Radar High Resolution Range Profiles Target Recognition Based On Wavelet Autoencoder-recurrent Neural Network

As the wideband radar develops,the resolution of radar is becoming much smaller than the target size,so the signal of target in the returned echo will occupy some range resolution cells.A radar high resolution range profile(HRRP)is the amplitude of coherent summations of the complex time returns from target scatterers using wideband radar,which represents as a one-dimensional signature.It infers some target structure signatures,such as scatterer distribution,target size,etc.Meanwhile,HRRP is very easy to acquire and efficent to process and of great importance for target recognition and classification.So that it received intensive attention in the field of radar automatic target recognition(RATR).In this paper,considering the spectral characteristics of radar echo data and the correlation between range cells,the target recognition based on radar high resolution profile is discussed and studied.The main content of this thesis is summarized as follows: The first part discusses and analyzes the physical properties of high resolution range images and preprocessing methods.Based on the scattering point model,the physical characteristics of the radar high resolution range image are studied,and the solutions to the three main problems of target-aspect,timeshift and amplitude-scale sensitivity of HRRP target recognition are proposed.Then,the non-stationary and time-frequency representation methods of the signal,such as short-time Fourier transform and wavelet transform,are introduced.The advantages and disadvantages of the both methods are discussed and analyzed.The second part focuses on the radar HRRP target recognition based on the recurrent neural network.First,the measured data of this paper is introduced in detail.However,the traditional methods ignore correlation and dependency between range cells should not be ignored.And though features based on the physical characteristics of the target has a clear physical meaning and is easy to extract,human intervention is required.There still be some limitations of the methods above in radar HRRP target recognition.The recurrent neural network(RNN)is an effective model for processing time sequence data.Using this kind of network for target recognition will make sense.In order to extract the features that are more effective for target recognition,take the temporal correlation between the range cells of HRRP samples into consideration.Firstly,the HRRP data is separated into time sequences,and then the RNN model is adjusted for HRRP target classification.Finally,the radar HRRP target recognition based on recurrent neural network is evaluated with measured data.The related experimental results are analyzed and discussed.The third part is contributed to wavelet autoencoder for radar HRRP target recognition with recurrent neural network.The second part mainly discusses the RNN and its application on radar HRRP target recognition,especially taking into account the correlation and dependency betweent range cells of HRRP.For non-stationary signals such as HRRP,the frequency domain often contains characteristics that the signal does not have in the time domain.If only the information in time domain is considered,other information beneficial to the identification will be ignored.Therefore,it is desirable to design an end-to-end model that can consider both the time correlation and frequency spectrum of the data,in order to obtain the features that are more favorable to the final classification task.Therefore,the wavelet autoencoder with RNN are proposed,combining the autoencoder and RNN,in which the weights of decoder are designed purposely.The experimental results show that the features extracted by the proposed model can not only have the frequency domain characteristics in a way,but also be directly affected by the final target classification task,and the recognition can be efficient and effective.