Uncertainty aware predictive maintenance using a hybrid Transformer with Monte Carlo Dropout and conformal prediction

Predictive maintenance (PdM) relies on accurate estimation of the remaining useful life (RUL) to support efficient industrial maintenance. However, most RUL models overlook uncertainty quantification (UQ), which is essential for safety–critical decision-making. This study presents a hybrid uncertainty-aware framework that combines a Transformer backbone with Monte Carlo Dropout (MC Dropout) and Conformal Prediction (CP). The Transformer architecture effectively learns long-range temporal dependencies in sensor data, while MC Dropout approximates epistemic uncertainty arising from model limitations. CP complements this by producing prediction intervals that capture aleatoric variability caused by noise and operating conditions. The framework is validated using NASA’s C-MAPSS FD001 and FD003 datasets. It achieves strong performance on FD001, with MAE 8.11, RMSE 11.71, and a predictive score of 193.6, and on FD003, with MAE 7.21, RMSE 10.50, and R2 0.926. By jointly addressing both uncertainty types, the method yields well-calibrated confidence intervals, enhancing reliability and interpretability in PdM applications.