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

Predictive maintenance (PdM) has seen significant advances through machine learning, yet its practical deployment remains challenged by data scarcity, system complexity, and uncertainty in cost-related decisions. The majority of current PdM strategies are concerned with enhancing Remaining Useful Life. Prediction accuracy of (RUL) in isolation and with maintenance scheduling as a problem (secondary or fixed). This study proposes a component based, decision oriented predictive maintenance (PdM) approach that links Remaining Useful Life (RUL) to optimization of maintenance. The two-stage framework proposed anticipates the component-specific RUL prediction where Long Short-Term Memory (LSTM) models was used to predict the Remaining Useful Life (RUL) of individual components. To address sparsity in failure data, Wasserstein Generative Adversarial Networks Gradient Penalty (WGAN-GP) were utilised to fill in run-to-failure sequences, stabilizing downstream modeling. In the second step, similar components in terms of Remaining Useful Life (RUL) degradation are clustered together by Density-Based Clustering Space (DBSCAN), which allows opportunistic maintenance. A decision on maintenance is then optimized cost-conscious grid search which works on fitted RUL distributions and a normalized. Not only based on point RUL estimates, but on a risk proxy. Empirical experimentation across multiple industrial components of a water bottling plant system indicates that the proposed approach continually reduces corrective failures and normalized maintenance costs as opposed to non-clustering approaches such as random choice and fixed choice of maintenance point. Sensitivity analysis also indicates that the optimal maintenance levels be consistent over a vast spectrum of cost assumptions, which emphasizes the resilience of the structure in economic uncertainty. Overall, this study contributes a robust and scalable maintenance literature that combines data augmentation, component based clustering, and risk aware optimization. This helps advance predictive maintenance practices into a more practical, cost aware decisions.