A Self-Supervised Contrastive Learning Method for Radar Signal Modulation Recognition
Shiya Li, Xiaolin Du · 2024
Effective implementation of supervised learning-based radar signal modulation recognition (RSMR) techniques is heavily dependent on the quantity and quality of labeled datasets. However, the high cost and difficulty involved in analyzing and labeling radar signal samples limit its development. To address this issue, a RSMR system that utilizes self-supervised contrastive learning (SSCL) methodology is proposed. In the classical contrastive learning framework MoCo V2, a custom data augmentation method is employed to capture time-frequency features of the radar signal. Furthermore, the feature extraction network ResNet50 is enhanced by separating spatial and channel filters, resulting in increased sensitivity to time-frequency features. To improve recognition accuracy, two loss functions, alignment and uniformity, are employed in place of the info noise contrastive estimation (InfoNCE) loss, and both loss functions are optimized directly. The experiments demonstrate the effectiveness of the proposed system.