Time-Validated Latent Fault Detection in Gas Turbines Using Physics-Based Features and Interpretable Decision Trees

Abstract

This study addresses fault detection in gas turbines under sensor-limited conditions, where diagnostically informative measurements such as emissions and internal variables are unavailable. A leakage-aware framework is proposed using only five readily available thermodynamic sensors (AT, AP, TAT, AFDP, TEY). To compensate for missing measurements, physics-based features grounded in Brayton cycle principles are constructed. A proxy fault label is generated offline using high-fidelity variables (TIT, GTEP, CO, NOx), which are assumed to be unavailable during real-time deployment. A Decision Tree classifier is selected to ensure interpretability in safety-critical environments. To reflect realistic industrial conditions, a strict time-based validation strategy is adopted. The results show that the proposed model achieves an F1-score of 0.703, a Recall of 0.881, and an AUC of 0.951. Furthermore, random split validation is found to overestimate performance by approximately 5.7% in F1-score, highlighting the risk of optimistic bias in conventional evaluation practices. The proposed framework provides a practical and interpretable solution for fault detection under constrained sensing conditions, with direct applicability to legacy turbine systems.