A vision transformer-based method for predicting seismic damage states of RC piers: Database development and efficient assessment

The structural safety of bridges, particularly the ability to predict the damage states of reinforced concrete (RC) piers under seismic action, has become a critical issue in structural engineering. This study employs deep learning techniques to enable efficient prediction and assessment of damage states in-service RC bridge piers subjected to seismic events. To support model training, a parametric sample set of 100 bridge piers is generated using Latin Hypercube Sampling, leading to the development of a comprehensive seismic response database containing 66,000 samples across 15 defined damage states. These databases account for inherent seismic randomness, complex failure modes, and time-dependent composite evaluation indicators. A novel deep learning framework, CC-ViT, based on the Vision Transformer architecture, is proposed. This framework integrates Continuous Wavelet Transform, Context Anchored Attention, and DropKey techniques to enhance feature extraction and model generalization. Multiple models are trained and evaluated in a supervised learning setting. Comparative analysis reveals that CC-ViT achieved the highest test accuracy at 85 %. Grad-CAM-based interpretability analysis further confirms that CC-ViT effectively captures critical regions in the seismic response spectrum, supporting informed and explainable decision-making. To facilitate practical implementation, an end-to-end interactive software tool has been developed for efficient prediction of pier damage states. The findings contribute valuable insights for data-driven decision-making aimed at enhancing infrastructure safety and maintenance in smart cities.