Using supervised learning techniques to automatically classify vortex-induced vibration in long-span bridges

Owing to a capacity for high flexibility and low damping, long-span bridges are subjected to vortex-induced vibrations (VIVs) under operational conditions. Longterm monitoring data with machine-learning algorithms indicate the potential for automating the VIV assessment of long-span bridges. These methods require a significant amount of labeled data, whereas obtaining such data is normally not feasible owing to the limited availability of VIV datasets. This study leverages supervised learning techniques to develop an automatic classification method for VIVs. To address manual data labeling and develop an optimum model, a three stage strategy is presented: 1) Semi-supervised labeling, 2) deep neural network (DNN) training, and 3) identification of an optimum parameter range. First, semi supervised labeling is employed to automatically label the dataset into either VIV or non-VIV classes. Second, a DNN model is trained using the wind and vibrational features of labeled data. Finally, the optimum parameter range is determined by analyzing the peak factor distribution, confusion matrix, and corresponding velocity-amplitude curve of the classified test datasets. An application of the model to a long-span, cable-stayed bridge is illustrated to assess the classification performance based on actual monitoring data. The DNN with the suggested labeling process demonstrates consistent and accurate detection of VIVs.

Automated summary

This study presents an automatic classification method for vortex-induced vibrations (VIVs) in long-span bridges using machine learning algorithms. A three-stage strategy, including semi-supervised labeling, deep neural network training, and determination of the optimum parameter range, is proposed. The method demonstrates consistent and accurate detection of VIVs, as illustrated by its application to a cable-stayed bridge based on actual monitoring data.