Predicting the Dynamic Behaviour of a Concrete Dam using Statistical and Machine Learning Models

Abstract

Operational Modal Analysis is a reliable methodology for the assessment of civil engineering structures, allowing for the accurate definition of their dynamic behavior. Additionally, since it does not require the use of artificial excitation, it becomes a cost-effective choice for the performance of singular tests, as well as a consistent option for the long-term continuous monitoring of structures. Nevertheless, the modal characteristics of structures are affected by environmental and operational conditions, concealing the variations that could emerge due to abnormal behavior. With respect to concrete dams, factors such as temperature and the level of water in the reservoir exert a pronounced influence in the evolution of natural frequencies, increasing data variability and camouflaging the behavior that would be discerned under stable conditions. In this context, the current study seeks to examine the ability of statistical and machine learning tools to mitigate the effects of external conditions on the modal properties of concrete dams, specifically natural frequencies. To achieve this objective, the efficiency of methods incorporating measurements of variables impacting the structure, such as Multiple Linear Regressions and Neural Networks, is compared to that of tools not needing these inputs, as is the case of Principal Components Analysis and the Minimum Mean Square Error estimator. Experimental data obtained during the continuous dynamic monitoring of a concrete dam in Portugal is used as a case study.