Distributed Machine Learning Algorithms For Electromagnetic Targets Identification

Electromagnetic targets identification(ETI)is an important research topic in the fields of electromagnetic cognition.Artificial intelligence has developed rapidly in recent years,so the ETI based on the neural network becomes a popular approach.Currently,it is mainly carried out under centralized conditions.However,the enrichment of electromagnetic targets(ET)types and the complexity of the electromagnetic environment cause great pressure on the storage and computing resources for centralized training.Then the ETI tasks urgently need the participation of distributed machine learning.Therefore,machine learning algorithm and distributed computing method are integrated in this thesis to solve the ETI problem efficiently,and the main work of this thesis is listed as follows:(1)A classification and identification algorithm for multiple same-model ETs is designed.The algorithm uses neural network to learn the time-frequency fingerprint features of ETs,and realizes the end-to-end identification.By comparing the performance of different data preprocessing methods and neural networks,the centralized identification approach is determined.Experimental results show that the proposed algorithm can efficiently converge,and the identification accuracy is above 96%.(2)Two data parallel based distributed machine learning ETI algorithms are designed.Aiming at the problem that the size of centralized training data gets larger and larger,the algorithms divide the global dataset into several small parts and distribute them to multiple nodes.All Reduce and Ring All Reduce strategies are used to aggregate model parameters and update the gradient.Experimental results show that the time efficiency of distributed training with data parallel under the two strategies is at least 44% and 65%higher than that of centralized training.(3)A decentralized machine learning ETI algorithm is designed.The data parallel distributed algorithms are difficult to adapt to the decentralized scenarios,where the nodes only have partial classes of data,and the communication bandwidth between different nodes is limited.To solve the problem above,nodes only use local data to train models,and complete model aggregation by exchanging model parameters instead of data using limited communication bandwidth.After the information exchange,the all nodes will get a convergent global model.Experimental results show that the algorithm can train local models and obtain convergent global models,the efficiency is at least 36% higher than that of centralized training.