Crowdsensing-based automatic bridge health condition assessment using drive-by measurements and deep learning

Abstract

In recent decades, assessing the structural health conditions of aging bridges has emerged as a significant concern. A recent drive-by measurement method has attracted substantial attention, in which only several sensors are installed on crowdsensing vehicles rather than bridges, providing a more economical and convenient solution. This paper proposes an automatic bridge condition assessment framework incorporating drive-by measurements and deep learning techniques. The methodology involves collecting and segmenting accelerations from a vehicle passing a healthy bridge into short-time overlapped frames. Over multiple vehicular passes, all frames are then transformed into frequency-domain responses, forming the input for training an unsupervised deep learning model. The model is then trained to reconstruct the input using these frequency-domain responses. In assessing the bridge with an unknown health state, the trained model is employed to reconstruct the passing vehicle's new short-time frames, and the response construction error automatically determines the bridge's health condition. Experimental validation utilizing a laboratory bridge and scaled truck demonstrated that the trained model could consistently identify a healthy bridge during passages, with larger reconstruction errors indicating that the bridge was damaged. The innovative framework showcased promise for efficient and reliable bridge health condition assessment.