A Physics-informed Consistency Learning Framework for Robust Condition Monitoring of AUVs in Extreme Environments

Ensuring the survivability of Autonomous Underwater Vehicles (AUVs) in unstructured, disturbance-prone environments leads to a significant challenge due to sensor degradation and physical uncertainties. Conventional deep learning models often suffers from rapid decline in performance and limited interpretability under such conditions. To address the sim-to-real gap, this study proposes the Physics-Informed Residual Transformer (PIR-Transformer), which integrates two key components: 1) a Physics-Based Residual Mechanism that enforces consistency among sensor and actuator signals, and 2) an Environment-Adaptive Consistency Learning strategy that introduces simulated "Microshift" noise during training. By minimizing divergence between clean and perturbed predictions, the model learns robustness to environmental variations. Evaluation on the Haizhe dataset shows that the proposed method achieves a macro F1-score of 0.7316 under extreme-noise (N3), outperforming the best baseline by 0.1759. Unlike conventional models, which exhibit severe degradation, the PIR-Transformer maintains stable performance, demonstrating that combining physics-based constraints with adaptive learning effectively enhances robustness in real-world conditions.