Interval-dependent maintenance policy optimisation of a system subject to multi-level preventive maintenance types: a non-parametric, severity-aware approach

Purpose This paper develops a unified framework for preventive maintenance (PM) modelling that simultaneously accounts for downtime severity and interval-dependent effectiveness. The aim is to generate maintenance policies that are both risk-aware and operationally efficient for repairable industrial systems with multiple PM types. Design/methodology/approach Downtime-based severity weights are embedded within a virtual-age non-homogeneous Poisson process (NHPP) with multiple PM types. Interval-dependent restoration factors are estimated non-parametrically using interval discretisation and isotonic regression. Model parameters are obtained by severity-weighted maximum pseudo-likelihood using global metaheuristics and local refinement. Where within-interval variation is weak, conservative constant restoration estimates are used in the constrained availability optimisation. Optimised schedules are compared with original equipment manufacturer (OEM) baseline policies. Findings All three optimisation algorithms favour PM intervals near the upper operational bounds. Under severity weighting, the best schedule, approximately (298.5, 547.2, 1000) h, increases availability from 0.7368 to 0.7513. The light (Type I) PM interval has the greatest influence on availability, while major (Type III) PM timing is comparatively insensitive. Severity-aware calibration also improves model fit relative to the unweighted baseline when evaluated under the severity-weighted pseudo-likelihood objective. Research limitations/implications The dataset is limited to one subsystem and four trucks, and severity is represented solely by downtime. Sparse PM observations constrain recovery of detailed non-parametric restoration profiles, particularly for the less frequent PM types. Broader fleets, richer severity measures and hierarchical treatment of fleet heterogeneity would strengthen inference, permit fuller estimation of interval-dependent restoration and improve generalisability. Practical implications For the studied fleet, severity-aware calibration combined with interval-effectiveness diagnostics supports longer PM intervals within the tested operational bounds, reducing planned downtime and increasing availability. Scheduling precision should be concentrated on moderate PM, which exerts the greatest influence on availability, while major PM can be scheduled more flexibly. The framework also provides practitioners with a data-driven means of testing whether maintenance effectiveness varies materially with intervention timing. Originality/value This study is the first to jointly integrate downtime-based failure severity, interval-dependent restoration and multiple PM types within a non-parametric NHPP virtual-age framework. It links severity-aware parameter estimation with constrained availability optimisation while retaining the interpretability of classical imperfect-maintenance models, providing a bridge between risk-sensitive reliability modelling and implementable industrial maintenance policy.