A modelling approach based on Bayesian networks for dam risk analysis: Integration of machine learning algorithm and domain knowledge

Abstract

The safety of dams, especially that of earthen dams, is threatened by various uncertain and interrelated risk factors. Consequently, dam risk analysis is vital for dam safety governance and failure prevention. A Bayesian network (BN) is an effective tool for this issue as its excellent ability in representing uncertainty and reasoning. Most previous studies have relied solely on domain knowledge (DK) to establish BN models, leading to inefficient and subjective results when solving complex systems. The increasing observations has improved the viability of using machine learning (ML) to automatically model complex systems. Herein, ML algorithms are used to develop automatic BN models for risk analysis of earthen dams in the USA, which are subsequently modified using DK. The results revealed that the automatic BN models can identify some potential causal relationships that are ignored by DK, whereas some impractical causalities identified in the automatic BN models can be modified by using DK. Moreover, the modified BN model has a better performance in the prediction of earthen dam failure with an average overall accuracy of 84.6%, compared to 80.3% with the automatic BN models, and 76.5% with a manual BN model created using only DK. Using the modified BN models, the three foremost risk factors based on their influence and sensitivity analysis were identified to be extreme flood, malfunction of spillway or gate, and slope instability. Our study highlights that the integration of ML algorithms and DK is an effective approach for developing reliable BN models for dam risk analysis.