Research On Distributed Machine Learning For Specific Emitter Identification

With the rapid growth in the number,type,and updating speed of emitter equipment,the electromagnetic environment presents highly complex and dynamic characteristics.Traditional specific emitter identification methods are unable to meet the current needs,and although intelligent machine learning methods have achieved certain results,most of them are centralized reasoning methods,while the current radiation source signals are many types,large in number and widely distributed,which causes problems such as low efficiency of centralized reasoning,high computational capacity requirements and large storage capacity requirements.Therefore,this paper focuses on the specific emitter intelligent identification method and the distributed specific emitter identification method.Among them,the specific emitter intelligent identification method includes dataset generation and neural network classifier module,and based on this,this paper researches and proposes different distributed specific emitter identification methods based on different distributed optimization methods and lightweight method,and constructs datasets for experimental validation using the actual harvested communication transmitter data.The main work and innovations of this paper are as follows:(1)A centralized specific emitter identification method and a classical distributed specific emitter identification method are designed.Firstly,based on the data of communication transmitter,the datasets are constructed based on the Short-Time Fourier Transform and the Continuous Wavelet Transform to extract the time-frequency features respectively.Then,different recognition models are used to complete the training and testing of specific emitter identification in centralized and distributed scenes respectively.The experimental results on the dataset of this paper show that the Short-Time Fourier Transform time-frequency method is better than the Continuous Wavelet Transform method;the Residual Neural Network has the best identification effect;the distributed method can effectively improve the training efficiency and reduce the memory requirement of a single machine.(2)To address the problems of slow convergence of models and reduced identification effect in distributed scenes for data parallelism,two improved distributed specific emitter identification methods are proposed in combination with distributed optimization methods.First,a distributed specific emitter identification method is proposed based on the Decentralized Stochastic Gradient Descent method and combined with the Momentum method.Second,based on the Alternating Direction Multiplier Method,a distributed specific emitter identification is achieved and the generalization ability of the local model is improved by introducing regular terms with global pairwise variables to effectively aggregate the models at each node.The experimental results show that the model convergence speed can be accelerated by both methods and the identification accuracy is improved by more than 2.2% and 4.5% in distributed scenarios,respectively.(3)Based on the Knowledge Distillation method,a lightweight distributed specific emitter identification method is proposed.For the problem of limited performance of a single working device in some scenes,data are loaded in the form of data streams and combined with the Knowledge Distillation method for distributed training,which achieves large performance for small model,further compresses the identification model and reduces memory consumption.The experimental results show that the identification accuracy is improved by more than 22% for the student model and is able to be deployed for testing on Jetson edge devices.