Research On Radar Additive Composite Jamming Recognition Based On Meta-learning

The current electronic warfare environment is becoming increasingly complex,with multiple signal modulation types,high signal density,variable signal parameters,and increasing overlap in time-frequency domains.Composite jamming is gradually becoming the main means of attacking radar receivers,which has led to traditional radar jamming recognition methods being unable to meet current radar anti-jamming needs.In response to the difficulty of effectively identifying composite jamming,this thesis focuses on intelligent recognition of radar composite jamming.The specific work overview is as follows:1.The generation mechanisms and mathematical models of noise suppression jamming and deception jamming,including velocity deception jamming,noise amplitude modulation jamming,distance deception jamming,noise frequency modulation jamming,and Smeared Spectrum are analyzed.Additive combination of noise suppression jamming and deception jamming is simulated to generate a dataset of radar additive composite jamming signals.2.Considering that traditional feature extraction methods have obvious limitations in recognizing composite jamming,and deep learning methods are often limited by the need for a large number of training samples,a dense convolutional network model fused with efficient channel attention mechanism is proposed to address this problem.The model is also improved with modelagnostic meta-learning algorithms to update the model parameters,enabling the model to quickly converge with insufficient samples and complete the recognition of composite jamming.Experimental results show that the proposed algorithm achieves average recognition rate of 98.73%for composite jamming under various jamming-to-noise ratios,and even when the sample size is reduced to one-fourth of the original,the average recognition rate still reaches 95.35%,validating the effectiveness of the algorithm.3.A self-calibrating prototype network is proposed to address the issue of class prototype effectiveness in traditional prototype network models.Compared to traditional prototype networks that simply use support set data to calculate class prototypes for predicting the types of query set samples,the proposed self-calibrating prototype network also predicts the support set samples in the same manner to obtain prediction losses,and then uses the prediction losses to update and calibrate the class prototypes,thereby improving their effectiveness.Meanwhile,the use of Mahalanobis distance as the measurement function strengthens the correlation of feature vectors in various dimensions.Finally,a mixed-domain attention mechanism is added to the model’s feature extraction structure to further enhance class prototype representation.Experimental results demonstrate that the proposed algorithm improves the performance of the traditional prototype network by an average of4.51% under a jamming-to-noise ratio of 3d B,confirming its effectiveness.