Data-driven underwater virtual sensing for deep-water jacket platforms using dual-stream Temporal-Frequency Attention and Distribution shift Adaptation framework

Deep-water jacket platforms (DJPs) operate under severe marine loads, making sustained observation of their underwater structural responses essential for integrity management. However, relying solely on physical sensors for long-term coverage is often constrained by their high installation costs, limited accessibility, and finite service life. Underwater virtual sensing offers a practical alternative by reconstructing submerged responses from above-water measurements. Its reliability is nevertheless challenged by the nonlinear dynamics of DJPs, which induce complex spatiotemporal heterogeneity and lead to distribution shifts across different sensing locations and temporal phases. To overcome these challenges, a novel Virtual Sensor framework with dual-stream Temporal-Frequency Attention and Distribution shift Adaptation (TFADA-VS) is proposed. The dual-stream attention mechanism is designed to capture complementary time-frequency dependencies for virtual sensing, while an adaptive distribution shift learning module is introduced to enhance awareness of distribution shifts. A co-optimization loss function is further designed to reduce the misalignment between time-domain and frequency-domain feature streams and the residual distribution discrepancies. Experimental evaluations on real-world monitoring data from a DJP in the South China Sea demonstrate the superior performance of TFADA-VS. Compared with the best-performing baseline, TFADA-VS achieved a 39.73% improvement in MAE, a 38.75% improvement in RMSE, and a 43.99% improvement in SMAPE. These results suggest that TFADA-VS provides a promising solution for cost-effective and intelligent long-term monitoring of underwater structures.