Research On Automatic Modulation Classification Based On Signal Capsule Network

Automatic Modulation Classification(AMC)of communication signals is widely used in military and civilian fields such as battlefield communication reconnaissance,countermeasures and radio spectrum monitoring.Recently,deep learning has been applied in AMC,which overcomes the defect of "feature engineering" that relies on expert experience in traditional AMC methods.However,most of the existing deep learning AMC methods rely on massive labeled data,and require strong consistency between the decisionmaking scene and the training scene,and have limited application in the complex electromagnetic environment.In response to this problem,this paper draws on the modular structure of the brain and the characteristics of automatic navigation cognition,encapsulates multiple neurons into capsules,and constructs a new signal capsule network(SCN)model.SCN uses modular signal capsules instead of simple neurons,adaptive routing instead of fixed pooling,and vector output instead of single prediction,which can achieve better generalization performance under less data volume.Specifically,this paper proposes three signal capsule network models to solve the AMC problem:1.A modulation signal recognition method based on Complex Signal Capsule Network(CSCN)is designed.Aiming at the problem of low recognition accuracy when the communication modulation signal has a low SNR,a complex convolution module was designed to jointly process the IQ dual-channel data of the modulation signal.Then,the complex convolution features obtained were constructed into signal capsules to form the main capsule layer,and then cascaded into digital capsule layer.Since complex convolution can extract rich characterization features of IQ data,the CSCN model can make full use of these features and achieve sufficient generalization performance.The proposed method was verified and compared on RML2016.10 a and RML2016.10 b,and the experimental results show that,compared with other deep networks,the CSCN model has an improvement of0.8%～6.92% and 4.64%～6.4% at the signal-to-noise ratio(SNR)of-6～10d B under the condition of only 10% training data.2.A modulation signal recognition method based on Time-series Attention Signal Capsule Network(TASCN)is designed.Aiming at the problem that the signal timing information was not fully excavated in the previous work,a signal capsule network based on LSTM unit and feedforward attention mechanism was designed.Firstly,the amplitude and phase information of IQ data were obtained by means of polar coordinate transformation,and then the temporal features were extracted by LSTM.Then,the feed-forward attention module was constructed for adaptive aggregation of features,and the classification results were output through the main capsule layer and the digital capsule layer.Since IQ dual channels contain temporal correlation information,temporal attention signal capsule network can make full use of the temporal correlation information to enhance the recognition performance of the model.The proposed method was verified and compared on RML2016.10 a and RML2016.10 b,and the experimental results show that compared with other deep networks,the TASCN model has an improvement of 11.16%～18.98% and15.18%～17.12% at the signal-to-noise ratio(SNR)of-6～10d B under the condition of only 10% training data.Compared with the method in the previous chapter,the improvement is 10.43% and 10.07%respectively.3.A modulation signal recognition method based on Siamses Bidirectional Time-series Signal Capsule Network(SBTSCN)is designed.Aiming at the problem that the capsule network could not obtain relatively accurate AMC results under the condition of few data,a bidirectional timing signal capsule network with twin structure was designed.Firstly,the bidirectional LSTM network is used to increase the extraction of time-sequence correlation information of data context;secondly,the siamese signal capsule network is designed to extract the discriminant features of the modulated signals through the twin structure;finally,the extracted features are classified through the random forest.Since the input of the siamese structure is in the form of data pairs,the number of data pairs can be increased by using as little data as possible through different combinations,and the siamese bidirectional time-series signal capsule network can extract effective features from less data.The proposed method was verified and compared on RML2016.10 a and RML2016.10 b,and the experimental results show that compared with other deep networks,the SBTSCN model has an improvement of 10.84%～23.79% and 10.13%～52.25% at the signal-to-noise ratio(SNR)of-6～10d B,respectively,under the condition of only 5% of the training data.