Data-based feature representation of traffic flow for predicting bridge displacement responses with ensemble learning model

Roadway traffic can significantly affect the behavior of long-span cable-supported bridges, thereby influencing their durability. Characterizing roadway traffic loads is essential for evaluating the performance of bridges. Typically, traffic flow is complex in nature and is often represented using variables such as vehicle volume, vehicle weight, and vehicle speed. However, these variables are rarely analyzed with consideration of the spatial and temporal aspects of traffic flow, despite variations in both dimensions. This study focuses on a data-driven representation of roadway traffic flow using on-site weigh-in-motion (WIM) data, accounting for both time and space scales. The aim is to better represent roadway traffic and improve the prediction of bridge responses. To this end, structural health monitoring (SHM) data from a long-span cable-stayed bridge is utilized, and a prediction task that incorporates roadway traffic characteristics at different temporal and spatial scales is established. The study reveals the necessity of characterizing traffic load in both spatial and temporal scales, as providing excessive details in the traffic flow can complicate the machine learning model and reduce prediction accuracy. With proper representation of traffic loads, accurate predictions of bridge girder and pylon displacement responses are achieved. Improved bridge response prediction can enhance the understanding of bridge behavior under traffic load conditions.