Remaining useful life prediction with limited run-to-failure data: A Bayesian ensemble approach combining...

Area of Science:

• Engineering
• Data Science
• Machine Learning

Background:

• Accurate remaining useful life (RUL) prediction is vital for industrial predictive maintenance.
• Data-driven RUL methods often require extensive run-to-failure (R2F) data, which is frequently scarce.
• Limited R2F data leads to significant RUL prediction errors and challenges in quantifying uncertainty.

Purpose of the Study:

• To develop a robust Bayesian ensemble RUL prediction method adaptable to limited R2F data.
• To enhance RUL prediction accuracy and uncertainty quantification in data-scarce environments.
• To address the practical limitations of traditional RUL prediction techniques.

Main Methods:

• A Bayesian ensemble approach combining mode-dependent relevance vector machine (RVM) and trajectory similarity.
• Clustering of historical R2F trajectories into distinct degradation modes.
• Utilizing mode-dependent kernel functions and similar trajectories for improved RVM and similarity-based predictions.
• Fusing RVM and similarity predictions to quantify uncertainty with limited training data.

Main Results:

• The proposed method demonstrated improved RUL prediction accuracy in case studies involving bearings and batteries.
• Achieved over 20% reduction in mean absolute percentage error compared to three existing methods.
• Successfully quantified prediction uncertainty even with limited R2F trajectories (11 for bearings, 16 for batteries).

Conclusions:

• The Bayesian ensemble RUL prediction method effectively handles limited R2F data.
• The approach offers a practical solution for predictive maintenance in data-scarce industrial settings.
• Enhanced accuracy and uncertainty quantification make the method valuable for real-world applications.