Research On Gridless DOA Estimation Based On Deep Learning

Direction of Arrival(DOA)estimation is an important research direction in the field of array signal processing,and has a wide range of applications in communication,radar,sonar and navigation.Traditional DOA estimation algorithms are usually based on complex mathematical models with strong interpretability.However,traditional algorithms require more a priori knowledge and larger computation,and are difficult to cope with strong noise environments and more complex scenarios.The DOA estimation method based on deep learning is a data-driven method,which uses deep neural networks to directly model and process the input signal,and then realize the direction finding.This method is more adaptive and scalable,and can automatically learn the features of the signal from the data,while optimizing for different scenes and noisy environments.In contrast,most of the existing DOA estimation methods based on deep learning treat the direction finding problem as a multi-label classification task,so these methods are essentially on-grid methods,and may encounter grid mismatch effect.To address the above problems,this dissertation focuses on deep learning-based gridless DOA estimation methods:This dissertation focuses on the generalized linear array,which can be regarded as a uniform linear array(ULA)with/without “holes”.By using the Toeplitz structure,a Covariance Reconstruction Network(CRN)is proposed to estimate the noiseless covariance matrix of the aforementioned ULA with “no holes”,based on which the DOAs can be retrieved by using root Multiple Signal Classification(root-MUSIC).To increase the generalization,the CRN is pre-trained for different number of sources and the network parameters are stored in the database.Then this dissertation proposes a target detection network for source enumeration in order to select the corresponding CRN parameters from the database.Simulation results show that the proposed method can detect more sources than the number of array elements and is not affected by grid mismatch effect.In order to improve the DOA estimation performance of the network in the high SNR region,this dissertation uses Residual Attention Network(RAN)to focus on the important features in the input and avoid the problems of gradient disappearance and network degradation when deepening the CRN network structure.Meanwhile,in order to make the network applicable to different number of sources scenarios,this dissertation has trained the network in different number of sources.However,due to the diversity of different number of sources,training the network is still very complicated and requires high cost.Therefore,this dissertation introduces the idea of transfer learning,which solves the problem of overly complex training without affecting the network performance and provides a feasible solution for larger array scenarios.