Active-jamming Recognition Method Of LFM Radar Based On Deep Learning

Since its inception,radar technology has been used widely on the battlefield.Radar active-jamming can blind radar,thus putting the enemy in a position of initiative in the war.Taking appropriate anti-jamming measures is the key to reclaiming the initiative in warfare.The first step against jamming is to accurately recognize jamming patterns.Being the passive party,the development of radar anti-jamming technology normally lags behind the advancement of jamming technology in the complex electromagnetic environment of modern warfare.However,Traditional methods based on feature extraction and machine learning rely on expert knowledge and manual feature selection.The process of manually extracting features is time-consuming,and the lack of features that can distinguish between dominant jamming patterns makes it difficult to meet the requirements.Therefore the intelligent recognition of radar active-jamming is of great practical significance and is a guarantee to accelerate the modernization of national defense forces and military construction and to build the people’s army into a world-class army.Deep learning,especially convolutional neural network,as a popular direction in the current artificial intelligence field,has proved its feature mining ability better than traditional methods in many recognition tasks.In this thesis,we study the radar activejamming recognition method based on deep learning,comparing with the traditional jamming recognition method based on feature extraction,eliminating the complicated preprocessing process and realizing the end-to-end recognition of radar active-jamming patterns,and achieving good recognition performance.The research work of this thesis has the following aspects:1.Mathematical modeling and waveform simulation of suppression jamming,deception jamming and smart noise jamming based on LFM high resolution radar are performed to compose a radar active-jamming data set containing seven active jamming patterns,and preprocessing is performed as the data base for subsequent training and testing of classification algorithms.2.A total of 10 feature parameters of the jamming signals are extracted from the time domain,frequency domain and time-frequency domain.By manually analyzing the differentiation degree of the feature parameters and the sensitivity to noise,six of them are selected to form a feature vector and input to a multi-classification SVM to train a classification model as a benchmark method for subsequent analysis and comparison.3.A RNN model for radar active jamming recognition is proposed.By effectively extracting bi-directional features for input data,the Bi GRU network can make full use of the contextual information of radar active-jamming time sequences,eventually achieving better accuracy than the traditional method.4.A radar active-jamming recognition model based on 1DCNN+Bi GRU is proposed.In order to more effectively combine the spatial feature extraction capability of CNN and the memory capability of RNN,the recognition accuracy of the cascade model and the parallel model were compared experimentally.The experimental results show that the cascade model achieves better performance on the dataset of this thesis,with an average accuracy of 97.71% for 7 radar active-jamming at-10～20dB JNR.