Spatio-temporal physics-informed digital twins for early fault prognostics of wind turbine generators and gearboxes.

Utility-scale wind energy conversion systems are subjected to highly stochastic aerodynamic loading profiles, making drivetrain components-particularly high-speed gearbox shaft bearings and generator stator windings-susceptible to premature failure. Traditional predictive maintenance frameworks rely either on pure data-driven deep learning models, which suffer from high false alarm rates during transient operational profiles and frequently violate thermodynamic boundaries, or first-principles physical models, which fail to capture dynamic boundary variations due to site-specific environmental factors. This paper introduces a hybrid Spatio-Temporal Physics-Informed Digital Twin (PIDT) framework that integrates physical thermodynamic principles and material fatigue crack propagation laws directly within deep learning architectures. For SCADA thermal anomaly detection, the proposed approach combines a steady-state thermodynamic baseline model with a causal, dilated Physics-Informed Temporal Convolutional Network with Self-Attention (PI-TCN-Attn). Anomaly thresholds are dynamically derived using Peaks-Over-Threshold Extreme Value Theory fitting a Generalized Pareto Distribution to tail excesses. For vibration-based structural diagnostics, amplitude envelopes are extracted using Hilbert demodulation to isolate mesh defect energies, which are subsequently reconstructed via a Multi-Scale Convolutional LSTM Autoencoder. Remaining Useful Life is forecasted using a Physics-Informed Gated Recurrent Unit that incorporates monotonicity constraints and the Paris-Erdogan crack propagation law directly within the loss function, enabling the online discovery of material fatigue parameters. Experimental validation on real SCADA datasets and vibration condition monitoring data demonstrates that the hybrid twin provides early warning lead times for critical drivetrain failures while maintaining low false alarm rates, statistically outperforming conventional unsupervised and supervised machine learning baselines. Furthermore, the learned fatigue parameters of the bearings converge to theoretical fatigue wear limits, validating the physics-guided parameter discovery mechanism.