Online Piecewise Convex-Optimization Interpretable Weight Learning for Machine Life Cycle Performance Assessment.

Area of Science:

• Engineering
• Machine Learning
• Condition Monitoring

Background:

• Machine life cycle performance assessment is crucial for predicting failures.
• Existing methods often rely on unexplainable models and historical fault data.
• A need exists for interpretable, data-driven approaches for real-time machine health assessment.

Purpose of the Study:

• To propose an online, interpretable framework for machine life cycle performance assessment.
• To develop a piecewise health index for detecting incipient faults and trending degradation.
• To overcome limitations of traditional methods requiring historical abnormal data.

Main Methods:

• An online piecewise convex-optimization framework with interpretable weight learning.
• A first submodel using a separation criterion to detect incipient fault initiation.
• A second submodel with monotonicity and fitness properties for fault trending, using online updated weights correlated with fault frequencies.

Main Results:

• The framework generates a piecewise health index using only online monitoring data.
• Interpretable model weights are updated online, correlating with fault characteristic frequencies.
• Effectiveness demonstrated through two run-to-failure case studies.

Conclusions:

• The proposed framework enables practical machine life cycle performance assessment.
• It allows for timely fault detection, identification, and trending using interpretable online learning.
• The method offers a superior alternative to traditional approaches by eliminating the need for historical fault data.