Decision-Focused Predictive Maintenance: Bridging the Gap between Data-Driven RUL Prognostics and Maintenance Planning

Data-driven predictive maintenance (PdM) increasingly leverages machine learning techniques to predict remaining useful life (RUL) using abundant sensor data, supporting effective maintenance planning. However, most existing research follows a predict-then-optimize (PtO) paradigm, focusing on prognostic accuracy while overlooking how RUL predictions affect maintenance decisions. We propose a novel Decision-Focused Predictive Maintenance framework that bridges the gap between RUL prognostics and maintenance planning. This framework creates an end-to-end pipeline that directly connects RUL estimation to maintenance actions. An experiment using the CMAPSS dataset demonstrates that our framework achieves a 9.3% reduction in maintenance costs compared to the PtO approach. This improvement is primarily attributed to the avoidance of unnecessary preventive maintenance, leading to a reduction in average lifetime waste due to preventive maintenance from 20.9 to 11.3 cycles. More importantly, we highlight the distinction between DFPdM and PtO by analyzing the quantile levels of RUL labels and maintenance decisions, demonstrating that DFPdM exhibits greater consistency in unifying estimation and optimization. Interestingly, we also observe that DFPdM achieves an acceptable prognostic accuracy, despite not being the primary training objective. This prognostics accomplished by recalibrating a specific quantile of the estimated distribution, rather than relying on the expectation or median as is common in conventional approaches.

Keywords: Predictive maintenance, Prognostics, Decision-focused learning, Contextual decision-making.