Application of Anomaly Detection to Real World Monitoring Data of a Rail-Road Bridge

Abstract

The growing importance of automatic bridge condition monitoring, driven by the need to sustain infrastructure and reduce construction-related CO2 emissions, has led to increased deployment of sensor technologies on bridges. However, the overwhelming volume of data generated necessitates innovative approaches. Machine learning (ML), particularly the Regression Based Thermal Response Prediction Method (RBTRPM), offers a promising solution. RBTRPM employs ML models to predict structural responses from temperature measurements. These predictions are compared with actual responses to identify anomalies. While introduced in 2014, its practical applicability has been limited. This paper successfully translates RBTRPM into practice, exploring key considerations. A measurement duration of at least six months is recommended, though a longer period would provide even better results. It is also important to emphasize the preference for using only structural influences in the response predictions. The study introduces the Peak over Threshold method for threshold determination applied on RBTRPM and employs moving metrics for concise result presentation, providing valuable insights for the implementation of RBTRPM in real-world bridge monitoring scenarios.