Construction-to-lifecycle risk intelligence: a noise-robust Physics-Informed Temporal Graph Transformer and Bayesian Markov-Switching sample-selection framework for reinforcement and repair decisions

Abstract

Construction projects require sequential decisions on structural reinforcement during execution and repair prioritization during operation under heterogeneous conditions, incomplete inspections, temporal dependence and non-engineering constraints. This study develops a construction-to-lifecycle risk-intelligence framework integrating a Physics-Informed Temporal Graph Transformer (PITGT), contrastive noise filtering, uncertainty-aware ensemble learning and a hierarchical Bayesian Markov-switching sample-selection econometric model. The PITGT captures nonlinear temporal, spatial and network dependencies across construction sections, structural components, environmental conditions and intervention histories, while the noise-filtering layer separates engineering-driven repair demand from financially or operationally distorted decisions. The econometric module estimates latent deterioration regimes, time-varying repair thresholds, project heterogeneity and endogenous selection between observed maintenance actions and actual structural need. The framework generates four-level reinforcement recommendations, probabilistic repair-urgency scores and lifecycle intervention priorities under budget uncertainty. Evaluation yielded 93.8% reinforcement-classification accuracy, a macro-F1 of 0.921, a transition-event F1 of 0.816 and a repair-urgency AUC of 0.956. The framework provides explainable, economically grounded decision support across the construction-to-maintenance lifecycle.