Evidence-theoretic multi-sensor fusion with spatial semantic coupling for real-time hazard assessment in dynamic construction environments

Abstract

Construction sites are characterized by heterogeneous, dynamically evolving hazard sources whose spatial interactions remain poorly captured by existing monitoring approaches. Two longstanding obstacles are the unreliable combination of conflicting evidence from multi-modal sensors and the absence of three-dimensional spatial reasoning in quantitative risk estimation. Here we present a unified sensing and warning framework that addresses both challenges. The framework introduces a credibility-weighted evidence fusion scheme based on the Jousselme distance, which remains robust under high-conflict sensor inputs; a spatial-semantic risk quantification model that embeds exponential distance attenuation and component-level occlusion correction into a three-dimensional building representation; and a four-level warning strategy whose thresholds adapt to construction phase, temporal context, and historical warning performance through an exponential moving average self-correction mechanism. Validation on two public benchmark datasets and a reproducible simulated construction scenario shows that the proposed fusion scheme achieves F1-scores of 0.862 and 0.842, compared with 0.743 for the classical Dempster–Shafer rule; the full spatial model attains a grade-agreement rate of 0.848 versus 0.749 without spatial semantics; and the adaptive warning strategy reduces the miss rate from 0.251 to 0.168 under dynamic risk evolution. End-to-end latency remains within 437 ms at a 500-node scale. The results establish a reproducible and generalizable pathway for translating multi-source sensing into spatially-grounded, quantitative safety decisions.

The authors' abstract, as published at the source. Scientific Reports, 2026 · DOI ↗