The Research On Robust DOA Estimation Algorithm Based On Deep Learning

The Direction of Arrival(DOA)is estimated to be a major problem in the field of array signal processing.It has a wide range of functions in automotive artificial intelligence,remote video conferencing,and home intelligent voice assistants.In the last ten years,the DOA estimation technology based on machine learning and deep learning has made great progress.The DOA estimation technology based on this type of algorithm has better generalization,robustness,and accuracy.This article studies in depth based on this DOA estimation problems for the following three different application scenarios and algorithms:(1)To solve the problems of poor stability,low positioning accuracy of traditional neural network DOA estimation algorithms,and the need for radio frequency transmitters for common drone positioning,a locally weighted long short-term memory(LWLSTM)The UAV DOA estimation algorithm for the network uses a circular array antenna to perform real-valued and feature extraction on the UAV OFDM signal power,and selects the ratio of the average received signal power of the nth array element to the sum of the average power values of all array elements as the network Input,restore the azimuth angle θ through the LWLSTM network,and realize the DOA estimation of the drone.(2)Aiming at the low accuracy of the DOA estimation of the traditional algorithm under low signal-to-noise ratio,a speech DOA estimation algorithm based on SSDAE-DNN is proposed.First,extract the features of the upper triangular array of the array covariance matrix as the input of the SSDAE neural network.Through the transfer learning strategy,the training weight of the SSDAE network is given as the initial weight of the DNN network,and the antinoise performance and convergence speed of the DNN framework are enhanced.Finally,The DNN framework is used to recover the azimuth angle θ to achieve speech DOA estimation.(3)Aiming at the problem of poor generalization of the traditional algorithm DOA estimation algorithm and inaccurate bit estimation accuracy in different situations,an Attentionbased mechanism and a Bi Locally Weighted Long Short memory network were proposed.-Term Memory(Bi LWLSTM)DOA estimation method,using the triangle array of the imaginary and real parts of the covariance matrix of the array received signal as the input to generate the adversarial neural network,using Attention's optimization classification problem to assign weights,and then improving the algorithm through Generalization.Finally,high-precision DOA estimation bearing results are obtained,which improves the bearing accuracy.Finally,the above three DOA estimation methods are implemented and compared with existing traditional algorithms to verify the effectiveness of the three proposed methods.