Research On Radar Signal Intra-pulse Signatures Extraction Algorithm Based On Deep Learning

On the modern battlefield,conventional radars have gradually been replaced by new system radars.Traditional algorithms for radar signal sorting and identification based on five parameters of radar characteristics have failed.By analyzing the intentional and unintentional characteristics of the radar pulse signal,it is possible to realize radar signal sorting and identification under the condition of short-term observation data.Therefore,in recent years,how to extract signals captured in dense environments and extract intra-pulse features has become an important research content of radar signal sorting and identification.Compared with the traditional method,the intra-pulse feature analysis of radar signals based on deep neural networks can not only reduce the huge workload of artificially extracting features,but also can achieve the same or better results than artificially extracting features in a low signal-to-noise ratio environment.By drawing on the application of deep learning in image recognition and speech signal recognition,and according to the respective characteristics of intentional and unintentional features of radar signals,this paper uses convolutional neural networks and sparse self-encoding networks to perform intentional feature extraction and unintentional features extraction of radar signals respectively.In terms of radar signal intentional feature analysis,CWD time-frequency analysis is first used to perform time-frequency analysis on 8 commonly used radar signals,and a small-scale time-frequency image data set composed of eight signal time-frequency images is established.This paper uses transfer learning to avoid the risk of neural network overfitting on small data set.The final experimental results show that the overall recognition rate of the three networks for transfer learning can reach more than 90%.In order to further improve the recognition rate,this paper uses features fusion method to improve the radar signal recognition algorithm.The improved algorithm improves the overall recognition rate of 8 signals to more than 95%.According to the characteristics of the unintentional modulation characteristics of radar signals,this paper first performs bispectrum analysis on the measured radar signals,and preprocesses the bispectral data,and uses the stacked Sparse Auto Encoder(s SAE)to characterize the preprocessed bispectral data.In order to improve the classification performance,XGBoost is used as a classifier to recognize and classify the special features extracted by neural network.Experimental results on three actual radar emitter signals show that the overall recognition rate can reach more than 95%.This algorithm proposed in this paper can well extract the unintentional features of the radar signal.