Research On Lightweight Multimode Emitter Individual Recognition Based On Deep Learning

Specific Emitter Identification is a key link in obtaining important intelligence in cognitive electronic warfare.Traditional methods complete the identification of radiation source individuals by measuring their fingerprint characteristic parameters.Currently,methods based on deep learning can automatically extract deeper individual characteristics of radiation sources and are widely used in the field of radiation source individual identification.However,radiation sources today can work in multiple modes and intelligently switch,which poses challenges to the deep learning-based individual identification of radiation sources,such as the requirement for a large number of training data samples and the need for a large-scale network to ensure the corresponding recognition accuracy.To address these challenges,this article conducts research on model embedding and lightweight compression methods,with the main work and innovations as follows:A model embedding-based end-to-end multi-mode individual identification method is proposed.This method combines the time-frequency feature extraction algorithm with Bayesian fusion theory,improves the dual-path network and bidirectional dual-constraint ranking loss function,and realizes the sharing of model parameters between different modes of the same radiation source to extract common individual characteristics of the same radiation source under different working modes.Experimental results show that the proposed algorithm is effective.Under the same conditions,compared with traditional methods,the identification accuracy of the algorithm proposed in this thesis reaches97.45% without using Bayesian decision fusion,which improves by 5%.After using Bayesian decision fusion,the identification accuracy reaches 99.73%,which further improves by 2%.A network optimization method based on a combination of various lightweight techniques is proposed.By comparing the effects of different lightweight techniques,including convolution structure optimization,model parameter quantization,and model pruning,on the model size,a method combining these three lightweight techniques is proposed.This method is applied to the Res Net50 model,and pre-training with group convolution,QAT quantization,and L2 weight pruning(sparsity of 0.5)are performed.Under the premise that the accuracy loss is not significant,this method achieves a compression rate of 94% on the model size.Experimental results show that,under the same input data,compared with the traditional lightweight network Mobile Net V3(with a recognition accuracy of 80.35% and a model size of 12.11MB),the optimized Res Net50 model has a recognition accuracy of 91.55% and a model size of only 5.90 MB.A lightweight model training method based on sparse optimization is proposed.From the perspective of sparse optimization,this method replaces the Adam method with the alternating direction method of multipliers(ADMM)during fast training and improves the model’s objective function by adding a penalty factor.Through sparse training of the model,the complexity and recognition accuracy loss of retraining after lightweighting(pruning)are reduced.Experiments show that compared with the traditional training method that uses cross-entropy loss,the proposed method can increase the recognition accuracy from 12.50% to 65.51% after pruning.After 20 epochs of retraining,the accuracy can reach 93.97%,which increases by 2.64%.At the same time,it exceeds the 93.84%accuracy achieved without using this method for 40 epochs of retraining,and the time required to restore accuracy is significantly shortened.