Research On Multifunction Radar Behavior Prediction Based On Deep Learning

Multi-Function Radar(MFR)is playing an increasingly important role in modern warfare,and knowing the working state and behavioral intention of the enemy MFR at the next moment in advance can dominate the radar countermeasures.In view of the powerful predictive capability of temporal deep learning models,it has become an important element in radar countermeasures research to study how to use deep learning techniques to uncover the behavioral transition patterns of enemy MFRs and further realize the cognitive empowerment of electronic intelligence reconnaissance systems.In this thesis,based on the existing radar word,phrase and sentence multi-level MFR signal models,we explore the MFR behavioral characterization methods applicable to the prediction task,and then focus on the deep learning prediction methods under two different time scales of MFR signal sequences and working patterns.The main work and contributions of the full thesis are as follows:(1)To address the problem of insufficient ability to capture long-range dependencies of existing MFR signal sequence prediction methods,a Transformer signal sequence prediction model based on a self-attentive mechanism is proposed,which has a It has stronger timing prediction ability than the existing Markov chain(MC)and Long-Short Term Memory(LSTM)models,and there is no backward and forward timing limitation in processing the input sequence,which can realize parallel processing.(2)In order to solve the problem of degradation of the representation that the sample embedding vectors are ”crowded” in the same narrow space,we add contrast learning as the regularization loss of the model,in which the uniformity and alignment of the nature of contrast learning improve the original pathological sample embedding distribution.Transformer prediction model to further improve the correct prediction rate of low frequency high threat MFR signals.(3)The proposed working mode prediction method based on deep hidden variable modeling can improve the accuracy of working mode prediction by estimating the potential radar behavioral intention hidden variables affecting working mode switching from MFR signal sequences and using the estimated hidden variable distribution to further assist the working mode prediction.In a typical airborne MFR ”air-surface” mission scenario,the prediction accuracy is improved from 77.45% to 85.46% compared with the direct cross-entropy optimization method,which ignores the influence of intention hidden variables.The research results of the thesis have been validated by simulation experiments,which are further extensions of the existing MFR behavior prediction methods.The proposed deep learning prediction models provide ideas for the study of MFR behavior in more complex scenarios such as dynamic games,and can provide technical support for strategic operational decisions in the closed loop of OODA operations.