Radar Signal Recognition Based On Deep Learning

With the rapid development of information technology and electronic warfare,automatic modulation recognition(AMR)for radar signals is becoming a hotspot,which is also a significant part of electronic countermeasure and reconnaissance.Conventional radar signal recognition methods mainly use shallow features such as pulse description words and then adopt template matching to achieve recognition.In the face of the increasingly complex battlefield electromagnetic environment,conventional methods always own higher computational complexity and rely on expert knowledge,and their recognition performance is hard to meet the requirements.In recent years,with the rise and wide applications of deep learning(DL),DL and neural networks have also been introduced into radar signal recognition methods.In spite of the fact that DL-based recognition methods have achieved satisfactory performance,the issue of feature redundancy and optimization of neural networks have been introduced as well.Besides,existing DL-based signal recognition methods mostly focus on one special domain transform or fine handcrafted features,which influence the recognition performance against low signal-to-noise ratio(SNR)for complex and multi-class electromagnetic signals.To address these mentioned problems,on the basis of research on the DL-based AMR for radar signals methods in recent years,and taking both intra-pulse modulation recognition and inter-pulse modulation recognition into consideration,radar signal recognition methods based on two-dimension feature images with convolutional neural networks(CNN)and one-dimension feature sequences with recurrent neural networks(RNN)are proposed in this paper.The main contribution and novelty of our works are as follows:1.Nine kinds of typical radar emitter intra-pulse modulation signals and six types of inter-pulse modulation signals have been simulated and modeled.Multi-domain features including the frequency-domain features,time-frequency-domain features,and autocorrelation features of each kind of modulation signals are extracted.By referring to the real electromagnetic environment,simulation datasets used for training,validation,and test of neural networks have been constructed.2.Aiming at radar intra-pulse modulation signal recognition,multi-domain fusion and CNN are designed.Through integrating the AC block,the Residual network,and the SE network,the recognition ability of the proposed CNN is improved.The structure of multi-domain fusion takes use of signal multi-domain features at the same time and promotes the recognition performance of the proposed method.The results on simulation signals and measured signals demonstrate that the proposed method achieves better recognition accuracy under low SNRs and outperforms the existing four methods.3.Aiming at radar intra-pulse modulation signal recognition,a method based on feature sequences and RNN is constructed.Via taking advantage of the feature extraction ability of CNN and timing processing of bi-directional long short-term memory networks(Bi-LSTM),input feature sequences are recognized,and the recognition performance of each domain feature is compared.The results on the simulation signals and measured signals prove the advantages of the proposed method at low SNRs and the proposed method is superior to the existing four neural networks.4.Aiming at radar inter-pulse modulation signal recognition,a method based on autocorrelation features and improved CNN is proposed.Compared with time-domain features and frequency-domain features,the robustness of autocorrelation features is verified.And the recognition performance under different SNRs and pulse missing rates are analyzed.The comparison results show that the proposed method owns satisfactory recognition performance even under conditions of the high pulse missing rates and low SNRs,and the proposed method outperforms the existing four methods.