Multi-type data anomaly detection for structural health monitoring using interpretable machine learning methodology

Structural health monitoring (SHM) systems utilize multi-type monitoring data, particularly data collected under extreme events, to assess safety and serviceability during structures’ lifetime. However, sensor faults can compromise the reliability of structural analysis. This study proposes a novel interpretable data anomaly detection framework to identify both extreme events and sensor faults across various data types in SHM systems. Sensitive parameters and features are first defined based on the similarities and differences between normal pattern, data under extreme events, and sensor faults. Two machine learning methods, Kernel density estimation and support vector data description, are then used to help determine anomaly detection thresholds for these parameters and features. Based on the thresholds, sensor faults and extreme events in acceleration data can be simultaneously detected. Subsequently, the results of extreme event detection in acceleration data can assist in separating extreme events from sensor faults in wind speed and strain data. Two datasets from the SHM systems of a footbridge and Tsing Ma Bridge are used to validate the proposed method. Results demonstrate that the proposed anomaly detection framework can identify anomalies across multiple types of sensors in SHM systems with high efficiency and accuracy.