Research And Application Of Radar Signal Recognition With Deep Learning Methods

Since radar was invented by the British in the 1930s,the related technologies and applications have changed rapidly.Radar plays an irreplaceable role in national defense,military and aerospace.While every country try to develop more powerful radar technology,they are also developing corresponding countermeasure technology.As an important function of radar countermeasure system,radar emitter signal recognition plays an incomparable role in electronic reconnaissance.The more information we get from radar emitter signal,the more support we can provide for our military operations.However,with the development of science and technology and radar system,most radars nowadays have strong anti-reconnaissance and anti-jamming characteristics,which makes it difficult for us to recognize radar emitters.How to improve the accuracy of radar emitter recognition is of great significance under the situation that traditional methods are gradually diff-icult to meet the task of radar emitter recognition.Firstly,the background and significance of radar emitter signal recognition are summarized,and then the related research methods inland and abroad are summarized and analyzed in this thesis.On the basis of understanding the relevant theory of radar emitter recognition,we apply machine learning knowledge such as deep learning and clustering to the classification and recognition task of radar emitter,and achieve good results.The main work of this thesis includes:(1)For the radar type classification problem,we propose a method based on one-dimensional convolutional neural network,and design the convolutional network structure which is most suitable for the needs of this project through experiments,so that we can classify radar type without knowing the type of signal modulation.(2)Aming at the problem that labeling radar signals has high labeling cost and complexity,we use semi-supervised learning method based on Generative Adversarial Networks.While learning labeled samples,we train the network using unlabeled samples by generating models,which greatly reduces the labeled data needed in training network process.(3)For the unknown signals that have not been seen in the training set,we use the deep metric learning to learn the metric relationship of the known categories,and use Generative Adversarial Networks to enhance the effect of feature learning,and then use this metric relationship to distinguish the known signal and unknown signal.Finally,we cluster the unknow samples with the deep metric networks and k-means++.