Physics-guided and data-driven Artificial Intelligence framework for aeroengine health monitoring

Reliable aeroengine health monitoring is essential for effective predictive maintenance and early fault localization. This study proposes a hybrid artificial intelligence framework that integrates unsupervised anomaly detection with supervised component-level fault classification. An Long Short-Term Memory (LSTM) autoencoder is first used to learn normal multivariate engine behaviour. Additionally, a physics-guided variant of the autoencoder is introduced to incorporate thermodynamically motivated residual features and reduce the influence of operating-condition variability. The detected anomaly information is subsequently processed by a multi-label Bidirectional Long Short-Term Memory (BiLSTM) classifier to identify simultaneous degradation across five major engine components. The framework is evaluated using degradation trajectories from nine turbofan engines in the N-CMAPSS dataset under realistic flight conditions. Across the evaluated datasets, the physics-guided autoencoder achieves Root Mean Square Error values (RMSE) of 0.0946 to 0.1100 and 𝑅 2 values of 0.9963 to 0.9976, showing better performance compared to the purely data-driven model, which presents RMSE values of 0.1268 to 0.1500 and 𝑅 2 values of 0.9918 to 0.9941. The BiLSTM classifier achieves accuracy values of 0.9448 to 0.9818 and recall values of 0.9321 to 0.9836 across the different datasets. These results demonstrate that combining physics-guided anomaly representation with bidirectional temporal learning improves reconstruction performance, supports reliable component-level fault localization, and provides a robust framework for aeroengine condition-based maintenance.