Research On Recognition Of Carrier-free UWB Radar Ground Targets

The high-range resolution characteristic of carrier-free ultra-wideband radar(UWBR)makes the target echoes contain rich target structure information,which can describe the outline and details of the target,which is of great significance for the automatic recognition of radar targets.However,radar target recognition based on high-resolution data has the problem of amplitude-scale sensitivity,which will affect the target recognition performance of the model.In addition,single-view observations carry limited information,which limits the improvement of model recognition performance to a certain extent.For this reason,Therefore,based on the carrier-free UWBR,the thesis considers practical application scenarios such as low signal-to-noise ratio(SNR)and non-full-labeled samples,and integrates multi-view observation information to conduct in-depth research and practice on the automatic recognition of ground targets.The main research contents are summarized as follows:1.The thesis first studies UWB electromagnetic scattering characteristics for targets.Based on the finite integration technique(FIT),the thesis calculates the UWB electromagnetic scattering field for different targets,analyzes the main scattering mechanisms,and gives approximate distributions of the strong scattering centers.As a result,the electromagnetic scattering characteristics of the target are described from different angles.Experiments show that FIT has the ability to calculate the transient electromagnetic scattering field for complex targets,and the target response can be modeled as the coherent superposition of multiple scattering structure responses.The above research has laid a foundation for subsequent target recognition research.2.Aiming at the problem that the amplitude-scale sensitivity will affect the recognition performance of the traditional KSVD algorithm,an improved KSVD(IKSVD)algorithm is proposed.The improvement of IKSVD is reflected in two aspects: 1)the traditional KSVD algorithm only considers the sparse representation of signals,resulting in the learned dictionary lacking distinguishability.For this reason,the improved algorithm improves the dictionary update procedure,making dictionary atoms have category attributes and distinguishability in the dictionary update stage;2)propose hierarchical coding constraints to replace the traditional consistent coding constraints,which can adaptively assign sublabels to the same type of samples according to the similarity between target echoes,relaxing the amplitude-scale sensitivity problem to a certain extent.The experimental results show that the algorithm can significantly improve the separability of dictionary atoms.When SNR is 20 d B,the target recognition accuracy of IKSVD is improved by about 3% compared with the traditional KSVD algorithm.3.Aiming at the problem that the traditional dictionary learning algorithm cannot guarantee the separability of dictionary atoms under the environment of low SNR,a hierarchical dictionary learning(HDL)algorithm is further proposed.HDL learns and optimizes dictionary atoms in two levels.Aiming at target recognition,the first level learns several signal sub-dictionaries with category attributes to ensure the distinguishability of dictionary atoms;the second level aims at a low SNR environment,learning the noise dictionary.As a result,an identifiable and robust dictionary can be learned.In addition,a "dictionary atom selection" mechanism is added to the learning process to improve model efficiency.Compared with other dictionary learning algorithms,the dictionary learning process in HDL does not depend on any regular term constraints,simplifying the system program and improving the interpretability of the model.When the SNR is only-10 d B,compared with traditional dictionary learning algorithms,the target recognition accuracy of the algorithm is improved by more than 10%.4.Aiming at the problem that the deep learning model cannot extract separable and robust representations through sufficient label information in the case of low SNR and nonfull-labeled samples,an improved multi-task self-supervised learning(IMTSSL)algorithm is proposed.IMTSSL integrates the contrastive learning network and the stacked convolutional denoising autoencoder network,which can simultaneously perform contrastive learning and signal denoising tasks.In the comparative learning task,a data enhancement method aiming at the carrier-free UWBR target echo data is proposed,which can improve the robustness of the model to azimuth changes.Experimental results show that the model is capable of learning robust,separable representations from large amounts of non-full-labeled data.When SNR is 10 d B and the proportion of labeled samples is only 20%,the model has a comparable recognition performance with the supervised learning model.When the SNR is only-10 d B and the proportion of labeled samples is only 20%,the target recognition accuracy of IMTSSL is improved by about 10% compared with the supervised learning model,having better anti-noise performance.5.Aiming at the problem that the single-view observations carry limited information,which leads to the performance limitation of models based on single-view observations,target recognition models based on multi-view observations are proposed.Consider the following multi-view observation scenarios: 1)multi-view samples are obtained through continuous observation in a small angle range;2)multi-view samples are obtained through random observations from all target directions,and there is no angle constraint between adjacent observations.For the first observation scenario,the bi-dimensional variational mode decomposition algorithm is introduced to extract image contour features.The obtained two-dimensional feature map is used as the input of the subsequent neural network,realizing image-based feature extraction and target recognition.For the second situation,a multi-view multi-task self-supervised learning(MV-MTSSL)algorithm is proposed.MV-MTSSL builds a contrastive learning model based on multi-view observations,which aims to extract representations that are robust to attitude changes from a large number of unlabeled multiview observations.The experimental results show that the effective use of multi-view observations can achieve feature enhancement and significantly improve the target recognition performance of the model.When SNR is only-10 d B,compared with the target recognition model based on single-view observation,the target recognition accuracy of two proposed models are improved by more than 8%.