Adaptive relevant vector machine based RUL prediction under uncertain conditions.

Area of Science:

• Engineering
• Machine Learning
• Reliability Engineering

Background:

• Engineering systems face performance degradation due to time-dependent uncertainties and operational changes.
• Accurate remaining useful life (RUL) prediction is crucial for system maintenance and reliability.

Purpose of the Study:

• To develop an improved relevance vector machine (RVM) approach for RUL prediction that explicitly addresses system uncertainties.
• To enhance the accuracy and reliability of RUL predictions in degrading engineering systems.

Main Methods:

• An offline multi-step RVM regression model was established using historical data, quantifying uncertainties via Gaussian distribution variances.
• An adaptive RVM model was trained, updating time-varying uncertainties using the expectation-maximization (EM) algorithm.
• Online RUL forecasting was performed using the first hitting time (FHT) method.

Main Results:

• The proposed RVM approach effectively models degradation processes from fault to failure.
• Demonstrated effectiveness through case studies on a high-speed train's traction system.
• The method successfully incorporates and quantifies common uncertainties in RUL prediction.

Conclusions:

• The improved RVM method provides accurate RUL predictions by considering system uncertainties.
• This approach enhances the reliability and predictive maintenance capabilities for engineering systems.
• The study validates the proposed method's effectiveness in real-world applications.