Abstract
Recent decades have witnessed a rise in interest in bridge health monitoring utilizing the vibrations of passing vehicles. However, existing studies commonly rely on constant speeds or tuning vehicular parameters, making their methods challenging to be used in practical engineering applications. Additionally, recent studies on the data-driven approach usually need labeled data for damage scenarios. Still, getting these labels in engineering is difficult or even impractical because the bridge is typically in a healthy state. This paper proposes a novel, damaged-label-free, machine-learning-based, indirect bridge-health monitoring method named the assumption accuracy method (A2M). Initially, the raw frequency responses of the vehicle are employed to train a classifier, and K-folder cross-validation accuracy scores are then used to calculate a threshold to specify the bridge's health state. Compared to merely focusing on low-band frequency responses (0-50 Hz), utilizing full-band vehicle responses can significantly improve the accuracy, meaning that the bridge's dynamic information exists in the higher frequency ranges and can contribute to detecting bridge damage. However, raw frequency responses are generally in a high-dimensional space, and the number of features is much greater than that of samples. To represent the frequency responses via latent representations in a low-dimension space, appropriate dimension-reduction techniques are therefore, needed. It was found that principal component analysis (PCA) and Mel-frequency cepstral coefficients (MFCCs) are suitable for the aforementioned issue, and MFCCs are more damage-sensitive. When the bridge is in a healthy condition, the accuracy values obtained using MFCCs are primarily dispersed around 0.5, but following the occurrence of damage, they increased significantly to 0.89-1.0 in this study.