Research On Range Profile Recognition Based On Deep Learning

The high resolution range profiles(HRRP)of a target provide the distribution of scatterers along the radar line of sight,which reflect the shape,size,structure of the target.The target recognition with HRRP has become a hot research area due to its advantages such as simple and easy profiling,small amount of data and low computational complexity.The HRRP recognition based on deep learning algorithm is investigated in this thesis.Firstly,the HRRP recognition with deep neural networks(DNN)is studied.Because deep neural network has the problem of error signal attenuation layer by layer,and easy to overfitting when the back propagation algorithm is used for training.This thesis studies how to improve the training process of deep neural network by unsupervised pre-training algorithm.The restricted boltzmann machine(RBM)and sparse autoencoder(SAE)are used to initialize the network parameters layer by layer to replace the traditional random initialization method.In the recognition experiments for four types of aircraft,the performance of the improved pre-training algorithm is better than that of the randomly initialized network,and the average recognition rates is improved by 13%.Among the two pre-training algorithms,the recognition performance based on sparse autoencoder is better than that based on restricted boltzmann machine.Then HRRP recognition based on convolution neural networks(CNN)and recurrent neural networks(RNN)is studied.A CNN-RNN-based HRRP recognition scheme is constructed by combining convolution neural network and recurrent neural network.This method can combine multiple HRRP sequences of the same target at different aspect angles by utilizeing the automatic feature extraction ability of convolution neural network and the time series data correlation processing of recurrent neural network.The simulation experiments show that the average recognition accuracy of CNN-RNN model is 99.4%.Finally,aiming at the aspect sensitivity of the high resolution range profile,the HRRP recognition using hidden markov model is studied.The sparse autoencoders is used to construct a deep neural network to extract features and reduce the dimension of HRRP for HMM classification.The recognition performance of deep neural network and hidden markov model is compared in the simulations.The influence of HMM model parameters on recognition performance is analyzed.