Semi-supervised method for ship-radiated noise recognition system of wave gliders via contrastive pseudo learning

Underwater acoustic target recognition plays a critical role in enhancing autonomous navigation and environmental sensing capabilities for unmanned mobile platforms. However, most existing methods cannot effectively cope with acoustic signal identification at low signal-to-noise ratios. In this study, a contrastive pseudo learning framework is proposed to design the data-driven ship-radiated noise recognition (SRNR) system, which is suitable for edge computing devices such as wave gliders. In particular, a time-frequency soft re-weighted module is applied to calculate soft thresholds in time and frequency domains to achieve feature fusion and adaptive suppression of interference. Additionally, fullband-subband contrastive learning is introduced to guide feature alignment and integrate global and local information for acoustic representation enhancement. Other than that, contrastive pseudo labeling is employed to use a large scale of unlabeled data with limited amounts of labeled data for semi-supervised learning. Extensive experiments along with sea trials show that the proposed method achieves the state-of-the-art performance on SRNR in noisy, real-world conditions.