Physics-Informed Digital Twins for Real-Time Health Monitoring Performance Prediction and Remaining Useful Life Assessment of Gas Turbines Foundations and Structural Support Systems

Gas turbines and their foundations and structural support systems operate as dynamically coupled thermo-mechanical assets whose degradation is governed by interacting aerodynamic, thermal, rotational, vibration, fatigue, settlement, stiffness-loss, and load-transfer mechanisms. Conventional condition-monitoring and remaining useful life (RUL) models frequently treat the turbine and its supporting structure independently and rely predominantly on historical sensor patterns, limiting their ability to maintain physical consistency and generalize under changing operating regimes. This study develops a physics-informed digital twin framework for real-time health monitoring, performance prediction, anomaly diagnosis, and RUL assessment of the integrated gas turbine–foundation–structural support system. A novel Physics-Informed Graph-Transformer Digital Twin Prognostics Network (PIGT-DTNet) is proposed. PIGT-DTNet combines thermodynamic gas-path constraints, rotor-dynamic relationships, structural dynamic equilibrium equations, fatigue and stiffness-degradation laws, graph neural message passing, multi-head temporal attention, and online sensor assimilation within a unified digital-twin architecture. Multivariate streams consisting of exhaust gas temperature, compressor pressure ratio, shaft speed, fuel flow, power output, bearing temperature, vibration amplitude, modal frequency, strain, displacement, foundation settlement, and support acceleration are synchronized with continuously updated virtual states. The physics-informed loss function constrains learned predictions using conservation of mass and energy, gas-turbine performance relationships, and the structural equation Mu¨+Cu˙+Ku=F(t)M\ddot{u}+C\dot{u}+Ku=F(t), thereby reducing physically implausible predictions and improving extrapolation during degradation conditions. The proposed algorithm is comparatively evaluated against LSTM, BiLSTM, GRU, CNN-LSTM, conventional Transformer, XGBoost, standalone Graph Neural Network, and conventional Physics-Informed Neural Network models using RMSE, MAE, R2R^2, RUL score, prognostic horizon, fault-classification accuracy, inference latency, and uncertainty calibration. Comparative line graphs, degradation trajectories, predicted-versus-actual plots, confusion matrices, error distributions, 3D response surfaces, sensor-to-health-state heat maps, and RUL confidence-bound plots are employed to establish performance differences across operating loads and degradation stages. The proposed framework is designed to achieve lower prediction error, earlier degradation detection, more stable RUL estimates, improved physical consistency, and stronger robustness to noisy and partially missing sensor measurements than purely data-driven and conventional physics-informed baselines. By simultaneously representing turbine performance deterioration and foundation/support-system degradation, the study establishes an integrated prognostic digital twin capable of supporting condition-based maintenance, structural integrity management, failure-risk reduction, and life-extension decisions for industrial gas turbine installations.