Anomaly detection method for experimental bridge monitoring data based on multimodal feature fusion

Structural Health Monitoring (SHM) systems have been widely deployed in large-scale bridge structures. However, due to factors such as sensor malfunctions, environmental disturbances, and signal noise, SHM data often contain various types of anomalies, which can significantly compromise the accuracy and reliability of structural condition assessments. To address the challenges posed by data imbalance and limited model generalisability in existing anomaly detection methods, this study proposes an anomaly detection and classification approach based on multimodal feature fusion. Specifically, SHM data are first transformed into visual representations, from which image features are extracted using Convolutional Neural Network (CNN). These features are then fused with statistical features extracted directly from the raw signals, enhancing the model’s capability and robustness in identifying complex anomaly patterns. A case study is conducted using real-world monitoring data collected from the SHM system of a long-span cable-stayed bridge, where a small-sample, imbalanced dataset is constructed to evaluate the proposed method. Furthermore, a mixed training strategy based on multi-bridge dataset is introduced, leveraging datasets from multiple bridges to improve the model’s generalisation performance across different structural systems. Experimental results demonstrate that the proposed method effectively detects and classifies anomalies in SHM data, achieving high accuracy and robustness.