A two-stage framework for cost-sensitive predictive maintenance using deep learning, GANs, and risk-aware clustering

Area of Science:

• Industrial Engineering
• Machine Learning
• Data Science

Background:

• Predictive maintenance (PdM) faces challenges in data scarcity and cost-uncertainty.
• Current PdM focuses on Remaining Useful Life (RUL) prediction, often neglecting maintenance scheduling optimization.
• A component-based, decision-oriented approach is needed to integrate RUL with maintenance optimization.

Purpose of the Study:

• To propose a novel two-stage PdM framework linking RUL prediction with maintenance optimization.
• To address data scarcity using Wasserstein Generative Adversarial Networks Gradient Penalty (WGAN-GP).
• To enable opportunistic maintenance through component clustering and optimize decisions using a cost-aware grid search.

Main Methods:

• Component-specific RUL prediction using Long Short-Term Memory (LSTM) models.
• Data augmentation for failure data sparsity with WGAN-GP.
• Component clustering using Density-Based Clustering Space (DBSCAN) for opportunistic maintenance.
• Cost-aware grid search optimization based on RUL distributions and risk proxy.

Main Results:

• The proposed approach significantly reduces corrective failures and normalized maintenance costs.
• Empirical results show continuous cost reduction compared to non-clustering methods.
• Sensitivity analysis confirms consistent optimal maintenance levels across various cost assumptions, highlighting economic resilience.

Conclusions:

• The study presents a robust and scalable PdM framework combining data augmentation, clustering, and risk-aware optimization.
• This approach advances PdM towards more practical and cost-effective decision-making in industrial settings.
• The integrated method enhances maintenance strategies by considering RUL, component similarity, and economic factors.