A Synthetic Reliability-Aware PINN Benchmark for Offshore Wind Turbine Support-Structure Monitoring with Bayesian Inverse Identification

Structural health monitoring (SHM) for offshore wind turbine (OWT) support structures is often hindered by the high computational cost of finite element analysis and the scarcity of sensor data. This paper addresses the need for a fast, physics-consistent surrogate model capable of real-time state estimation and reliability assessment, specifically targeting the challenges of inverse parameter identification and computational latency in digital twin applications.

The authors introduce DigiTurbine, a synthetic benchmark that integrates three core components into a single pipeline. First, a forward PINN uses the Euler-Bernoulli beam equation to approximate structural displacement from sparse measurements. Second, to address the 'gradient direction problem'—where standard PINNs fail to identify physical parameters like stiffness when initial weights are far from the truth—the authors implement a Bayesian-prior-informed inverse identification strategy. Finally, the pipeline incorporates the First-Order Reliability Method (FORM) to provide near-real-time structural reliability indices, bypassing the need for computationally expensive Monte Carlo simulations.

The study demonstrates that the DigiTurbine framework achieves an end-to-end inference latency of less than 7 ms, significantly faster than the 10 ms real-time target. The forward PINN model achieves high accuracy with minimal sensor input (50–100 points). Crucially, the authors identify that standard inverse PINNs fail on fourth-order PDE problems due to conflicting gradient directions; however, applying a weak log-normal Bayesian prior centered on the as-built specifications allows for robust identification of material properties (Young's modulus and soil stiffness) even under 10% measurement noise. The integrated FORM layer provides reliability estimates in approximately 1 ms, enabling rapid structural health screening.