Research On HRRP Unknown Target Discrimination Based On Deep Shallow Feature Fusion

Traditional radar high-resolution one-dimensional range profile(HRRP)target recognition requires training to complete classification and discrimination.However,in actual recognition,the target may be an out of library target that cannot participate in training,leading to being misjudged as a known category.Therefore,unknown target identification needs to be carried out first before conventional known target identification can be completed.The target discrimination of radar HRRP is relatively complex,and it is difficult to achieve the requirements using a single feature recognition.Different feature fusion can complement each other’s advantages and improve the target discrimination rate.Therefore,this thesis conducts research based on the method of deep shallow feature fusion,and the research content is as follows:1.Propose a method for identifying unknown targets based on the fusion of Bidirectional Gated Recurrent Neural Network(Bi-GRU)features and subspace features.This method improves the target recognition rate by fusing the time-domain correlation features extracted by the bidirectional GRU network with the global shallow features extracted by the four subspace learning method(principal component features,linear discriminant features,kernel principal component features,and kernel discriminant analysis features).The simulation experiment results show that when the signal-to-noise ratio is 0dB,the discrimination rate of the discrimination method based on the fusion of deep features and subspace features of the bidirectional GRU network is increased by more than 1.94% compared to this single deep feature discrimination method.Among them,the discrimination rate of the feature fusion discrimination method based on kernel discriminant analysis and bidirectional GRU network reaches86.45%.2.A method for identifying unknown targets based on the fusion of parallel residual network features and manifold learning features is proposed.By fusing the deep features extracted by parallel residual networks with four local shallow features extracted by manifold learning algorithms: local linear embedding features,equidistant mapping features,local preserving projection features,and T-distribution random neighborhood embedding features,richer local information is obtained,making the target description clearer and further improving the discrimination rate.The experiment shows that when the signal-to-noise ratio is 0dB,the discrimination rate of the fusion method based on parallel residual network deep features and manifold learning features is increased by more than 4.83% compared to this single deep feature discrimination method.Among them,the feature fusion discrimination method based on T-distribution random neighborhood embedding and parallel residual network fluctuates within 7.65%,which has good noise resistance.3.Propose a method for identifying unknown targets based on the fusion of attention convolutional network features and correlation analysis features.Considering the differences in the contribution of different local features to recognition,a convolutional network based on attention mechanism is used to extract features that are more advantageous for classification.At the same time,a correlation analysis algorithm is introduced to extract the shallow discriminative features with the highest correlation between the training samples and the class matrix,obtaining more information related to the class.The attention convolutional network features are fused with the extracted typical correlation features,kernel typical correlation features,and discriminative typical correlation features,thereby further improving the discrimination rate.The experiment shows that when the signal-to-noise ratio is 0dB,the discrimination rate of the convolutional network feature fusion method based on attention mechanism and related features is improved by more than 4.5% compared to this single network feature discrimination method,and the discrimination rates are all above 71.5%.