A drift-aware explainable artificial intelligence framework for dynamic construction risk governance using multi-domain fuzzy inference and Bayesian updating

Abstract

Construction projects operate in highly dynamic environments where project risks evolve continuously as new information becomes available. Conventional risk assessment approaches are predominantly static and provide limited capability to adapt to temporal changes or generate transparent decision support. This study aims to propose a drift-aware dynamic explainable artificial intelligence (DXAI) framework that integrates multi-domain fuzzy inference with Bayesian updating to enable adaptive and interpretable construction risk governance. The proposed framework was evaluated using a publicly available Building Information Modelling–Artificial Intelligence (BIM–AI) Integrated Dataset comprising 1,000 construction projects. Four latent stress domains – performance, safety, structural–operational and environmental – were extracted from project variables and integrated through a Mamdani fuzzy inference system to generate prior risk estimates. Bayesian updating incorporated evolving project evidence to produce posterior risk estimates, while temporal risk drift quantified dynamic changes in project conditions. An explainability layer generated interpretable governance recommendations supported by statistical validation, sensitivity analysis, bootstrap resampling and ranking stability assessment. The framework demonstrated adaptive risk estimation through Bayesian belief updating, identified projects exhibiting significant temporal risk drift and generated transparent governance recommendations. Validation results confirmed robust model behaviour, stable project prioritisation and statistically significant improvement in dynamic risk assessment. The study introduces an integrated DXAI framework that combines fuzzy inference, Bayesian updating, temporal risk drift analysis and explainable governance support for dynamic construction risk management.