Uncertainty-quantified unsupervised transfer learning for ultrasonic-based corrosion detection in underwater steel pipes

Automation-ready subsea and coastal asset inspection requires reliable, low-touch structural health monitoring, critical for digitalized underwater pipeline management. This paper introduces a deep-ensemble unsupervised transfer-learning (DE-UTL) framework to automatically quantify corrosion-induced thickness loss from ultrasonic signals while quantifying aleatoric and epistemic uncertainties. DE-UTL employs an enhanced transformer architecture with adaptive weighting and multi-scale feature extraction, integrated with a Wasserstein-distance domain adapter supported by consistency regularization and reconstruction, ensuring environmental transferability and mitigating retraining needs. Ten UTL models are integrated via a deep ensemble to evaluate uncertainty. Underwater corrosion and ultrasonic experiments on two pipe types are conducted, revealing that pipe thickness, weight, and ultrasonic signal amplitude and phase exhibit trend-like variation with corrosion severity. Experimental validation confirms DE-UTL's effective corrosion thickness quantification in both geometry-only and geometry-environmental (air-to-water) transfer scenarios. Integrating corrosion experimentation, unsupervised transfer learning, and uncertainty quantification, DE-UTL shows strong potential for automated, scalable, real-time underwater asset inspection.