Research On Deep Learning-Based Methods For Identification Of Communication Radiation Sources

With the continuous integration of the electromagnetic and cyberspace,new types of wireless services are emerging,leading to an increasingly scarce spectrum resource and growing concerns about wireless radio order and spectrum security.In the complex and dynamic electromagnetic environment,there is a higher demand for the detection and analysis of wireless signals.This has expanded from spectrum energy domain monitoring to the identification of information in the domain of multiple system communication radiation sources.It is crucial for ensuring national electromagnetic spectrum security and the development of the communication industry.The identification of communication radiation source information refers to the process of extracting features about the transmitting end from received signals,mainly involving two aspects:modulation recognition and identification of the radio frequency fingerprints of devices.Modulation recognition helps in understanding the transmission characteristics of signals and supports signal demodulation,while radio frequency fingerprints reveal the unique characteristics of hardware devices,aiding in accurate identification and tracking of radiation source equipment.In summary,the identification of communication radiation source information extends the traditional radio monitoring from the spectrum domain observation dimension to the radiation source information dimension,enhancing the ability to quickly respond to and track illegal radiation source devices.In the presence of noise interference,the individual differences in signals and radiation sources pose a challenge for the identification of communication radiation source information.Traditional feature-based recognition approaches have limited performance in these tasks.As a result,signal recognition based on deep neural networks has become a hot research topic due to its strong capability for autonomous feature representation.However,existing deep learning-based identification schemes for communication radiation source information suffer from poor real-time performance and insufficient detection accuracy.Furthermore,they overlook the similarity between modulation recognition and specific emitter identification,resulting in a lack of research on the fusion of these two tasks.This paper addresses the issues existing in the identification of communication radiation source information in wireless monitoring scenarios and conducts in-depth research.The main research objectives can be summarized as follows:(1)Modulation mode identification algorithm based on self-attentive mechanismThis thesis addresses the performance limitations and the slow inference speed of common convolutional neural networks(CNNs)and long short-term memory networks(LSTMs)in automatic modulation classification tasks.It proposes a high-performance classification model based on self-attention mechanisms.The proposed model innovatively utilizes large convolutional kernels and dense connections to capture multi-scale receptive field features.It also employs self-attention mechanisms to extract global features,achieving parallel acceleration capabilities compared to LSTM structures.Experimental results demonstrate that the performance of this model surpasses that of traditional deep learning-based optimal models,with an average accuracy improvement of 1.7%on public datasets,while achieving a 40%increase in computation speed compared to LSTM network models.Furthermore,to further reduce computational complexity and enable deployment in energy-constrained systems,a compressed version of the model is proposed.Compared to the original model,the compressed model reduces the number of model parameters and floatingpoint computations by approximately 7 times,with a loss of approximately 0.5%in model accuracy.(2)Joint identification method of radiation source information based on multi-task learningThis thesis addresses the problem of performance loss in traditional radio frequency signal monitoring systems due to the lack of feature fusion between two signal feature classification tasks.Specifically,when both automatic modulation classification and specific emitter identification tasks need to be performed simultaneously,this thesis proposes an improved approach through a new dual-task joint classification method.The dual-task model utilizes a high-capacity backbone network to learn common features for both tasks and then employs a classifier with inter-task information interaction capability to learn unique features for each task.Experimental results demonstrate that automatic modulation classification and specific emitter identification are two feature-correlated tasks,and the dual-task learning model can fully leverage this correlation to improve the detection accuracy of specific emitter identification by 2.5%while ensuring the performance of the modulation classification task.Furthermore,the computational structure of the dual-task model is the same for both tasks,simplifying the training and deployment of multi-task monitoring systems.Compared to the traditional sequential execution framework that performs each task separately,the dual-task learning framework reduces the processing time by half.Through this research,it is discovered that multi-task learning has great potential in the field of signal monitoring,where the more comprehensive the information acquired during the training process,the more accurate the signal recognition becomes.(3)Joint identification method of communication radiation source information under time-varying multipath channelConsidering the significant impact of time-varying channels on the communication process in a orthogonal frequency division multiplexing(OFDM)system,this paper proposes a communication radiation source information recognition algorithm suitable for time-varying multipath channels.Firstly,a novel subcarrier symbol information representation method is designed to suppress the interference of multipath channels on signal features while retaining the recognizability of modulation information and RF fingerprint information.Since this symbol representation sequence is inherently unordered,a point cloud feature classification network is proposed for its recognition.In the model design,a new dual-task classification model is proposed.This model utilizes attention mechanisms to extract local features layer by layer from the point cloud input and finally forms global features,providing prediction basis for two downstream classifiers.In the experiments,the model is trained only in Gaussian channels and then tested on a new dataset in time-varying channels.The results from the simulation dataset demonstrate that the model ensures the robustness of the radiation source information recognition system in time-varying multipath channels,achieving an accuracy improvement of over 40%compared to traditional recognition approaches.