Modulation Recognition of Digital Signals Based on Contrastive Self-Supervised Learning

Existing deep learning-based modulation recognition algorithms often require ample labeled samples. Considering the significant time and human resources needed for labeling signal samples, this paper proposes a recognition model based on contrastive self-supervised learning (SSL). The network is built upon the SSL framework SimCLR and introduces a novel feature extraction module comprising improved multi-head self-attention mechanism and memory storage module. To enhance the model's self-supervised performance, a new loss function is designed. Experimental results demonstrate tha t the proposed self-supervised learning model, SMRA-N et, achieves effective recognition of nine modulation signals even in the absence of labeled samples, outperforming other models in terms of performance.