Adaptative warning thresholds in dam safety using kernel density estimation

Abstract

This study presents a novel adaptive warning threshold approach for dam safety monitoring, employing Kernel Density Estimation (KDE), a statistical method for estimating the probability density function of continuous random variables. This technique adjusts thresholds based on data density and residual variability. Traditional fixed-threshold methods often fail to account for variability caused by environmental and operational changes. The proposed approach addresses these limitations by enhancing sensitivity in high-density regions and providing flexibility in sparse data areas. It operates on the premise that test set points in low-density regions indicate higher model error —not due to abnormal measurements of operational variables, but because external variables, such as reservoir levels or temperatures, fall outside the range of the training set—.

The methodology integrates a Boosted Regression Tree (BRT) model for predictive analysis, identifying key factors such as reservoir levels and temperatures. KDE is employed to estimate data density, enabling dynamic threshold calibration. In low-density regions, where prediction uncertainty is higher, thresholds are widened, while high-density regions maintain stricter thresholds to ensure precision.

This framework was applied to monitoring data from a double-curvature arch dam, utilizing reservoir levels and temperatures over multi-scale moving averages. Results demonstrated a substantial reduction in false alarms while retaining robust anomaly detection. The adaptive thresholds improved reliability by accommodating deviations caused by external factors not captured in the predictive model, such as extreme environmental conditions.

By leveraging KDE, this approach offers an interpretable, scalable solution for modern dam safety monitoring, particularly under varying load conditions. It enhances the detection of structural anomalies without excessive false positives, making it a practical advancement in early warning systems. This study underscores the importance of adaptive methodologies in addressing the complex and evolving challenges of dam safety.