Severity-aware, age-based preventive maintenance optimization under a virtual-age NHPP: evidence from an underground LHD fleet

Preventive maintenance (PM) policies for repairable systems often assume constant effectiveness and uniform failure consequences. This paper integrates age-dependent PM effectiveness with severity-aware estimation to produce availability-optimizing PM schedules that are both condition- and consequence-sensitive.

We develop a virtual-age non-homogeneous Poisson process model in which each PM type \(k\) induces an age-dependent restoration factor φ_k(t) $∈$ [0,1]. Corrective-maintenance (CM) failures enter the likelihood with downtime-derived weights f_c, yielding a severity-weighted pseudo-likelihood. Two calibrated baselines are estimated at OEM intervals: (1) unweighted (OEM) and (2) severity-aware (weighted). We then optimize PM schedules under both unweighted and weighted φ_k(t) using a genetic algorithm, with feasibility bounds subject to operational constraints t₁ $∈$ [200,300]\), t₂ $∈$ [450,550]\), t₃ $∈$ [800,1000]\) hours.

Relative to OEM estimation, severity-aware fitting increases inferred failure intensity and systematically lowers φ_k(t) for light and moderate PMs, producing more conservative reliability and improved model fit. Incorporating age-dependent φ_k(t) reshapes the optimization landscape: weighted schedules favor slightly earlier moderate PM (Type II) while maintaining comparable availability. Back-testing shows closer alignment between predicted and observed failure trajectories, particularly for high-downtime events, demonstrating the value of consequence-aware, age-based PM modeling beyond availability alone.

Severity is proxied by downtime; multidimensional consequence metrics, i.e. safety, cost, quality and schedule, are future work. Sparse counts at extreme ages can widen uncertainty in φ_k(t). Unit heterogeneity suggests hierarchical extensions.

The method operationalizes risk-sensitive PM: planners can jointly tune timing and consequence-aware effectiveness to reduce severe disruptions without over-servicing. The approach is data-driven and CMMS-ready.

This is the first synthesis to jointly (1) model PM effectiveness as a function of age and (2) estimate it via a severity-weighted virtual-age likelihood, then (3) optimize schedules under real feasibility constraints using real downtimes in chronological order.