Research On Signal DOA Estimation Based On Artificial Intelligence

Direction of Arrival(DOA)estimation is a key issue in the field of signal processing,especially in radar and wireless communication.Despite receiving widespread attention in the past few decades,traditional DOA estimation algorithms are still limited by factors such as source number variation,array errors,etc.Deep learning extracts array signal features in a data-driven manner to achieve DOA estimation.In recent years,many researches have confirmed the potential and feasibility of applying deep learning to DOA estimation.Based on multi-layer perceptron and convolutional neural network,solutions for coherent source angle measurement,broadband DOA estimation,array error and other scenarios have been proposed.However,the application of Transformer model,which has become the standard in the field of natural language processing,in the field of DOA estimation has not yet been studied.This thesis first studies the framework of applying deep learning to DOA estimation,and models the DOA estimation problem as a multi label classification problem according to the signal model,MUSIC,compressed sensing and other traditional algorithms.At the same time,the dataset for simulation training was studied,and a workflow for solving DOA estimation problems using deep learning was provided.On these basis,a Transformer based DOA estimation model was introduced,describing the customization and adjustment of Transformer for DOA estimation problems.The experiment proved that the model has effective application in single source scenarios.A weighted cross entropy algorithm is proposed to solve the problem of low spectral peaks of model output in multi source scenes,which effectively improves the performance of the model under multi source conditions.In addition,this thesis studied how to handle the broadband DOA estimation problem and provided a method that is easy to implement.Due to the difficulty of obtaining a large amount of data in practical applications,a transfer learning-based solution is provided to the correction problem of array errors.In the experiment,the model pre-trained on a large amount of simulation data is transferred to the array data with specific errors.The results show that the method has good error correction ability and requires less computing resources.