Deep autoencoder model for direct monitoring of bridges subjected to a moving vehicle load under varying temperature conditions

Abstract

Structural health monitoring (SHM) approaches have offered potential solutions for monitoring and analyzing the behavior of transport infrastructures. Recently, vibration-based damage detection using deep learning approaches has received considerable attention in the SHM community. Previous studies on structural damage detection have employed supervised deep learning models, which require large amounts of labeled data. However, acquiring labeled data in practical engineering is challenging, costly, and sometimes impractical. This study uses a deep autoencoder model to extract the damage-sensitive features from acceleration data in the frequency domain without any labels. For this, a numerical example of a concrete highway bridge model subjected to a single-vehicle load under varying temperatures, low-extent damages, vehicle speeds, road surface conditions, and measurement noises was used to evaluate the effectiveness of the proposed method. This study demonstrated that the trained model is sensitive to damage in terms of reconstruction loss. In addition, the damage index (DI) between different damage-sensitive features was calculated using the Gaussian process-based z-scores. The results show that the proposed method has good damage detection capability, with the model struggling only at a higher speed (V3 = 8.33 m/s) with poor road surface roughness, where damage becomes evident after 10% or 15% damage severity. These results emphasize the potential of using the proposed method in practical engineering.