Physics Constrained Multi-Sensor Data Fusion for Steel Catenary Riser Shape Reconstruction

The steel catenary riser (SCR) is a critical component in floating production systems (FPS). It serves as the primary infrastructure for transporting hydrocarbons between subsea production system and topside processing units. Its geometric simplicity, structural robustness, and adaptability to large water depths have led to its widespread adoption in deepwater oil and gas fields around the world. The SCR is subjected to complex environmental forces such as wave loading, platform-induced motion, and ocean currents. These dynamic excitations can induce significant deformation, fatigue accumulation, and localized stress variations. To ensure safe and continuous operation, the oil and gas industry has implemented various monitoring strategies to assess riser integrity and motion response. However, realizing accurate, high-resolution monitoring of SCR deformations in real-time remains a major challenge. The proposed framework integrates multi-sensor data from Inertial Measurement Units (IMUs) with physics constrained machine learning. The denoising autoencoder (DAE) is first applied to refine raw sensor signals, mitigating noise and preserving critical motion characteristics. Riser orientation is then estimated with the Madgwick filter, followed by an Unscented Kalman Filter (UKF) for displacement reconstruction. The graph neural network (GNN) is employed to ensure structural consistency, incorporating physics-based constraints to model spatial dependencies and optimize riser shape reconstruction. The framework is trained and validated using real-world monitoring data from a deepwater SCR, incorporating platform motion, ocean environmental conditions, and multi-axis angular velocity and acceleration. The proposed framework accurately reconstructs SCR dynamic behavior and is validated with OrcaFlex simulations under identical environmental and operational conditions. Quantitative assessments across three riser areas (hang-off, midpoint, and touchdown) show consistently high performance. The root mean square error (RMSE) remains below 0.06 m in all spatial directions, with R2 values exceeding 0.86 and mean absolute errors (MAE) maintained under 0.045 m. These metrics confirm that the fusion of IMU-derived dynamic features with physics-informed modeling achieves precise spatial tracking across the full riser. Notably, the framework preserves both global shape characteristics and localized vibration trajectories, such as elliptical oscillations near the hang-off area, figure-eight lateral loops at midpoint, and pipe–soil interaction signatures at the touchdown zone. This validates the robustness of the model in capturing non-linear, multi-modal responses critical for real-world offshore monitoring and structural integrity assessment. This study introduces a novel physics constrained multi-sensor fusion framework for SCR shape reconstruction, integrating a GNN to enforce physical consistency within a deep learning paradigm. This approach surpasses traditional numerical simulations in accuracy and adaptability by bridging the gap between sensor-driven data processing and physics-informed modeling. The proposed framework offers a scalable and generalizable solution for offshore structural health monitoring, enabling real-time motion estimation and providing critical insights for the safe and efficient operation of deepwater risers in dynamic marine environments.