A Metric-Driven Evaluation Framework for Remaining Useful Life Prognosis with Quantified Uncertainty.

Area of Science:

• Mechanical Engineering
• Data Science
• Artificial Intelligence

Background:

• Accurate Remaining Useful Life (RUL) prediction is crucial for effective maintenance and preventing equipment failures in rotating machinery.
• Existing RUL prognosis methods often fail to adequately address complex nonlinear degradation patterns and provide reliable uncertainty quantification.

Purpose of the Study:

• To introduce a novel metric-driven evaluation framework for RUL prognosis in rotating machinery.
• To enhance RUL prediction accuracy and provide robust uncertainty quantification for industrial applications.

Main Methods:

• Integration of a probabilistic Deep State Space Model (DSSM) with variational inference to model nonlinear degradation and aleatoric uncertainty.
• Utilization of the Slime Mold Algorithm (SMA) for efficient hyperparameter optimization.
• Implementation of an online adaptation mechanism using a heuristic reinforcement learning agent for real-time model updates and concept drift management.

Main Results:

• Demonstrated superior RUL prediction accuracy on the IMS bearing dataset.
• Achieved the lowest Root Mean Square Error (RMSE) of 8.1829 cycles.
• Obtained a Prediction Interval Coverage Probability (PICP) of 0.59416, indicating reliable uncertainty estimates.

Conclusions:

• The proposed framework offers a significant advancement in RUL prognosis for rotating machinery.
• Its dual capability of accurate prediction and uncertainty quantification makes it highly suitable for real-world predictive maintenance.
• The framework enhances operational safety and equipment reliability through improved maintenance strategies.