Research On Recognition Of Radar PRI Modulation Type Based On Deep Learning And Extreme Value Theory

The modulation type of the radar PRI is an important means of analysing the operational status and mission of the radar and is an important basis for the recognition of radiation sources in electronic support systems.By accurately recognizing the PRI modulation type of the intercepted radar pulse signal,the threat radiation source will be more easily identified.With the development of microelectronics technology,the forms of radar signals vary,and accurate recognition of radar PRI modulation types is a key focus of research in the field of electronic warfare.Based on some issues in current research on PRI modulation type recognition,the following research has been conducted in this thesis:Aiming at the problem that the traditional PRI modulation type recognition method may have insufficient completeness in the artificial construction of PRI sequence feature set,a closed set recognition method of radar PRI modulation types based on ResNetBiLSTM-Attention is proposed in this thesis.Combined with deep learning,this method takes advantage of residual network(ResNet)and bi-directional long short-term memory(BiLSTM)to mitigate the vanishing gradient.In addition,two attention mechanisms are introduced,namely squeeze-and-excitation attention in ResNet and sequential attention behind BiLSTM,to focus on the key information of PRI sequence and improve network performance.The simulation experimental results show that the proposed method has better recognition performance under the consideration of pulse loss environment and pulse spurious environment.Aiming at the problem of unknown PRI modulation types in complex electromagnetic environments,this thesis needs to implement the open set recognition function.First,on the basis of closed set recognition,the proposed ResNet-BiLSTMAttention network model is used for feature extraction of PRI sequences.In combination with prototype learning,a loss function combining distance based cross entropy loss and prototype loss is used to train the feature extraction network.Compared with the crossentropy loss function,the features extracted from the model trained by the loss function used in this thesis are more compact in the same class of features and more separated in different classes of features.Second,in the open set recognition process,the Extreme Value Machine(EVM)model is based on the Weibull distribution in extreme value theory to obtain the decision function.In this thesis,the linear combination of known class features is introduced into the feature space to simulate the behaviour of the unknown classes during the training phase.An open set recognition method of radar PRI modulation type based on improved EVM is proposed.The simulation experiment results show that compared with EVM,the proposed method can improve the recognition accuracy of radar PRI modulation types in the case of pulse loss and pulse spurious,and has good open set adaptation in open electromagnetic environment.