Remaining Useful Life Prediction Method for Bearings Based on LSTM with Uncertainty Quantification.

Area of Science:

• Mechanical Engineering
• Reliability Engineering
• Data Science

Background:

• Bearing failures cause significant economic losses and safety risks.
• Real-time monitoring of internal bearing degradation is challenging.
• External uncertainties complicate accurate prediction of bearing degradation.

Purpose of the Study:

• To develop an effective remaining useful life (RUL) prediction method for rolling bearings.
• To address the challenges of real-time degradation monitoring and external uncertainties.
• To incorporate uncertainty quantification into bearing RUL prediction.

Main Methods:

• Proposed a novel bearing RUL prediction method utilizing long-short term memory (LSTM) networks.
• Introduced a fusion metric to reflect the latent degradation process based on runtime.
• Developed an improved dropout method using nonparametric kernel density for enhanced RUL estimation accuracy.

Main Results:

• The proposed method accurately predicts the point estimation of bearing RUL.
• The model successfully provides the probability distribution of the bearing RUL.
• Validation on the PHM2012 dataset confirmed the method's effectiveness.

Conclusions:

• The developed LSTM-based method with uncertainty quantification offers accurate bearing RUL prediction.
• The fusion metric and improved dropout enhance the model's ability to handle degradation and uncertainty.
• This approach contributes to reducing economic losses and improving safety by enabling proactive maintenance.