A Hybrid Machine Learning Framework for Construction Site Safety Assessment and Accident Risk Prediction in Infrastructure Projects

Abstract

The status quo safety management approach is reactive, relying on the traditional, checklist-driven approach to safety in workplaces, including infrastructure projects like highways, bridges, tunnels and metro systems. This paper suggests a hybrid machine learning approach through combining three complementary information streams that are often treated separately: on-site observation of visual information, historical accident narratives and structured project risk factors. A vision stream identifies workers and personal protective equipment (PPE), detects proximity hazards, and derives site imagery based on employees' locations; a language stream extracts unstructured information about incidents and inspections from a free-text database for the purpose of latent causal factor mining; a structured stream models project, task and environmental attributes using gradient-boosted ensembles. A late-fusion reasoning layer is used to fuse the three streams into a single, interpretable site-level accident risk score, and a model-agnostic explainability module extracts the factors responsible for each score, which can then be used by site engineers to inform their actions. We explain the end-to-end architecture, an evaluation protocol for which construction safety data from the public is reported, and a series of baselines from the single-modality and classical approaches. Reported results are sample values intended to illustrate and not measured. The framework will be used as an early, explainable warning proactive decision support tool rather than a post incident analysis tool for infrastructure safety management.