are approximated by the network architecture.The main innovative results achieved in the thesis include:1.This thesis proposes the concept of image tensor of the covariance matrix and uses it in the DOA estimation method based on the deep neural network.In order to design the architecture of deep DOA estimation network,this thesis analysises the depth features of the covariance matrix extracted by the network.With the help of classification-type and regression-type of network models based on supervised learning,this thesis propose a variety of estimated network prototypes for different antenna arrays.The concept of multi-classification and multi-label（MCML）problem is introduced into the architecture of deep DOA networks,which not only improves the accuracy of the network but also enables the network to deal with the DOA information of multiple signals.For the regression-type of deep neural network,this thesis proposes a novel prior power spectrum to effectively suppress the noise in the two-dimensional spatial spectrum,then expands the array size of the uniform circular array and its degree of freedom（DOF）through the Khatri-Rao（KR）product transformation.2.The labels used in the supervised learning are difficult to collect.In order to solve this problem,this thesis proposes the unsupervised training strategy of DOA estimation network for the first time.To accurately recover the DOAs of signals under the condition of one-bit sampling,this thesis combines the subspace decomposition algorithm with the deep learning technology and proposes an unsupervised one-bit DOA estimation network.By replacing the complex operations such as matrix decomposition and spatial spectrum search in the MUSIC algorithm through the iterative optimization networks,this thesis further reduces the complexity of the algorithm and greatly improves the real-time per-formance of the subspace decomposition based estimation algorithm.Furthermore,this thesis perfectly integrates the signal sparse reconstruction technology in the compressed sensing with the deep learning technologies and designs an unsupervised sparse power spectrum recovery network based on the constraint?₁norm based optimization problem.The unsupervised learning strategy greatly improves the generalization of the network,which helps the large-scale networks to be fully optimized without the constraint of the size of the label space.At the same time,this thesis vectorlizes the covariance matrix of the array received signal in the network objective function and introduces the KR product transformation into the network model,therefore,the sparse power spectrum recovery network can normally output the correct DOAs when the number of sensors is less than the number of signals,and this extension need no further processing for network input and output.3.Based on the above-mentioned network design ideas,this thesis also proposes two kinds of DOA estimation network prototypes for the off-grid signals,which are based on the supervised and unsupervised learning,respectively.By further optimizing and im-proving the network training label and the binary cross-entropy（BCE）loss function,the method based on supervised learning proposed in this thesis enables the network to use the category and amplitude in the output spatial spectrum to predict the integer and fractional part of DOAs.Furthermore,this thesis uses a new network training strategy to increase the capacity of the network,so that the network with a simple structure can also achieve estimating task for the off-grid signals.The off-grid DOA estimation technology based on unsupervised learning is implemented with basic idea of the sparse power spectrum recovery network.This thesis proposes to use the parallel fractional estimation network to assist the sparse power spectrum recovery network and output the integer and fractional space spectra respectively.With the principle of?₂norm to recover the covariance ma-trix of the received signal from the array realizes the unsupervised training mode of the network.A variety of representative algorithms are selected for in-depth comparison with the methods proposed in this thesis.Based on the simulation results,the characteristics of all methods are described and analyzed in detail.All results show that,compared with tradi-tional algorithms,the DOA estimation method based on deep learning can still efficiently obtain satisfactory estimation performance under the estimation environment of low SNR and small snapshots.