Research On Key Technologies Of Radar Emitter Recognition Based On Machine Learning

Radar emitter identification is a key link in electronic countermeasure reconnaissance system,which provides an important basis for subsequent implementation of electromagnetic interference and defense.Due to increasingly complex and changeable battlefield electromagnetic environment and continuous development of radar signal system,non-cooperative radar signal modulation recognition has become a significant challenging research topic.Traditional modulation recognition methods are extremely dependent on artificial experience and have poor adaptability to complex electromagnetic environment,which is difficult to meet the needs of modern battlefield environment.In recent years,machine learning has made great achievements in the field of signal processing with its strong self-learning and generalization ability.In order to solve the problem of poor adaptability of traditional recognition methods,this thesis focuses on the application of machine learning in the field of radar emitter modulation recognition.The main contents are as follows:1.In the field of non-cooperative signals,there are few signal samples,which makes it difficult to train deep networks.To solve this problem,a radar emitter modulation recognition method based on transfer learning and residual network is proposed.The improved fusion time-frequency images are used to train the pre-trained residual network to realize the intra-pulse modulation recognition of radar emitter signal.In this thesis,the trained models are tested in Gaussian white noise and Rician fading channel respectively,and it is found that the residual network trained by white noise signal can better identify the multipath fading signal at low signal-to-noise（SNR）ratio.2.In order to solve the problem that deep network model requires high computing resources of deployed equipment,an intra-pulse modulation recognition method based on lightweight neural network was proposed.Due to the high computational complexity of deep network,it is not conducive to real-time signal recognition.An improved lightweight residual network is constructed by using asymmetric convolution kernel and grouping convolution.The multi-scale information of radar time-frequency image is extracted through the core residual module to improve the recognition effect of radar signal under low SNR.The simulation results show that compared with other deep models,the lightweight neural network proposed in this thesis reduces the computational power requirements of deployed equipment while ensuring the recognition performance,and the recognition accuracy of 13 radar signals reaches 90 % when the SNR is-8d B.3.Aiming at the problems of redundant information and high dimension of manually extracted time-frequency images,a time-series signal recognition method of radar emitter based on machine learning is proposed.The attention mechanism is introduced to select and optimize the features extracted by the residual module,and multiple pooling shortcut connections are used to fuse the signal features in different spaces,so as to improve network performance.Through the comparison of simulation experiments,the following conclusions are obtained: 1)the recognition performance of the proposed method achieves superior performance than traditional method and three deep learning methods; 2)Under the condition of very low SNR,the recognition effect of using sequential signal method is better than time-frequency images ; 3)For phase modulation signals,the recognition performance of time-frequency images as network input is better.For normal pulse signal and frequency modulation signals,the overall effect of time-frequency images as network input is better.