Incoherently Distributed Sources Localization Based On Deep Learning

Direction of arrival(DOA)estimation is a basic problem in array signal processing,which is widely used in military and civilian fields such as radar,sonar,and mobile communications.Therefore,DOA estimation algorithm is of great significance in both theoretical research and practical application.However,in many practical application scenarios,multipath propagation or scattering of the source signal often occurs,giving rise to a certain spatial distribution.In this case,it is necessary to consider the spatial distribution information of the source signal and establish a distributed source model.The traditional subspace-based incoherently distributed source localization algorithm has some problems,such as it is difficult to solve the signal subspace and noise subspace accurately,the distribution function of source signal needs to be known accurately,and the distribution types of multiple source signals need to be consistent.In recent years,deep learning has been studied and applied in the field of DOA estimation,but the existing DOA methods based on deep learning are only for point source models which does not consider the angular spread parameter estimation.In view of the above problems,this paper studies the incoherently distributed source localization method based on deep learning,which provides theoretical and application reference for the incoherently distributed source localization technology based on deep learning.The main research work and achievements of this paper are as follows:1.A DNN-based incoherently distributed source localization method is proposed.This method designs the DNN model into two parts,the central angle estimator and the angular spread estimator,and introduces a Lambda layer to connect the two estimators to form a complete DNN model.The model decouples the central angle parameter and the angular spread parameter,which effectively solves the problem of matching of the parameters for multiple source signals.In addition,the method of transfer learning is introduced to solve the problem of performance degradation when the data sample distribution of the pre-trained model does not contain the target distribution type.Simulation results show that the proposed method has better performance than subspace-based algorithms,and can perform DOA estimation when multiple source signals have different distribution types.2.A CNN-based incoherently distributed source localization method is proposed.The method uses a CNN model to robustly output an angular power density spectrum.Based on the angular spectrum,we propose two methods to estimate the localization parameters.The first method assumes that the angular power density function of each source has one peak and do not overlap greatly with other sources,then a method of peak detection and segmentation is used to achieve parameter estimation.The second method proposes a spatial spectral function independent of the specific array manifold,and then achieve parameter estimation through spectral peak search.Moreover,the transfer learning is used to address the mismatch of the source signal distribution type with the pretrained model distribution.The simulation results show that the proposed CNN model can output a robust angular power density curve,and the computational space spectrum method and the parameter estimation method of peak detection segmentation both have better performance than subspace algorithms.3.The DNN model and the CNN model are further optimized respectively,so that it can judge whether there is an effective source signal incident in the observation range of the array when applied in the actual scene.Then a set of microphone array experimental platform hardware equipment is built,and the corresponding microphone array audio acquisition tool is developed to realize the function of acquisition and storage of measured audio data.The proposed distributed source localization method based on DNN model and CNN model is tested using the measured data collected by the experimental platform,and the experimental results verify the effectiveness and good generalization ability of the proposed method.