Vibration-Based Structural Damage Detection with a Convolutional Transformer Network

Deep learning (DL) algorithms have significantly improved the accuracy of vibration-based structural damage detection (SDD). However, conventional DL methods are difficult to extract complex discriminative damage features, and these methods cannot quantify damage severity for more accurate damage assessment. To address these limitations, this study proposes a vibration-based SDD method with a convolutional transformer (ConvFormer) network, which is designed for damage identification, localization, and quantification. This approach integrates one-dimensional convolutional neural networks with Transformer encoders to effectively extract spatial and temporal correlation features from raw vibration acceleration signals. Specifically, the network is employed for multi-classification tasks to identify and localize damage, while also performing regression tasks to predict damage severity. The proposed method is compared with several baseline models and state-of-the-art methods using simulated data from a steel frame, laboratory data from a grandstand simulator, and field experimental data from the full-scale Z24 Bridge. The ConvFormer model exhibits superior detection performance in different scenarios. Furthermore, the reliability of the proposed method is validated through experiments on noise immunity, few-shot learning, and unbalanced sample.