Multi-Horizon Machine Learning for Bridge Deterioration Prediction and Risk-Based Maintenance Decision-Making
TL;DR
An integrated multi-horizon framework for bridge management agencies that eliminates trajectory incoherence entirely, and allocates a per-period budget by maximising discounted, consequence-weighted risk averted, solved as a linear-programming relaxation of a multiple-choice knapsack is proposed.
Abstract
Bridge management agencies must decide not only which structures to treat, but when to treat them over a multi-year programme. Most machine-learning pipelines for bridge condition, however, forecast a single fixed horizon, report only marginal uncertainty, and feed a top-k selection rule that ignores treatment cost and timing. This paper proposes an integrated multi-horizon framework. (i) A horizon-conditioned learner (HCM) is trained on horizon-stacked data with a monotone constraint on the horizon variable, which guarantees coherent deterioration trajectories and, as a by-product, a valid discrete-time hazard and remaining-useful-life estimate. (ii) Predictive uncertainty is certified at the trajectory level by a locally adaptive max-norm conformal band, and calibrated first-passage risks are kept monotone by a coherence-preserving isotonic projection. (iii) A multi-horizon scheduling policy (MHSP) allocates a per-period budget by maximising discounted, consequence-weighted risk averted, solved as a linear-programming relaxation of a multiple-choice knapsack. On a 11,963-bridge NBI-schema longitudinal benchmark with explicit intervention counterfactuals, the HCM eliminates trajectory incoherence entirely (0.0% of bridges, versus 32.2–53.5% for independent per-horizon models and 43.8% for a recursive model) while also being the most accurate learner at every horizon (ROC-AUC 0.952 to 0.903 for h = 2 to 10 years). Per-horizon 90% conformal bands attain only 73.6% joint trajectory coverage; the proposed adaptive max-norm band reaches 89.9% joint coverage while being 9.9% narrower than a Bonferroni band. At a 15% programme budget, MHSP averts 51.9% of consequence-weighted Poor bridge-years, against 33.1% for a single-horizon risk×consequence index, 22.8% for condition-first and 10.6% for age-first triage, reaching 79.9% of a perfect-foresight upper bound. Code and benchmark are fully seeded and reproducible.