Anomaly detection on a concrete arch dam: Exploring the influence of training periods

Abstract

The condition monitoring of civil engineering structures involves systematic assessment of their performance and integrity, which may decrease over time due to external actions and material degradation. Therefore, the early detection of structural changes is essential to guarantee well-planned management of the structure’s operation and to perform timely interventions when needed. To detect damage or ongoing processes of deterioration, techniques such as visual inspections, non-destructive testing and continuous data-based structural health monitoring (SHM) can be used. In the case of the latter, key properties can be tracked, initially to build databases and subsequently to create damage sensitive features that can support the identification of statistical anomalies. The approach most used in SHM for the detection of anomalies is based on the comparison between new data and data previously recorded during a period when the structure is thought to be in a perfectly healthy state (known as training period). In this context, this work takes advantage of the combination of numerically simulated damage with real experimental data obtained from the continuous monitoring of a concrete dam, to explore the relation between the length of the training period, the intensity of the simulated damages and the ability of statistical tools to detect these anomalies. Additionally, the impact of including damaged-state data in the training period is explored.