Direction Of Arrival Estimation Base On Machine Learning Under Array Model Imperfection

Spatial spectrum estimation is an important branch of array signal processing,widely useed in radar and communications.The traditional spatial spectrum estimation algorithm is model-driven,that is,spatial parameters are estimated based on the match between pre-established function model and received signal.However,the matching between model and data becomes invalid in presence of antenna array imperfection,which significantly reduces the performance of this type of algorithm.In order to solve this problem,this thesis uses machine learning methods to construct a network model to directly learn the nonlinear relationship between received signal and spatial parameters,and realize the direction of arrival(DOA)estimation in presence of antenna array imperfection.This thesis focuses on the problem of spatial spectrum estimation,combined with machine learning to carry out the research of DOA estimation in presence of antenna array imperfection.The specific research content is as follows:(1)Constructed the array receiving signal data set under the conditions of different antenna array imperfections,signal-to-noise ratio,and number of training samples.First,based on the array model,generate an array steering vector including three array imperfections of channel amplitude and phase,element position deviation,and mutual coupling;then,according to the array steering vector and random spatial signals,generate received signals with different signal-to-noise ratios and different strength of the imperfections;finally,extract features from data and reduce dimensionality of data,and calibrate the randomly generated incoming angle to obtain a received signal data set containing multiple array imperfection at the same time,which provides a complete data sample set for network model based on machine learning in the following.(2)Designed the classification scheme and fitting scheme of the direction of arrival estimation based on neural network in presence of antenna array imperfection.In the classification scheme,the direction finding problem is converted into a multi-classification problem by dividing the effective airspace angle range into multiple areas,and linear interpolation is used to eliminate the gridding influence of the classification model.In the fitting scheme,based on the continuity and boundedness of the trigonometric function of the coming angle,the network model uses two-node output representing the sine and cosine values of the angle,directly obtain the output information including the coming angle parameter,which avoids peak searching.The simulation results show that the performance of the proposed classification estimation method is not affected by the strength of the array imperfections,and it can achieve stable and high-precision angle prediction.Compared with the classification method,the proposed fitting estimation method has faster convergence speed,higher estimation accuracy and more stable performance.(3)Three DOA estimation schemes based on complex convolutional network,Wide-Shallow complex residual network and Narrow-Deep complex residual network in presence of antenna array imperfection are proposed.Aiming at the problem of low utilization of received signal phase information,firstly,a complex convolutional network framework with a complex-valued covariance matrix as input is built for the direction angle estimation.Then,using different receptive fields and network depths,two scales of complex residual network models: Wide-Shallow,Narrow-Deep,are designed to enhance the characterization ability of complex convolutional networks.The simulation results show that when the signal-to-noise ratio reaches a certain threshold,the estimation accuracy of the complex convolutional network is not affected by the strength of the array imperfection,and can quickly converge to the global optimum.The Wide-Shallow complex residual network has a slightly lower estimation accuracy when the array imperfection is strong,but it can always maintain a high value.The estimation performance of the Narrow-Deep complex residual network is significantly better than the Wide-Shallow complex residual network,and is not affected by the strength of the array imperfection,In some cases,the estimation accuracy is higher than real residual network.