Real-time drive-by bridge damage detection using deep auto-encoder

TLDR

Structural health monitoring of aging bridges focuses on damage detection, and drive‑by methods have recently gained attention because they require only a few sensors on a passing vehicle. The paper proposes an automatic, real‑time bridge damage detection method that operates while a vehicle passes over the bridge. The method trains a deep auto‑encoder on short‑time, frequency‑domain vibration frames from a passing vehicle, extracts damage indicators for unknown bridges, and determines health states in real time, validated on a U‑shaped beam and truck simulator, with an identified damage ratio index used to assess severity. The approach achieved 86.2 % accuracy in detecting bridge damage, and the identified damage ratio index increased with severity, approaching 100 % as damage worsened, although raw indicator values did not correlate directly with severity.

Abstract

Structural health condition monitoring of bridge structures has been a concern in the last decades due to their aging and deterioration, in which the core task is damage detection. Recently, the drive-by method has gained much attention as it only needs several sensors installed on the passing vehicle. In this paper, we proposed an automatic damage detection method, which can be exploited in real time when the vehicle is passing the bridge. There are three steps in the proposed method: (1) The vehicle’s framed short-time vibrations instead of full-length data are utilized for training a deep auto-encoder model; at this stage, not commonly used time-domain accelerations of the passing vehicle, but its selected frequency-domain responses are employed to circumvent the influence of noises, (2) For the bridge with unknown health conditions, damage indicators can be extracted from its passing vehicle’s short-time vibration data using the trained model, and (3) The bridge’s health states are determined by real-time extracted damage indicators. To verify the proposed idea, a U-shaped continuous beam and a model truck are used to simulate the vehicle bridge interaction system in engineering. Results showed that the proposed method could identify the bridge’s damage with an accuracy of 86.2% when different severity was considered. In addition, it was observed that higher damage severity could not be revealed by greater values of damage indicators in the laboratory test. Instead, a novel index called identified damage ratios was employed as a reference for assessing the severity of the bridge’s damage. It was shown that with the increase in damage severity, the index would increase and gradually approach 100%.