Open Set Recognition Of Communication Interference Signals Based On Deep Learning

Since the electromagnetic environment is becoming more and more complex,the communication system will face more and more interference.Adopting corresponding antiinterference measures according to the identified interference signals can effectively improve the communication quality.Therefore,communication interference signals recognition will be one of the key technologies of intelligent anti-interference communication in the future and has attracted the attention of many researchers.Open set recognition algorithms for communication interference signals based on deep learning are mainly studied.The existing deep learning-based recognition of communication interference signals requires a large amount of labeled sample data.For the identification of unlabeled communication interference signals,a method for unsupervised interference signal recognition based on contrastive learning-double phases and double dimensions contrastive clustering(DDCC)has been proposed.In the first phase,a data-augmentation strategy is used to generate positive and negative sample pairs for communication interference signals,and contrastive learning based on data-augmentation is performed.In the second phase,the pre-training network obtained by the first stage of contrastive learning is used to obtain the k-nearest neighbor sample set of the original dataset and the positive and negative sample pairs are constructed.Then,a double dimensions contrastive learning based on the k-nearest neighbor sample is performed,in which the feature dimensional contrastive learning improves the network feature extraction ability,and the cluster dimensional contrastive learning completes clustering.In addition,a dynamic entropy parameter training strategy is proposed.Simulation experiments on nine types of communication interference signals show that the performance of DDCC is superior to the other five deep clustering algorithms and close to the supervised learning algorithm.Aiming at the problem of poor robustness in rejecting unknown interference signals using existing open set recognition methods based on deep learning for communication interference signals,a hollow convolutional prototype learning(HCPL)is proposed.The prototypes are updated to the periphery of the feature space by using inner-dot-based cross-entropy loss,thereby the internal space is left for unknown class samples.The intra-class compactness of various sample features is improved by using center loss,and known class sample features are allocated to the surrounding of the corresponding prototype.The impact of prototype norms on the rejection rate of unknown classes is reduced by using radius loss.A Hybrid Attention and Feature Reuse Net(HAFRNet)suitable for communication interference signals has been designed,which includes a feature reuse structure and a Hybrid Domain Attention Module(HDAM).HDAM can recalibrate both time-wise and channel-wise feature responses by constructing a global attention matrix automatically.Simulation experiments on nine types of communication interference signals show that this method achieves high unknown class rejection rates while maintaining high known class classification accuracy,has better performance than existing methods,and requires less storage resources.The existing closed set identification and open set identification of communication interference signals belong to static identification and do not have incremental identification capabilities.Therefore,a task-specific deep open set class incremental learning(TOCIL)algorithm is proposed to solve the problem of incremental recognition of open set interference signals.Firstly,a TOCIL network structure is designed,which decomposes class incremental learning into intra-task and inter-task classification problems.Each subtask is assigned an independent projection head to obtain a different embedding space.Meanwhile,the hollow convolution prototype learning method is used to classify intra-task interference signals,enabling the network to have certain open set recognition capabilities.Secondly,an open set loss constraint is adopted to solve the inter-task classification problem.Then,a memory set is constructed by using a portion of the samples in the training set,whose features are closed to the corresponding prototype,and applied to subsequent incremental training to further improve the incremental recognition performance of open set interference signals.Finally,simulation experiments on 30 types of communication interference signals show that,TOCIL not only exhibits better class incremental ability for interference signals than the four incremental learning methods that do not retain all samples in different incremental scenarios and all jamming-noise-ratio(JNR)environments,but also outperforms the five comparison methods in interference signal open set recognition performance.