Ensemble-based anomaly detection for dam safety monitoring data

Safety monitoring functions as the “doctor” of concrete dams, ensuring their safety by providing long-term, periodic, or continuous historical observation data. Through mathematical statistics, data mining, and expert knowledge, the working state of concrete dams can be comprehended, their health status diagnosed, and real-time safety monitoring realized. However, raw monitoring data is complex and requires meticulous analysis to identify outliers. Mining and analyzing this data are important for accurately detecting anomalies. This study explores the application and comparative efficiency of three advanced outlier detection techniques: Isolation Forest (IF), Local Outlier Factor (LOF), and the ensemble approach “cascade” IF-LOF. The ensemble approach leverages the complementary capabilities of IF and LOF techniques to provide a more robust and accurate anomaly detection (AD) system. The proposed hierarchical IF-LOF model was evaluated on real-world dam displacement datasets, where it demonstrated superior performance in identifying subtle anomalies that individual methods often missed, achieving an accuracy rate of 98% and F1-scores consistently above 92% across all monitoring points. This study emphasizes the value of combining techniques to overcome single-algorithm limitations, especially with high-dimensional, noisy data. It enhances AD reliability, advancing safety monitoring for concrete dams. The findings highlight the potential of the ensemble methods in structural health monitoring and support further research on infrastructure safety and maintenance optimization.