Deep learning-based gross vehicle weight estimation in Bridge Weigh-in-Motion by using sensors in one cross-section

Abstract

Gross Vehicle Weight (GVW) estimation plays a crucial role in ensuring the safety, maintainability, and sustainability of road transportation by identifying and filtering out overloaded vehicles. Bridge Weigh-in-Motion (B-WIM) systems enable the determination of axle loads, vehicle speeds, axle spacings, and other vehicle parameters as they cross a bridge, using data from strain gauges installed beneath the bridge deck. This paper proposes a novel deep learning-based GVW estimation method designed for B-WIM systems equipped with sensors at a single cross-section. Unlike conventional axle load estimation-based GVW estimators, the proposed method does not rely on vehicle speed estimation or axle detection steps. The method is evaluated on an annotated dataset of 91 vehicles measured on the Monostori Bridge. Results demonstrate B+ accuracy with a Mean Absolute Percentage Error of 2.47% for GVW estimation in accordance with the COST 323 Weigh-in-Motion classification standard. Furthermore, the proposed solution can be integrated into standard B-WIM pipelines using ensemble models. Tests on the same dataset indicate that the ensemble approach may outperform existing B-WIM pipelines in GVW estimation accuracy by reducing the Mean Absolute Percentage Error by 0.1%.