Research On The Key Technology Of Intelligent Recognition Of Signal Modulation Mode Under Sample-Constrained Conditions

With the successful application of fifth-generation wireless communication technology and extensive research on next-generation wireless communication technology,the communication spectrum environment has become increasingly complex.Therefore,many researchers have attempted to manage communication spectrum resources by investigating cognitive radio technology.Modulation recognition plays a crucial role in cognitive radio systems,enabling the automatic classification and identification of radio signals to ensure communication security and reliability.Currently,modulation recognition technology faces challenges such as limited samples,scattered and isolated data,and low signal-to-noise ratios.In view of the difficulty of traditional recognition algorithms in dealing with the recognition of complex electromagnetic signals under sample-limited conditions,this thesis explores the key technologies for intelligent recognition of electromagnetic signal modulation under sample-limited conditions.We investigate topics including recognition with limited samples,zero-shot recognition,distributed modulation recognition,and semi-supervised modulation recognition under low signal-to-noise ratios.The contributions of this thesis primarily encompass the following three aspects:(1)Few-shot modulation recognition algorithm under limited-sample conditions for radio signals.To address the challenge of training wireless signal data with limited samples,this thesis proposes two solutions:distance-based recognition and sample generation using generative adversarial networks.Firstly,the distance-based method accurately identifies unknown modulation types by comparing the differences between the limited sample signals and the signals to be tested.Experimental results demonstrate that the proposed algorithm improves the classification accuracy by 10%to 50%compared to traditional methods,with the highest classification accuracy reaching 93%.Secondly,to tackle the problem of zero-shot learning in wireless signals,the sample generation approach utilizes semantic space mapping and feature encoding.Specifically,by utilizing the feature encoding of different modulation types and generating the feature vectors of modulated signals using generative adversarial networks,the training dataset is enriched.Experimental results show that in zero-shot categories,the proposed method achieves an average accuracy of over 85%,with classification accuracies exceeding 76%in both zero-shot and common categories.When two modulation signals are missing,the proposed algorithm achieves an average accuracy of over 69%in the zero-shot category.(2)Distributed modulation recognition under few samples conditions.A federated transfer learning framework is proposed for the training challenges of each node in a distributed context.In order to reduce the data dependency of each recognition node,pre-trained modulated modality recognition depth models are sent to each node for fine-tuned transfer learning.Then,each node uploads the updated model weights to the server,which accumulates the received weights and averages them to obtain the new weights,and then distributes this weight file to each node,which updates the weights to achieve less-sample learning under distributed conditions.The experimental results show that the federated transfer learning algorithm improves the classification accuracy of each node by 2%to 7%with less training data compared to the centralized training method.(3)Semi-supervised wireless signal modulation recognition.A semi-supervised automatic modulation mode recognition network is designed to achieve label-free training at low signal-to-noise ratios for the label-free problem.The proposed algorithm can be adapted to another low SNR domain without the corresponding labeled data pre-training.The source domain is mapped by the source encoder to the classification domain and can be classified by the classifier,where the labelled source data enables the process.The discriminator and target encoder are then trained by determining whether the data is from the target or source domain.Next,the target encoder and classifier are used to infer the target domain data labels.Finally,it is verified that the domain adaptation adjustment algorithm proposed in this thesis improves the classification accuracy of the target domain by 3%to 27%in a signal-to-noise ratio environment of-20 dB to-4 dB.Extensive experimental show that this thesis provides an effective solution for cognitive radio by considering the sample-constrained scenario of wireless signals,with high accuracy and robustness of sample less recognition,zero sample recognition,distributed sample recognition,and low signal-to-noise ratio semi-supervised recognition under sampleconstrained conditions.