Modulation Signal Recognition of Small Sample Radar Based on Attention Mechanism Prototype Network

In the modern battlefield environment, the signals of new radar systems are often short of enough samples to train deep learning models. At the same time, hostile parties may use various jamming and spoofing techniques, resulting in an extremely complex radar signal environment. In this case, how to quickly and accurately identify enemy radar signals is of vital significance for the acquisition of electronic countermeasures and air superiority. In this paper, a channel attention-mechanism module (Squeezed and Excitation Network) is introduced to improve the prototype network. By explicitly modeling the interdependence between convolutional feature channels, SE-Net can make the network pay more attention to important feature channels. The time-frequency images of thirteen kinds of LPI radar signals are extracted, and prototype network training based on attention mechanism is carried out to improve the prototype network by optimizing the mapping of embedded space and the design of classifier, which can further improve the accuracy and efficiency of radar signal recognition in the case of small samples. According to the simulation results, when the training set is only 8 classes and the test set is 13 classes, the recognition accuracy of SE-protonet network can reach 85% at OdB signal-to-noise ratio, which is 2 % higher than the pure protonet network and 20% higher than the traditional neural network, and has a good recognition accuracy at -12c <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\sim$</tex> SdB signal-to-noise ratio. At the same time, when the number of support set samples reaches 15-shot, the recognition rate of 8dB signal to the whole signal can reach 98%. This further shows that the attention-mechanism-based prototype network has a good recognition accuracy for radar modulated signals in the case of small samples, and performs well in multi-class signal recognition and dynamic signal environment adaptation.