A two-stage machine learning framework for fault detection and severity estimation in industrial gas turbines
Loading...
Date
2026-06-22
Authors
Καλδής, Μάριος Βασίλειος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Heavy-duty gas turbines operating in continuous power generation accumulate degradation across compressor and turbine components, making systematic fault diagnosis a necessary capability for reliable operation. This thesis develops and evaluates a two-stage machine learning diagnostic system for fault detection and severity estimation, targeting the single-shaft GE 7HA.03 gas turbine as a representative industrial platform. The thermodynamic cycle of the engine was modeled in the GasTurb trial version, under three load conditions, and a damage propagation model adapted from established NASA methodology was used to simulate three fault types: compressor fouling, turbine erosion, and their combination as a double fault. Monte Carlo sampling of ambient conditions produced a dataset spanning eight degradation severity levels across three load conditions, while multiplicative Gaussian noise was injected for data augmentation purposes. The datasets produced contained 34,560 and 2,880 operating points for training and testing, respectively.
The first diagnostic stage is a selective classifier, calibrated via isotonic regression, that assigns fault class labels only when the estimated class probability exceeds a defined confidence threshold. After comparing ensemble tree-based classifiers, XGBoost was selected for its accuracy and abstention behavior at the target confidence threshold. Samples falling below this threshold are routed to Monitoring Mode for further human review. At a threshold of 0.85, the XGBoost achieved 98.6% classification accuracy on committed predictions with a coverage of 94.7%.
The second stage is a Multi-Task Multilayered Perceptron regressor with two dedicated output heads. This architecture reflects the physical separation between the two fault domains and produced an aggregate R² of 0.985 and a mean absolute error of 0.156 percentage points across all four degradation parameters.
Explainability analysis via SHAP and Permutation Importance confirmed that both stages rely on physically interpretable sensor signals, with attribution patterns consistent with the fault signatures identified during exploratory data analysis. Finally, robustness evaluation included testing at interpolated degradation levels, elevated sensor noise up to 100% above the training distribution, and incipient fault levels that fall within the sensor noise floor, collectively defining the conditions under which the diagnostic system operates reliably and the boundaries beyond which its performance degrades.
Description
Keywords
Industrial gas turbines, Health monitoring, Selective classification, Multi-task learning, Degradation severity estimation