Abstract: To improve the identification accuracy of vortex-induced vibration (VIV) in the main girder of long-span cable-stayed bridges, this study proposes a VIV identification method that combines Gramian angular fields (GAF) with a convolutional neural network (CNN), aiming to overcome the limitations of traditional statistical-feature-based methods in capturing transient vibration characteristics. First, one-dimensional vibration time series are transformed into two-dimensional grayscale images through polar coordinate mapping and Gramian matrix construction. This representation preserves both global temporal relationships and local transient features, enabling a structured mapping from the time-series domain to the image domain and addressing the lack of globally unified representation in conventional time-frequency feature fusion methods. The resulting GAF images are then fed into a structured CNN for end-to-end feature learning and classification, with model robustness and generalization further improved through hyperparameter tuning. Validation based on field monitoring data from an actual bridge shows that the proposed method is more effective than conventional statistical-feature-based approaches in identifying sudden VIV events. It also outperforms several representative machine learning and deep learning baseline models in terms of accuracy, recall, and F1-score. Further correlation analysis with wind field data indicates that the detected VIV events mainly occur within the low-to-moderate wind speed range of 2.0~9.0 m·s⁻¹, which is consistent with the physical excitation mechanism of VIV.