Turbofan Engine Remaining Useful Life Prediction with Reliable Prediction Intervals via LSTM-Based Quantile...

Area of Science:

• Aerospace Engineering
• Machine Learning
• Reliability Engineering

Background:

• Accurate prediction of remaining useful life (RUL) is crucial for turbofan engine maintenance.
• Point estimates and traditional prediction intervals often fail to adequately capture uncertainty and maintain stable coverage.
• Existing methods struggle with reliable RUL prediction under varying operating conditions.

Purpose of the Study:

• To develop a robust framework for turbofan engine RUL prediction that provides reliable uncertainty quantification.
• To enhance the stability and coverage of prediction intervals for RUL estimation.
• To support effective decision-making in predictive maintenance through interpretable RUL predictions.

Main Methods:

• Proposed a Long Short-Term Memory-based Quantile Regression (LSTM-QR) framework for RUL prediction.
• Integrated a weighted pinball loss and an overestimation penalty during training for robust quantile estimation.
• Applied Conformalized Quantile Regression (CQR) for post hoc calibration of prediction intervals.

Main Results:

• The LSTM-QR framework demonstrated stable point-prediction performance.
• Post hoc calibration using CQR significantly improved prediction interval reliability, achieving nominal coverage.
• Substantial gains in prediction interval probability of correct coverage (PICP) were observed under both same-condition and cross-condition transfer scenarios.
• Root Mean Square Error (RMSE) was maintained at competitive levels post-calibration.

Conclusions:

• The proposed LSTM-QR with CQR offers a reliable and interpretable approach to RUL prediction with quantified uncertainty.
• The framework effectively addresses the limitations of point estimates and unstable prediction intervals in turbofan engines.
• This method provides valuable support for predictive maintenance strategies, especially under challenging operating conditions.