Research On Radar Emitter Recognition Based On Deep Learning

Radar emitter recognition is a vital task for radar countermeasures,which has an important technical support role in determining radar usage,operating mode,threat level,behavioral intent and loading platforms.Based on the deep learning methods,this project focuses on two research contents of radar intra-pulse modulation type recognition and radar individual recognition,and the main work done is as follows.For the problem of radar intra-pulse modulation type recognition,two investigations are carried out in this thesis from different perspectives.From the perspective of data preprocessing,the radar emitter intra-pulse modulation type recognition methods based on encoding of signal images and transfer learning is proposed.These methods first transform the radar signal into a two-dimensional image,and then adopt transfer learning techniques to achieve accurate recognition of the radar emitter intra-pulse modulation type.Compared with traditional time-frequency transform algorithms,the encoding of signal images methods consume less time and computational resources and can be combined with pre-trained models to achieve higher recognition accuracy with lower training expenses.From the perspective of sampling signal recognition,an intra-pulse modulation type recognition method based adaptive deep residual shrinkage network is proposed.By comparing and analyzing the recognition performance of the soft threshold function with other traditional activation functions,the improved adaptive leaky soft threshold function is proposed as a new nonlinear activation function to construct an adaptive deep residual shrinkage network model more suitable for sampling signal recognition,which can filter noise features while retaining signal features and improving the feature learning capability.The experimental results show that compared with the traditional deep learning networks,the adaptive deep residual shrinkage network improves the recognition accuracy and is more suitable for sampling signal recognition.For the problem of individual recognition of radar emitter,this thesis proposes a method of one-step recognition combining residual neural network and Transformer in parallel,which is named Resformer.The signal does not need to be pre-processed,Resformer directly takes the original one-dimensional signal as the input,and uses its powerful feature extraction ability to realize the one-step recognition of radar emitter individuals by combining the local correlation features extracted by the residual neural network and the sequence important discriminative features extracted by Transformer.