Research On Radar Signal Recognition Technology Based On Deep Learning

With the rapid development of radar technology and the continuous improvement of anti-jamming capability.Low intercept,ultra-wideband,fast signal agility,and other technologies are widely used.The signal environment that electronic reconnaissance is facing is becoming more and more complex,which brings great difficulties to radar signal interception and analysis.Since radar signal waveform is the main feature of radar,the recognition of radar signal waveform is an important research content of electronic reconnaissance.In this paper,the research on radar signal recognition technology based on deep learning will be carried out.Given the difficulties in collecting radar signals and the scarcity of samples in the real environment,this paper investigates two methods: generation model and transfer learning based on the analysis of the effect of classical deep learning algorithm on radar signal recognition under the condition of small samples.The main work of this paper is as follows:1.Establish the radar signal model,simulate and analyze six common radar signals from time and frequency domains.Then,two time-frequency analysis techniques,short-time Fourier analysis and wavelet transformation are used to analyze the time-frequency of six radar signals.By comparing experiments under different signal-to-noise ratios,it is found that short-time Fourier transform has better performance on radar signals under different signal-to-noise ratios.2.The radar signal recognition method based on the classical deep learning network is analyzed.CNN is analyzed in detail.Under the conditions of 20000 and 500 training sets,two CNNs,VGG16 and Res Net18,are used to identify the time-frequency map of the radar signal.The results show that Res Net18 network has good performance,but neither network can be used for radar signal recognition when training samples are scarce.3.The generative model is used to solve the problem of sample scarcity from the data level,and two kinds of generation models,VAE and GAN,are analyzed.Based on the time-frequency diagram,a large number of training samples are generated from a small number of real radar signals using VAE and GAN,and Res Net18 network is trained separately for comparison experiments.Contrast analysis shows that the generation of models improves the recognition accuracy of deep learning models.but because the data generated by GAN is unstable and the data generated by VAE is less variant,two types of networks are less effective in identifying radar signals.4.A method of using the transfer learning and optimized the Softmax classifier is presented.Radar signal recognition under the condition of scarce samples is also based on the time-frequency map of the radar signal.The results show that the recognition rate of the network can reach 89.1% under-10 d B condition,and the training time is usually shorter.The experimental results show that the method can effectively improve the network recognition rate under the condition of scarce samples.