DOA Estimation Algorithmbased On Neural Network

DOA estimation refers to solving the Angle of incoming wave signal by processing the received data of the antenna.After obtaining the Angle of the signal,it can better conduct beam forming or detect the position of users.Therefore,DOA estimation is an important research Direction in many fields.The classical DOA estimation algorithms include subspace class algorithm and sparse reconstruction class method.Among them,the most representative method of subspace class algorithm is multi-signal classification(MUSIC)method,which has better performance at high signal-to-noise ratio,but higher computational complexity and is easily affected by noise.Although the sparse reconstruction algorithm can estimate parameters in a single snapshot,it will bring gridding effect,and the real Angle is difficult to be located in the selected grid,or it is almost impossible to be located exactly in a certain grid.Therefore,we need to find a gridless DOA estimation method,so that it still has good estimation performance at low SNR.In recent years,neural network,with its strong nonlinear fitting ability of application in many fields,the DOA estimation problem can also be regarded as essentially supervising them problems,namely by neural network to learn the array receiving data and the nonlinear mapping relationship between the wave signal Angle,this article mainly studies how to establish a suitable neural network structure to realize the gridless DOA estimation.Specific research contents are as follows.For narrow-band signal DOA estimation,this paper established a depth of neural network model,the model including pretreatment,spatial filtering the encoder,multi-layer classifier and linear interpolation of four modules,including spatial filtering the encoder mainly complete the spatial filtering of the signal,in the process can not only to rough classification of Angle,can also reduce the influence of the noise,and multi-layer classification is used to estimate the real Angle and the interval of adjacent grid,finally through linear interpolation with multi-layer classifier output can realize gridless DOA estimation.In addition,in order to reduce the cost of hardware in practice,we introduce the idea of single-bit quantization into the neural network model,and use the single-bit data to train the established neural network model.Finally,the neural network based on original data training,training of neural network based on single bit data were compared with MUSIC algorithm,simulation results show that the neural network method can realize the gridless DOA estimation,and under the condition of low SNR,the neural network model has higher estimation accuracy,under the condition of high SNR,using single bit datatrained neural network to estimate performance than MUSIC algorithm is a bit poor,however,neural network model has lower computational complexity.For DOA estimation of broadband signal,this paper combines RSS focusing method and neural network method to establish a neural network model based on rough classification.As a result of the broadband signal DOA estimation performance as the focus area reduction and improve,so this article use the softmax and probabilistic neural network(PNN)of two kinds of classification is used in the wideband DOA coarse classification,and compared the classification of the two success rate and the computational complexity,the simulation results show that under the condition of low SNR,success rate of PNN classification than softmax classifier,but the computational complexity of PNN is also higher.We then to focus the signals are analyzed based on the results of the coarse classification,and USES the multi-layer classifier and linear interpolation gridless DOA estimation,in the end,the neural network based on rough classification method and the traditional spinning,comparing the signal subspace method(RSS)simulation results show that the neural network method has higher estimation accuracy.