Research On Radar Intra-Pulse Modulated Signal Identification Based On Self-Attention Feature Fusion Network

The intra-pulse modulated signal recognition of radar emitter is one of the important technologies in electronic warfare,aiming to provide technical support for our electronic warfare and ensure its smooth progress.Currently,most radar emitter signal recognition technologies have achieved good recognition results,but the complex and changing electromagnetic environment still makes the technology unable to meet the high-precision recognition requirements under low signal-to-noise ratio(SNR).In addition,most existing algorithms were proposed under the condition of sufficient data samples,ignoring the problem of small signal sample sets and few labeled data,resulting in algorithms lacking practical application value.To better apply radar emitter signal recognition algorithms to practical electromagnetic environments,this paper studies the radar emitter signal recognition algorithm from two aspects: signal features and network structure.The main work of this paper is as follows:1.In this paper,feature extraction is performed on radar radiation source signals,and the time-frequency images and conventional parametric features of the signals are extracted separately.Due to the wide variety of traditional parametric features that can be extracted from the radar radiation source signal,a feature importance ranking algorithm is proposed to filter the features with stronger noise immunity and more favorable signal classification by comparing the influence of features on the signal stratification degree.In this way,seven conventional parametric features of the radar radiation source signal are selected from the frequency domain and bispectrum,and a 7-dimensional feature vector is composed for the subsequent signal identification task.2.An Improved Multi-headed Self-Attention Network(IMSA-Net)is designed to address the problem of low recognition rate of radar radiation source signals under low SNR.The network uses residual connectivity as the base architecture of the network,and firstly,the improved multi-headed self-attention mechanism algorithm is introduced to take the timefrequency image of the signal as the input of the network,which improves the capture of the global features of the time-frequency image during the training process,thus improving the signal recognition accuracy under low SNR.Second,the proposed activation function with parameter adaptive learning(A-LeakyReLU)accelerates the convergence of the network and improves the stability of the network.Experimental results show that the proposed network can achieve high accuracy in recognizing 10 typical radar radiation source signals at a SNR of-10 dB.3.For the problems of incomplete signal information extraction by single feature,poor recognition of multiple types of signals,and small number of samples of data sets in the actual electromagnetic environment,a Multimodal Feature Converging Network(MFC-Net)is designed based on the IMSA-Net network.The network takes multimodal features as the input of the network,i.e.,the time-frequency images of signals and traditional parametric features,and improves the representation capability of the network for signals through feature fusion,which further improves the recognition accuracy of each type of radar radiation source signals under low SNR.In addition,a random matching mechanism is proposed to randomly match the time-frequency images of signals and traditional parametric features within classes during the training process,which disguisedly expands the number of signal samples in the dataset and satisfies the practical application significance.The experimental results show that the proposed network can further improve the recognition accuracy of signals in each class at-10 dB and ensure a high recognition accuracy even with a limited number of signal samples.4.To address the problem that actual electronic warfare contains only a very small amount of labeled data,a novel semi-supervised learning model(All Mean teachers,AMT)is designed to closely match the radar radiation source signal recognition task,and a student vector is added to the framework of the Mean teachers model specifically to calculate the contrast loss of timefrequency images during the training process.This loss term directionally transfers the subtle features of the time-frequency images from the teachers’ network to the students’ network,which acts as a knowledge distillation and solves the problem that the time-frequency images are not easily trained in a semi-supervised signal recognition task.The experimental results show that the proposed semi-supervised learning model can effectively improve the recognition accuracy of radar radiation source signals and has better performance compared with other semi-supervised models.