A Tri-Objective Optimization Framework for Selecting Infrastructure Construction Safety Interventions Balancing Risk, Complexity, and Generalizability

Abstract

The selection of safety interventions in infrastructure projects is a complex decision-making process involving multiple competing objectives. Existing quantitative models often focus on a bi-objective trade-off, such as risk versus cost, neglecting the strategic value of an intervention's applicability across various scenarios. This paper introduces a novel tri-objective optimization framework to address this gap, using data from OSHA accident narratives in heavy construction. The framework balances three objectives for each intervention: (1) Risk, quantified by a severity score derived using a Large Language Model (LLM); (2) Complexity, a metric termed "Semantic Friction," defined as the cosine distance between the textual embeddings of an unsafe action and its safe alternative; and (3) Generalizability, a metric representing an intervention's broad applicability, calculated as its average semantic similarity to all other safe alternatives. Interventions were grouped into homogenous hazard clusters using k-means clustering. The NSGA-III multi-objective evolutionary algorithm was then employed to identify the 3D Pareto-optimal front for each cluster. The framework’s metrics were validated internally, with the generalizability metric demonstrating a statistically significant correlation with linguistic features indicative of broad applicability. This research contributes a third quantifiable criterion, generalizability, to construction safety optimization, providing managers with a nuanced portfolio of solutions balancing risk, implementation complexity, and cross-hazard applicability.