Learning to Balance the Performance and Deterioration of Aging Systems Through Derating.

Area of Science:

• Engineering
• Operations Research
• Machine Learning

Background:

• System aging necessitates strategies like derating to extend operational lifetime.
• Derating involves a trade-off between reduced performance and slowed deterioration.
• The performance-deterioration (pd) relationship is crucial but often unknown a priori.

Purpose of the Study:

• To optimize workload in derating scenarios where the pd-relationship is learned adaptively.
• To develop a framework for real-time decision-making considering performance, deterioration, and learning.
• To characterize the optimal derating policy under uncertainty.

Main Methods:

• Formulation as a partially observable Markov decision process (POMDP).
• Adaptive learning through sequential experimentation (active learning).
• Development of an efficient algorithm using the fast Gauss transform.

Main Results:

• The workload influences performance, deterioration, and learning speed.
• The optimal workload is consistently lower than the myopic load.
• The proposed algorithm efficiently computes optimal policies.

Conclusions:

• Optimal derating requires balancing immediate performance with long-term learning and system health.
• Active learning enables effective workload optimization even with unknown pd-relationships.
• The study provides a computationally efficient method for optimal derating policy computation.