Performance of data-based models for early detection of damage in concrete dams

Abstract

A failure of a massive concrete dam could cause catastrophic consequences. The purpose of monitoring is to detect anomalies and damage at an early stage to prevent failure. Data-based models for anomaly detection are based on the hypothesis that the behaviour of an undamaged dam will follow an expected pattern, and deviation from this pattern is an indication of damage. In this study, simulations were used to create time series for an undamaged dam and three different damage scenarios at three different locations in the dam body. Three common data-based models were used to predict a dams crest displacements, both on the generated artificial data and the corresponding measurements from the dam. Prediction bands for future measurements were created, and the ten time-series were used to test the ability to detect damage. All models could detect instantaneous damage but struggle to detect progressive damage; the Neural network outperforms the two regression models. The choice of the mathematically optimal threshold limit leads to a large number of false alerts. Requiring three consecutive values outside the threshold before an alert is issued, increases the possibility to receive an early alert compared to the standard approach where observations are classified individually.