ABSTRACT
This study introduces an unsupervised approach to categorize the acceleration time history of a highway bridge into three distinct classes: vehicle events, the bridge’s free vibration, and noise. We employ a combination of wavelet scattering networks and principal component analysis for extracting features, while the Gaussian mixture model is utilized for clustering the data segments into these classes. The results are compared to a manually adjusted database of events identified through computer vision, aligned with the acceleration data to define event intervals. The effectiveness of this method in identifying events from the bridge’s acceleration data is evident. It successfully detected all events without omission and accurately separated the free vibrations.
