Anomaly detection of sensor faults and extreme events based on support vector data description

Abstract

Structural health monitoring (SHM) systems generate a massive amount of sensing data. On one hand, sensor faults may cause the measurement data to have low fidelity. On the other hand, extreme events, such as typhoons or earthquakes, may cause the monitoring data look “abnormal.” These abnormal data, however, are closely related to the structural safety condition and require special attention. This study proposes an automatic and efficient anomaly detection methodology based on support vector data description (SVDD) to simultaneously detect anomalies caused by sensor faults and extreme events. The SVDD trained by a single pattern can divide the feature space into one-versus-the rest. Several decision boundaries are defined to enclose normal data and common sensor fault patterns, forming an equivalent multi-class classifier to classify common sensor fault types and detect unknown patterns. Next, multiple sensor faults and extreme events are separated from the unknown patterns. Multi-label data are detected based on the local features, while extreme events are recognized by the correlation of different sensors. The proposed method is finally applied to datasets collected from two SHM systems. Results show that the sensor anomalies in the systems are detected with high efficiency and accuracy, and extreme events are separated as a special pattern from the normal, common abnormal, and unknown patterns.