Data Anomaly Detection for Structural Health Monitoring Based on Computer Vision and Transfer Learning
Cheng Pan, Seyedmilad Komarizadehasl, Ye Xia · Report · 2025
During the operation of structural health monitoring (SHM) systems, monitoring data often contains anomalies due to sensor failures and complex environmental factors, which can severely impact structural analysis and evaluation. This paper proposes a data anomaly detection method based on computer vision (CV) and transfer learning (TL). The method can accurately identify various anomalies with only a small amount of labelled data. First, data augmentation techniques are applied to create different types of anomalous data. Then, the generated anomaly data is visualized into images, which are used to pre-train a vision transformer (ViT) network. Finally, a small sample of labelled data from the target bridge is used to fine-tune the pre-trained model for the real-world application. Verification results on the actual bridge demonstrate that the proposed method exhibits strong classification performance.