Research On Array Error Calibration And DOA Estimation Based On Neural Network

Direction of Arrival(DOA)technique is widely used in wireless communication,radar localization and other fields by processing the received signals of antenna array to obtain the direction of incoming waves in space.However,in actual engineering practice,the array carries array errors due to external or internal factors,making the array flow matrix unknown a priori,which leads to the rapidly poor performance of some traditional model-based DOA estimation algorithms.In contrast,neural networks based on data-driven have the advantages of low online computational complexity and strong data fitting capability.Therefore,the research on the DOA estimation algorithm based on neural network under the array with error conditions has research significance.This thesis focuses on array error calibration and DOA estimation under array error conditions,considering two different scenarios:the array with gain and phase errors as well as the partially impaired antenna array.In these scenarios,effective neural network-based DOA estimation algorithms are explored.Firstly,an active calibration algorithm based on cascaded neural network is proposed for the DOA estimation under the condition of array with gain and phase errors.The proposed cascade network consists of a signal-to-noise classification network and two sets of error calibration networks applicable to high and low signal-to-noise ranges.DOA estimation is performed by applying the traditional DOA estimation algorithm by compensating the array mainfold matrix with the gain and phase errors estimated by the cascaded network.To address the problem that the neural network-based algorithm does not generalize to the number of array and array shape,a group calibration strategy is proposed to make the proposed cascade network can be directly applied to antenna arrays of any shape or any number of array after training.Simulation results show that under the same conditions,the proposed cascaded neural networkbased active calibration algorithm can achieve a better balance between calibration performance and calibration computational complexity than the rest of existing mature algorithms.Secondly,to address the DOA estimation problem under the scenario of partially impaired antenna array,aDOA estimation algorithm based on Convolutional Neural Network-Vision Transformer(CNN-ViT)is proposed in this thesis.The algorithm is composed of a CNN-based impaired array elements localization algorithm,a ViT network-based DOA estimation algorithm under the condition that the location information of the impaired array elements are known,and an amplitude interpolation algorithm in series.In addition,the network structures of CNN and improved ViT and the way of data set construction are elaborated in this paper.Simulation results show that under the same conditions,the DOA estimation algorithm based on CNN-ViT proposed in this paper can achieve higher calibration accuracy compared with other algorithms.