Research on Enhancing Open-Set Generalizability for Active Jamming Recognition on Measured Dataset

With the rapid development of radar interference technology, the diversity of active jamming and the variety of interference strategies increase continuously. As a consequence, active jamming recognition (AJR) in the real world is gradually becoming an open-set classification problem. Thus, the generalizability of AJR methods based on deep learning is challenged severely. In this paper, a dataset is constructed with real-world outfield signals to reflect the generalization capability of AJR methods. A convolutional neural network (CNN) based time-frequency domain AJR method is established as the research baseline, whose recognition and generalizability are evaluated on the dataset. To enhance the recognition and rejection abilities on test datasets containing unforeseen jamming patterns, a data augmentation strategy based on chirp characteristics and a metric learning loss optimizing feature clustering is designed. Experiments demonstrate that the proposed methods possess superior recognition and rejection accuracy on the test dataset to those of the baseline method.