Research On Active Jamming Identification Method In Complex Electromagnetic Environment

Since the advent of radar,the control of electromagnetic power in the electronic battlefield,as an important resource of modern military operations,has always been the target of both combatants.Radar is facing increasingly serious electromagnetic interference.At present,with the investment of scholars in electronic countermeasure research,radar jamming technology has developed rapidly.In order to enable the radar system to quickly modulate and forward the jamming signal after intercepting the enemy’s radar signal,the Digital Radio Frequency Memory jamming technology was born.Radar jamming signal identification is the key to radar anti-jamming steps.Although the previous radars have some anti-jamming functions,the lack of efficient radar jamming automatic detection methods is still not enough to complete the radar anti-jamming goal.For radar active jamming identification,traditional identification methods have limited types of identification,large amount of calculation for feature extraction,and low identification performance in noisy environments,which makes it difficult for jamming identification to meet the requirements of anti-jamming.This paper takes radar active jamming identification as the research object,and carries out specific research.The main research contents are as follows:(1)Aiming at the problems of large amount of calculation and complex model in feature extraction of conventional methods,an active blancket jamming identification method based on time-domain feature and optimized weighted random forest model is proposed.The method first extracts multidimensional features such as statistical features,instantaneous features,and entropy features of jamming signals in the time domain to form a dataset.In order to improve the generalization ability of the random forest,the Correlation similarity is used to evaluate the correlation degree of the decision tree in the classification performance,and the decision tree is weighted after removing the low precision decision tree.Finally,the weight is further optimized through the particle swarm optimization.Experimental results show that the proposed algorithm has higher recognition rate and better noise robustness than other improved random forest and other jamming recognition methods.(2)Aiming at the problems of strong subjectivity and large uncertainties in the jamming identification under manual features,a jamming identification network using time-frequency distribution map as input is proposed.Firstly,a variety of time-frequency analysis methods were analyzed and compared,and the PWVD was selected to generate the time-frequency image of the jamming signal.After that,considering that the local features of the time-frequency distribution map are easily overwhelmed by the noise features in the noisy environment,the dilated convolution is introduced and the dilated convolution module with increasing dilation rate is constructed to perceive the local features.At the same time,in order to allow the network to focus on the important features in the two dimensions of channel and space in the time-frequency distribution graph,channel attention and spatial attention mechanisms are respectively introduced to further improve the feature extraction ability of the neural network.The simulation results show that the recognition rate of the proposed network is better than other networks under low JNR,which provides a new idea for the recognition of radar active jamming signals in strong noise environment.(3)In order to adapt the above network to devices with insufficient computing power,a model compression method based on network pruning and knowledge distillation is proposed.First,use the dynamic network pruning method to prune the connection.Compared with the static pruning method,dynamic pruning can reduce the amount of model parameters more safely,and has advantages in real-time.Then,the above-mentioned time-frequency image recognition network is used as the teacher network,and soft targets containing wrongly predicted category information are generated through the teacher network.Finally,the soft target and the hard target jointly supervise the pruned network for learning.Experimental results show that the proposed lightweight network outperforms other lightweight networks with higher parameters in terms of recognition rate,and the recognition performance is close to that of the teacher network.