SCAW: Exploring Semantic Structures of Radio Signals via Soft Top-k Contrastive and Adaptive Consistency With Wavelet Refinement

Abstract

Self-supervised learning (SSL) methods have recently been applied to cognitive radio technology (CRT), yielding promising results by utilizing unlabeled data. Nevertheless, most existing SSL approaches primarily focus on a single aspect, either adopting contrastive paradigms that emphasize discriminative feature learning or generative paradigms that excel at extracting the semantic structure of fine-grained information. In this paper, we introduce a novel SSL method, Soft Top-k Contrastive and Adaptive Consistency with Wavelet-Based Refinement (SCAW), which effectively integrates the advantages of both discriminative feature learning and semantic, fine-grained structural information extraction, thereby enhancing robustness and generalization in CRT scenarios. First, a wavelet-based feature refinement module is incorporated into the encoder to enable full-frequency and fine-grained semantic feature extraction. Second, a soft top-k contrastive loss is introduced to exploit inherent inter-sample relationships, thereby preserving discriminative capability and enhancing the semantic structural integrity of the learned representations. Moreover, an adaptive consistency loss is formulated to mitigate noise-induced perturbations while maintaining the underlying structure of the representations. The effectiveness of SCAW is evaluated on four public datasets across multiple downstream tasks. Experimental results show that SCAW consistently outperforms existing state-of-the-art methods.