Radar One-dimensional Range Profile Target Recognition Based On Transformer And Contrastive Learning

Radar one-dimensional high-resolution range profile(HRRP)reflects the vital information of the target,such as size and structure signatures.It also has the advantages of small calculation burden,easy data acquisition and processing.Therefore,HRRP has gradually become a hot topic in the radar automatic target recognition(RATR)field in recent years.Most of the traditional HRRP-based radar target recognition methods simply treat the radar signal as an overall envelope,ignoring the sequence correlation embedded in the HRRP which reflects the inherent physical structural features of the radar target.In experiments based on the measured data,we found that this method not only limits the recognition accuracy and generalization performance of the model,but also is more prone to the sudden drop of recognition performance when the number of training data samples is small,which is not conducive to the practical engineering application of the algorithm.In order to improve the defects of past methods,this thesis proposes new deep learning methods for HRRP target recognition from the potential sequence correlation within HRRP,and the main elements can be summarized as follows.(1)A unified data preprocessing process is designed to overcome the sensitivity of HRRP samples to reduce the difficulty of subsequent classifier design.Meanwhile,the principles and implementation processes of several typical HRRP recognition methods are analyzed to demonstrate the algorithmic process framework in this target recognition domain and pave the way for comparative experiments later on.(2)An HRRP radar target recognition model is proposed based on temporal convolutional network(TCN)and multi-level attention mechanism.The model firstly extracts the local features of HRRP by TCN to avoid the problem of high redundancy of the acquired input sequences due to sequence slicing in the data preprocessing process in traditional RNN;then,based on TCN,a stacked residual module is constructed to further learn the serial correlation among the local features;finally,a multi-level attention mechanism is introduced at the back end of the model to adaptively improve the recognition performance.Experiments based on measured data show the effectiveness of TCN and multi-level attention mechanism and the superiority of the model in recognition performance compared with previous methods.(3)This thesis proposes an MCNN-Transformer model based on multi-scale convolutional neural networks(MCNN),SE module,and Transformer encoder.The model first uses the multi-scale convolution module to extract the multi-level spatial structure features of HRRP,and automatically adjusts the weights of each channel through the SE layer to highlight the channels with stronger differentiability,so as to obtain the features that can reflect the local spatial structure of the HRRP target for subsequent network input;then,the Transformer encoder is used to model the HRRP features temporally and capture the bi-directional correlation within the HRRP Finally,the attention mechanism is introduced at the back end of the Transformer encoder to further highlight the features with strong differentiability for subsequent classification.Compared with the TCN-based HRRP recognition model proposed in the previous thesis,this method not only considers the bi-directional correlation of sequence timing in sequence modeling,but also establishes the long-range dependency between local and global features,and obtains better recognition results.(4)This thesis further introduces the contrast learning framework to propose the TransformerTCN model based on the previous one.First,the contrast learning framework can close the distance between HRRP samples of the same types in the feature space and farther the distance between HRRP samples of different types,so as to obtain more distinguishable features for recognition and improve the feature extraction ability of the model;moreover,the new loss function is introduced in the model to control the feature distribution to achieve better recognition results.The experiments based on the measured data show that the features extracted by the contrast learning framework have strong cohesiveness,and the application of the features can further improve the recognition accuracy of the model.(5)Based on the Jetson Xavier NX board,this thesis implements a set of hardware framework for HRRP target recognition.Through this framework,we verify the real-time and noise robustness of the model proposed in this thesis and realize the deep learning algorithm on small edge devices,which provides some references for future radar hardware deployment.Finally,the experimental results based on three types of aircraft HRRP measured data sets verify that the various methods proposed in this thesis have better recognition performance under the condition of relatively complete data sets and are more robust under the extreme condition less training samples.And we verify the realizability of the proposed algorithm in the natural edge computing hardware,which lays a solid foundation for the practical application of the algorithm in the future.