Data anomaly detection for structural health monitoring using the Mixture of Bridge Experts

Structure health monitoring systems (SHMs) play a crucial role in understanding the condition of structures. However, owing to various uncertain factors, sensor data may be anomalous, posing a great challenge to the real-time capture of dynamic characteristics of bridges. Hence, detecting anomalous data is crucial for SHM systems. This paper proposes a lightweight model to enable SHMs to detect anomalous data automatically and efficiently. This method combines three modules: Bridge Signal Transformer (BST), Mobile Vision Transformer (MobileViT), and Mixture of Bridge Experts (MoBE). Firstly, the acceleration data is converted to an image, and the MobileViT module is used to detect the anomalous data in the converted image. Since the converted image loses the absolute information of the data, this paper proposes a BST module, which uses directly the information of the data to identify the data type. MoBE module combines perfectly the advantages of the two modules to detect various types of anomalous data accurately. Experiments on an acceleration dataset from an extra-long-span railway cable-stayed bridge validate the advantages of the proposed method. This method performs well in any period, proving its generalization ability.