| As a significant research direction in the field of electromagnetic cognition,electromagnetic target identification has been widely used in the field of military,civilian,and industrial manufacturing.With the rapid development of electronics and information technology,the electromagnetic environment has become increasingly complex.Many achievements have been made by utilizing machine learning for the identification and classification of electromagnetic target signals.However,centralized training for large and complex electromagnetic target identification tasks suffers from issues such as insufficient timeliness and high computing power requirements.Therefore,this thesis conducts research on distributed machine learning algorithms for electromagnetic target identification to achieve fast and efficient distributed electromagnetic target identification algorithms.The main contributions of this thesis are as follows.1.A centralized identification algorithm for electromagnetic targets is designed.For actual radio and mobile phone signals which have been data-preprocessed separately,the fingerprint features are extracted using the short-time Fourier transform and HilbertHuang transform,and then a dataset is constructed.Then,the convolutional neural network and residual network are used to complete the identification tasks for the two types of electromagnetic signals,and the algorithm performance of feature extraction and identification classification is compared.Experimental results show that identification performance of the combination of short-time Fourier transform and residual network is better,with identification rates of 99.63%and 89.49%for the two datasets,respectively,demonstrating the effectiveness of the centralized electromagnetic target identification approach.2.A distributed electromagnetic target identification framework based on alternating direction method of multipliers(ADMM)is designed.In order to solve the problems of insufficient computing resources in centralized electromagnetic target identification and poor convergence and low time efficiency of traditional stochastic gradient descent-based distributed algorithms,a distributed identification algorithm based on ADMM is proposed based on distributed optimization theory.Experimental results show that the ADMMbased identification framework has better convergence and higher time efficiency.Compared with the stochastic gradient descent-based distributed algorithms,this algorithm has a time efficiency improvement of 20.71%and 18.06%for the two datasets,respectively.3.Two improved versions of the ADMM-based distributed electromagnetic target identification algorithm are designed.Specifically,for large-scale distributed clusters,the Ring-AllReduce communication mode is designed for the distributed ADMM algorithm.The RA-ADMM algorithm based on ring structure effectively solves the problem of communication bottleneck in the master-worker structure,and the time efficiency of the two datasets is improved by 37.22%and 24.20%,respectively.Then,for scenarios where the distributed cluster nodes are heterogeneous,an asynchronous communication mechanism-based ADMM algorithm is designed,and convergence conditions are given.The improved asynchronous algorithm effectively solves the communication waiting problem caused by slow nodes,and the time efficiency of the two datasets is improved by 17.73%and 20.54%,respectively.The results show that the improved algorithms can better handle the distributed electromagnetic target identification tasks in complex and variable electromagnetic environments. |
