Attention-Based Quantile Regression for RUL Uncertainty Prediction

Remaining Useful Life prediction is a core challenge in the field of Prognostics and Health Management. Traditional point prediction methods only provide a single estimate and cannot quantify prediction uncertainty, limiting their application in critical decision-making. This paper proposes a probabilistic prediction model integrating an LSTM-attention mechanism and quantile regression, aiming to achieve interval prediction and uncertainty quantification for RUL. The model employs a bidirectional LSTM network to capture temporal dependencies, focuses on key degradation features through a six-head self-attention mechanism, and outputs predictions for three quantiles (10%, 50%, 90%) simultaneously based on the quantile regression framework, constructing an 80% confidence interval with clear physical meaning. Experimental validation on a public dataset shows that the proposed model performs excellently in both point prediction accuracy and interval prediction quality: RMSE reaches 11.245, NASA Score is 166.414, while the interval coverage remains at 86.7%, and the interval width is 30.398. Ablation experiments further confirm the importance of each component, where removing the attention mechanism causes a 24.4% increase in RMSE, and removing the LSTM module worsens the NASA Score by 164.8%. The research results provide an effective solution for probabilistic remaining life prediction of complex equipment.