Multi-task Open-set Radar Jamming Recognition Based on Convolutional Prototype Network

In electronic warfare, radar inevitably receives unknown jamming signals released by jammers. Due to incomplete class information, closed-set jamming recognition paradigms fail to infer and learn class knowledge, resulting in continuous performance degradation in dynamic scenarios. In this article, a complete dynamic open-set jamming recognition (OSJR) framework is proposed to realize multiple recognition tasks in dynamic open environments. The framework contains two modules: the OSJR module first identifies known jamming and infers unknown jamming, and then the jamming class incremental learning (JCIL) module learns class knowledge of the detected unknown jamming. Utilizing the advantages of complex-valued networks and prototype learning in separable representation learning and data reduction, a generalized complex-valued convolutional prototype network for OSJR (GCVCPNet-OSJR) and an extended CVCPNet for JCIL (ECVCPNet-JCIL) are developed to implement module functions. GCVCPNet-OSJR combines the prototype distance and reconstruction error to provide a more robust metric, and adopts extreme value theory with threshold calibration to adaptively optimize discrimination thresholds for jamming inference. ECVCPNet-JCIL combines regularization and few-shot replay strategies to improve the retention of old class knowledge while continually learning unknown class knowledge. Experiments show that the proposed method has superior performance in OSJR and JCIL, especially on small scale samples.