Predictive maintenance optimization for industrial equipment via reliable prognosis and risk-aware reinforcement...

Area of Science:

• Industrial Engineering
• Machine Learning
• Reliability Engineering

Background:

• Predictive maintenance (PdM) is vital for industrial equipment performance and cost reduction.
• Remaining Useful Life (RUL) prediction is a key component of effective PdM strategies.
• Existing methods often struggle to adequately address uncertainty and risk in maintenance decisions.

Purpose of the Study:

• To develop an intelligent PdM framework integrating RUL prediction and distributional reinforcement learning.
• To improve the accuracy of RUL predictions and quantify associated uncertainties.
• To optimize maintenance scheduling by considering long-term returns, risk, and cost-efficiency.

Main Methods:

• A probabilistic neural network was employed for accurate RUL prediction and uncertainty quantification.
• A Quantile Regression Deep Q-Network (QR-DQN) agent was utilized for distributional reinforcement learning.
• Risk-sensitive decision rules were incorporated to manage uncertainty and failure probability.

Main Results:

• The proposed framework demonstrated superior performance compared to conventional baselines in complex system degradation scenarios.
• Achieved significant reductions in catastrophic failures and optimized maintenance schedules.
• Enhanced overall system reliability through effective risk management and timely interventions.

Conclusions:

• The intelligent PdM framework offers a robust approach to managing industrial equipment maintenance.
• Integrating RUL prediction with distributional reinforcement learning effectively addresses uncertainty and risk.
• The approach leads to improved operational efficiency, cost savings, and enhanced equipment longevity.