Pointwise vortex-induced vibration detection: Learning from synthetic time-series data

Long-span bridges, which are characterized by high flexibility, low natural frequencies, and damping ratios, are inherently susceptible to vortex-induced vibrations (VIVs). Traditional serviceability assessments rely on statistical threshold-based approaches, which are impractical for covering the various dynamic characteristics and operational conditions of bridges. Despite the successful implementation of deep-learning algorithms, challenges in acquiring sufficient and accurate training labels must be resolved for real-world applications. This study proposes a framework for pointwise VIV detection that is free from manual annotations through data synthesis. In contrast to traditional supervised methods, involving manual labeling and field data acquisition, the proposed framework synthesizes two vibration cases: (a) VIVs by generating amplitude-modulated sinusoidal waves; and (b) non-VIV datasets using generative deep-learning networks. The feasibility of data synthesis is first demonstrated by comparing the similarity of synthetic data with the corresponding field-monitoring datasets. Subsequently, the effectiveness of the sequence-to-sequence classification model trained solely on synthetic datasets is validated through a comparative analysis with a counterpart trained on manually labeled datasets, obtained from a cable-supported bridge. Validation on an additional bridge confirms the potential for expansion of the proposed framework alongside its potential early warning applications based on probability distributions. The proposed framework achieves a performance comparable to that of conventional supervised approaches while resolving labor-intensive and subjective aspects.