Intelligent Signal Recognition Methods Based On Deep Attention Capsule

Signal recognition is a key step in intelligence reconnaissance.Traditional signal identification methods rely heavily on domain priors and experience.Not only the process is rigid,but the automation and intelligence levels are low,and the performance is severely limited by the battlefield electromagnetic environment.In this paper,the modern machine learning method is used for signal recognition.,and three new deep neural network models are constructed to mine signal characteristics and then realize end-to-end automatic signal feature extraction and recognition.The main research contents and results are as follows:(1)A deep dilated attention capsule convolution network model is proposed.Firstly,the onedimensional dilated convolution is designed to extract the hierarchical features of the original signal.Then,for the characteristics of small difference between signal classes and redundancy of depth features,a spatial attention mechanism is used to assign different weights to the features to enhance the feature representation ability.The capsule network reduces the dependence of the network on the number of samples,improves the generalization performance of the model,and thus achieves effective signal identification.We do validity tests for signal recognition on ELS,JCCM,RSS data sets: experiments show that the method can achieve a minimum accuracy of 90% on the three data sets.The method can realize signal coding,frequency parameters and modulation methods,and no manual experience and specific parameter calculations are required,proving that the effectiveness of the method.(2)Aiming at the problem that the signal timing characteristics in the previous chapter,a residual recurrent channel attention capsule network is proposed.Firstly,the timing characteristics of the signal are extracted by the gated recurrent unit to fully exploit the time correlation of the signal.Then use channel attention to model the dependence between channels to make full use of global information to achieve adaptive calibration of the channel.Finally,the capsule network is used to improve the generalization ability of the model.We do validity tests for signal recognition on ELS,JCCM,RSS data sets: the experiment proves that the minimum recognition rate of 95% can be reached on the three data sets,which proves that the method is effective in the unified identification task of signal coding,frequency parameters and modulation methods.(3)Aiming at the problem of large model parameters in the previous chapter,a light weight multi channel separable attention capsule network is proposed.Firstly,the separable convolution is used to construct the separable residual module to reduce the parameter amount of the network.Secondly,use multi-channel network structure to expand the number of network channels and enrich the diversity of convolution kernels to obtain multi-level characteristics of signals.The channel attention module is introduced to enhance the beneficial features;on the other hand,the feature representation capability is enhanced by the spatial attention module;finally,the capsule network is used to improve the generalization ability of the model.This method greatly reduces the network complexity while taking into account the high-dimensional feature extraction ability of deep learning.Compared with the previous chapter,the number of parameters of the network is reduced by 70%,and the lowest overall recognition rate is 94.5%.