A Temperature-driven One-Class Support Vector Machine Method for Anomaly Detection

Competent data-driven anomaly detection methods should catch the meaningful changes in measurements due to the structural abnormity. However, the distinct thermal effect may produce significant temperature-related fluctuations in the strain measurements when compared with the reflections due to the real structural damages. This paper presents a temperature-driven one-class support vector machine method, designated as Td-OCSVM, for anomaly detection, which introduces the idea of blind source separation (BSS) for thermal feature extraction and further employs OCSVM for anomaly detection. First of all, the temperature-related strain variations can be investigated and revealed by employing independent component analysis based on the criteria of maximization nongaussianity, which is one of the most popular solutions for BSS problem. Afterwards, the OCSVM is adopted for anomaly detection on those separated temperature-induced responses. In case study, the Ricciolo curved viaduct in Switzerland is utilized for the purpose of evaluating the proposed method. This viaduct had been monitored in both construction and in-service period. The data obtaied during the construction period is labelled as anomalous condition, which is expected to be identified by Td-OCSVM. The performance of Td-OCSVM is also compared with OCSVM only with temperature-driven process. The interpretation results can demonstrate the outperformance of Td-OCSVM with the higher detectable ability when compared with OCSVM.