Engine health monitoring - diagnostics using data driven / AI methods
Loading...
Date
2026-07-07
Authors
Αναγνώστου, Κωνσταντίνος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Aircraft gas turbine engines are highly complex thermodynamic systems that experience gradual performance degradation throughout their operational life due to mechanisms such as compressor fouling and turbine erosion. Early detection and accurate diagnosis of these degradations are essential for ensuring flight safety, reducing maintenance costs, and supporting condition-based maintenance. However, strong thermodynamic coupling between engine components, together with continuously varying flight conditions, makes reliable gas-path diagnostics a challenging task for conventional machine-learning approaches.
This thesis presents a hierarchical artificial intelligence framework for aircraft engine gas-path diagnostics, combining physics-informed feature engineering with multi-stage machine learning and transfer learning. A high-fidelity turbofan engine model was employed to generate large-scale datasets representing healthy and degraded engine operation under multiple flight conditions through Monte Carlo simulation. A data-driven Healthy Baseline Interpolator was developed to estimate nominal engine behavior across the operating envelope, enabling the computation of residual-based features that effectively separate degradation signatures from operating-condition variability.
The proposed framework decomposes the diagnostic process into four sequential stages: fault detection, fault multiplicity classification, component isolation, and fault severity estimation. Dedicated machine-learning models, including LightGBM, XGBoost, multilayer perceptrons, and convolutional neural networks, were evaluated and optimized for each diagnostic task. The methodology was assessed through three experimental campaigns investigating same-distribution performance, cross-operating-point generalization, and transfer learning for target-domain adaptation.
The results demonstrate that the proposed hierarchical framework achieves near-perfect diagnostic performance under identical operating conditions while significantly outperforming conventional approaches in terms of diagnostic robustness and interpretability. Cross-operating-point experiments reveal a substantial deterioration in performance caused by domain shift, highlighting the limitations of models trained under fixed operating conditions. Finally, supervised transfer learning successfully restores diagnostic performance using only a limited amount of labeled data from the target operating condition, demonstrating a practical and data-efficient solution for real-world deployment.
Overall, this work provides a comprehensive methodology for AI-based aircraft engine diagnostics that combines hierarchical decision-making, physics-informed feature engineering, and transfer learning to enable reliable gas-path diagnostics across realistic operating conditions.
Description
Keywords
Aircraft engine diagnostics, Engine health monitoring, Machine learning, Physics-informed feature engineering, Transfer learning