Unknown Signal Class Discovery Based on Feature Constraints and Knowledge Distillation

This letter proposes a novel approach for automatic modulation recognition and novel class discovery, aiming to address the challenge of modulation classification in open-world scenarios where unknown samples exist, thereby enabling continuous signal reconnaissance. To optimize the network's feature space, we introduce two loss functions: one enforces inter-class constraints for different types of samples, while the other imposes intra-class constraints between samples and their regularized representations. To alleviate the forgetting problem in the discovery stage and to better exploit the knowledge of the supervised model, we employ knowledge distillation from the perspective of semantic correlations between labeled and novel samples to guide the clustering process. An adaptive weighting mechanism is used to adjust the contribution of the pretrained model, thereby mitigating its interference with the clustering of irrelevant samples. Comparative experiments conducted on public datasets demonstrate the effectiveness of the proposed method against state-of-the-art approaches.