Pointwise multiclass vibration classification for cable-supported bridges using a signal-segmentation deep network

Long-span bridges are inherently vulnerable to vibration, which could result in serviceability failure. Many design codes typically apply a one-size-fits-all approach with predefined thresholds for determining the occurrence of excessive vibration, but these often fail to work properly due to complex vibration patterns. Assessments that are based on vibration type could provide a more accurate assessment. To this end, this study proposes a deep-learning-based framework for the pointwise multiclass vibration classification of long-span bridges. Two abnormal vibration classes are considered here: vortex-induced vibration (VIV) and traffic-induced vibration (TIV). The proposed framework employs a Fourier Synchrosqueezed Transform (FSST) to accomplish time–frequency estimation, a Bidirectional Long-Short-Term-Memory (Bi-LSTM) model for network training, and a median filter for post-processing. In this study, field-monitored wind and weigh-in-motion data from a case study of the Jindo Bridge in South Korea was used to demonstrate the feasibility of the developed framework. The results show that the trained network classified TIVs and VIVs sample-by-sample with an accuracy of 94.35 %. In addition to significant VIVs, the network also successfully detected limited vibrations, which allowed it to outperform statistic-based approaches.