Early detection of dam instability risk using a CNN–BiGRU model

Abstract

Reliable operation of hydraulic structures is a critical requirement for ensuring infrastructure safety, particularly under increasingly variable environmental conditions. However, early signs of dam instability are often subtle and difficult to detect using conventional monitoring approaches. This paper presents a data-driven framework for early detection of dam instability based on multivariate hydrometeorological time-series data. The proposed model combines convolutional neural networks (CNN) with a bidirectional gated recurrent unit (BiGRU) to jointly capture local temporal patterns and long-range dependencies in system dynamics. The framework is validated using real monitoring data, including reservoir water levels and meteorological variables. To address the imbalance between normal and high-risk states, a cost-sensitive training strategy is employed. The proposed model achieves an accuracy of 0.96, precision of 0.90, recall of 0.75, F1-score of 0.82, and ROC-AUC of 0.91. The results demonstrate that the model is capable of identifying early deviations from normal operating conditions while maintaining a low false alarm rate. The proposed approach can be integrated into real-time monitoring systems and contributes to improving the reliability and safety of hydraulic infrastructure.