Transfer learning-based open-set data anomaly detection for structural health monitoring practice

The rapid growth of sensing technologies has enabled large-scale deployment of structural health monitoring (SHM) systems and produced massive data streams. However, sensor malfunctions and environmental disturbances introduce anomalies that impede reliable structural assessment. This paper proposes a transfer learning (TL)-based data anomaly detection method using vision transformer (ViT) model. This methodology can effectively identify various types of anomalies while requiring only minimal labeled data. Furthermore, the integration of the ViT model with the class anchor clustering open set recognition method addresses the issue of incomplete training sample categories. The method begins with the generation of acceleration data through numerical simulation, followed by data augmentation techniques to construct 10 representative types of anomalies that reflect common time-domain anomaly patterns. These time-series signals are then transformed into scatter-time images and used as input to pre-train a benchmark and transferable ViT for anomaly detection (BAT-ViT). Subsequently, a small amount of labeled data from the target bridge is used to fine-tune the BAT-ViT for real-world application. To demonstrate the robustness and transferability of the proposed method, SHM data from two different large-span bridges are utilized for validation across bridge structures. The results demonstrate that with only 100 labeled data samples from the actual bridge, the fine-tuned model achieves an accuracy of 95.6% through TL, significantly outperforming conventional convolutional neural network and long short-term memory models. Moreover, the model also achieves over 92% accuracy in open set recognition, even when using a limited transferred dataset. The method delivers robust and adaptable performance, especially when training data are scarce or resolution settings vary, and greatly reduces the need for manual data labeling.