Physics-Informed Hybrid Transformer with Uncertainty Quantification for Remaining Useful Life Prediction of Turbofan Engines

Turbofan engines operate under demanding conditions where unexpected failures can lead to serious safety andeconomic consequences. Accurately estimating their Remaining Useful Life (RUL) is therefore essential for enablingpredictive maintenance and improving system reliability. Physics based approaches provide interpretability andconsistency but they often fall short when modeling the nonlinear and dynamic nature of engine degradation. On theother hand, purely data-driven deep learning models, including LSTM, TCN, and Transformer architectures, achieveimpressive accuracy but tend to overlook physical constraints and uncertainty awareness. To address these limitations,this study introduces a Physics-Informed Hybrid Transformer (Pi-HT) that combines physical knowledge with datadriven learning. The proposed hybrid model integrates a CNN component to capture localized temporal features anda Transformer encoder to model long-range dependencies across sensor sequences. This study has incorporateda physics-inspired degradation rate module to enforce monotonic RUL trends. Additionally, uncertainty quantificationtechniques—including negative log, quantile regression, and conformal calibration—were used to improve the reliabilityand interpretability of the model’s predictions. Experiments on the NASA C-MAPSS dataset show that Pi-HT achieveshigh predictive accuracy, with an R2 of 0.902 and an RMSE of 12.40, better performing over 1D CNN, LSTM, TCN, andTransformer baselines. Beyond accuracy improvements, the framework provides physically consistent and statisticallycalibrated estimates, making Pi-HT a practical and trustworthy tool for real-world prognostic health managementapplications.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.