Radar One-dimensional Range Profile Target Recognition Based On Graph Neural Network

The range resolution of high-resolution wideband radar is much smaller than the target size.Its radar echo signal is called one-dimensional high resolution range profile(HRRP)of the target.Because it can reflect important information such as the physical structure of the target,and the amount of data is small and easy to obtain,radar automatic target recognition(RATR)based on HRRP has become a hot research direction in this field.Most of the traditional methods regard the HRRP signal as a whole envelope signal,and do not pay attention to the target structure features contained in the signal,so it is difficult to extract rich structural features for classification.Previous deep learning-based methods focus on extracting local structural features or global features of signals,and rarely consider the correlation between local structures and global features at the same time.Moreover,these methods only use the structural features within the sample and ignore the common features between samples that can be used for classification.In order to improve the problems existing in the past methods,this paper proposes a series of recognition methods of high resolution radar range profile based on graph neural network.The content of this paper can be summarized as the following five points :(1)In order to reduce the adverse effects of sample sensitivity on subsequent classification,this paper analyzes the causes of three sensitivity problems of HRRP signals in detail,and designs a preprocessing method to process the HRRP raw data to weaken the adverse effects of sensitivity on subsequent model training and testing stages.In order to verify the recognition performance of the proposed model,five representative classification methods are selected as comparison methods for subsequent comparison with the proposed method.(2)In order to better extract the correlation and global features between the local structures of the signal,this paper designs a global feature extraction network model based on multi-scale convolution neural network and graph convolution network(MCNN-GCN).Firstly,the local feature extraction module is used to extract the shallow local features of HRRP signal.Then,the global feature extraction module based on graph convolutional network is used to aggregate local features to fully extract the correlation between local structures and global features of the signal.Finally,the classification network is used to obtain the prediction results.Experiments based on measured data show that the multi-channel graph convolution global feature extraction network model can better extract the correlation and global features between local structures of HRRP,and has better recognition accuracy and small sample robustness than traditional recognition methods.(3)In order to more effectively extract the rich physical structure features in the signal,this paper proposes a Graph Transformer radar HRRP recognition model based on dynamic construction graph(DCGT).This method first designs a module to extract local features;then through the Transformer and Top-K method to dynamically construct the graph,the relationship between the local structural features is established.Finally,the graph is learned based on the Graph Transformer network to effectively extract the rich physical structure features in the signal.Compared with the traditional deep learning model,the method of dynamically constructing the graph reduces the number of edges,which not only reduces the model ’s demand for computing resources,but also aggregates two local features that are far apart but have a large relationship.Experiments based on measured data show that the effectiveness of extracting HRRP structure features based on this model,and the recognition accuracy of the model is further improved.(4)In order to make the model make better use of the common features between samples,this paper proposes a new graph neural network based on multi-feature fusion(DCGT-GE)from the common features between samples and the structural features within samples.The network first uses the graph embedding(GE)representation method to extract the common features between samples and uses the DCGT model to extract the structural features within the sample,and then fuses the common features and structural features to obtain a better feature representation.Finally,the recognition result is output by the classifier.Experiments on measured data show that the common features between samples and the structural features within samples are complementary,and the fusion of the two improves the recognition accuracy.(5)In order to make the model proposed in this paper can be applied to embedded devices with low computing power to achieve specific algorithm engineering goals,this paper builds a software and hardware platform for radar HRRP recognition based on i TOP-3588 development board and RKNN inference engine,and based on this platform,the model is accelerated on edge devices.The experimental results verify that the HRRP recognition method proposed in this paper also has high real-time performance and noise robustness on low-power embedded devices,which lays a foundation for the future application of the algorithm in practical engineering.