MM-CL: Enhancing Modulation Classification Through Multi-Modal Contrastive Learning With Phase Density

Automatic modulation classification (AMC) plays a pivotal role in cognitive communication systems. Recent advances in contrastive learning have shown promising feature extraction capabilities for AMC tasks using large-scale unlabeled signal samples. However, the effectiveness of these approaches remains constrained by two fundamental challenges: 1) the mono-modal representation constraints that limit feature diversity, and 2) the suboptimal handling of negative samples leading to feature dispersion.To address these issues, a multi-modal contrastive learning (MM-CL) framework for self-supervised signal modulation classification is proposed in this paper. In detail, we propose a novel signal modality, the phase density (PD), which enriches the phase distribution representation of the original signal. Additionally, we employ a top-k instance contrastive loss to filter out false negative samples, which results in feature dispersion. Experimental results on the four datasets show that MM-CL outperforms existing methods, achieving an average improvement of 4.7% in 0-18 dB signal-to-noise ratio and 1-shot scenarios. Above all, MM-CL provides an effective framework for multi-modal AMC, suggesting potential avenues for optimizing model efficiency and exploring diverse feature representations.