DAM SAFETY: AN INTELLIGENT MODEL FOR ANOMALY DETECTION IN INSTRUMENTATION DATA

Abstract

Dam safety is an area of high criticality in engineering, where early detection of anomalies is critical to prevent disasters. Traditional monitoring methods often analyze instrumentation data in isolation, failing to identify complex, contextual deviations that may foreshadow structural failures. This work proposes and validates an advanced methodology for the detection of anomalies in dam monitoring data, based on the synergy between a robust attribute engineering and the Local Outlier Factor (LOF) unsupervised machine learning algorithm. Using a synthetic dataset that emulates the behavior of multiple instruments (piezometers, water level indicators, Pars-hall flume and rain gauges.) over three years, the model was trained to identify normality patterns in a multidimensional space. The results demonstrate the conclusive superiority of the proposed approach, which achieved an F1-Score of 0.868 and, crucially, an accuracy of 100%, eliminating the occurrence of false positives. In contrast, traditional methods such as Boxplot and Linear Regression performed significantly less. The qualitative analysis confirmed the model's ability to detect contimental anomalies, such as flow peaks in periods of drought, which would be ignored by univariate analyses. It is concluded that the methodology not only increases the reliability of detection, but also provides a basis for diagnosing the nature of anomalies, representing a significant advance for the intelligent automation of dam structural health monitoring.

DOI: https://doi.org/10.56238/sevened2025.029-112