Research On Risk Analysis Method Of Earth Dam Failure Based On Bayesian Network Integrated With Machine Learning

The risk analysis of earth dam failure is a major concern in the field of dam safety management.A Bayesian network(BN)is an effective tool for dam risk analysis.The risk analysis of earth dam failure using BNs involves three important aspects: the variable selection of risk networks,the modeling approach of risk networks,and the reasoning of dam failure risk.Most previous studies have relied on expert opinions and frequency statistics,which leads to three critical limitations including unclear variable selection mechanism,inadequate expression of causal relationships,and insufficient explainability of dam failure risk reasoning.To overcome these shortcomings,machine learning technologies were integrated into BNs to conduct the risk analysis of earth dam failure based on the American earth dam incident dataset.The paper aims at improving the rationality of variable selection,establishing a more objective and thorough risk network model,and enhancing the interpretability of dam failure risk reasoning,which can thus provide a theoretical foundation for risk prevention and management and technical supports for safe operation of earth dams.The main research contents of this paper consists of the following three aspects:(1)Selecting critical risk factors is a prerequisite to ensure the quality of the risk network model.The traditional variable selection methods are mainly based on domain knowledge and frequency statistics,which lack quantitative basis of measuring the importance of risk factors and might lead to inaccuracy and unreasonable results in the process of variable selection.Based on the American earth dam incident dataset and the advantages of boosted regression trees(i.e.,BRT,a machine learning model)in complex system modeling,an optimized variable selection using BRT was proposed for earth dam risk network.In the process of optimized variable selection,the binary cross-entropy loss function was selected to measure the prediction deviance of the BRT model,the contribution of risk factors to the prediction deviance was used as the variable ranking standard,and the noise threshold criterion was served as the discrimination method to eliminate redundant risk factors,,which can thus provide a quantitative basis for determining the number and type of critical risk factors.Using the optimized variable selection,12 critical risk factors of earth dam failure in the United States,such as insufficient spillway capacity,extreme flood and slope instability,were selected form 25 candidate risk factors,which provided variables with strong predictive power for the eastablishment of risk network model.The results showed that the optimal variable selection using BRT effectively fitted the characteristics of the earth dam system,which can help to reduce the adverse effects of noise in data,reduces the complexity of risk network model and provides a new approach for selecting key variables of risk network model in risk analysis of earth dam failure.(2)Establishing a risk network model is a crucial basis for risk analysis of earth dam failure.For this issue,traditional methods mainly rely on domain knowledge to manually construct the network structure and determine the probability parameters.The established risk networks lack expression of causality and further reduce the reliability of risk analysis.Based on the optimized variable selection,a modeling approach of risk network integrated with data mining and domain knowledge was proposed.In the modeling process,the structure learning algorithms were used to mine the causality of the earth dam system,and the domain knowledge was used to optimize the network structure by producing causal structural constraints.The network parameters were determined by the parameter learning algorithm,and finally the risk network model of earth dam failure was established.The model comparision shows that the model established by the proposed approach can identify the potential causality ignored by domain knowledge,and has an appropriate complexity through the correction effect of domain knowledge,which highlights the main causality of the earth dam system.The evaluation of all indexes(e.g.,OA,AUC and F1-Score)showed that the model integrated with data mining and domain knowledge has significantly better prediction compared to the models constructed by a single method.Moreover,the influence strength of 12 critical risk factors,three failure modes and the safety of earth dam system were calculated,and the main earth dam failure path(i.e.,insufficient spillway capacity→overtopping→earth dam failure)was revealed.These results show that the integration of data mining and domain knowledge can effectively improve the predictive performance and causal expression ability of the risk network model,which provides an effective way for the modeling of risk network.(3)Reasoning based on the risk network model is a significant task of risk analysis of earth dam failure.Previous studies on risk reasoning of dam failure mainly involves the prediction of dam failure risk and the diagnosis of risk factors,which lack the interpretability and the whole process analysis of reasoning.Based on the established risk network model,the counterfactual framework in the field of artificial was introduced into the reasoning process of risk network model,and the counterfactual reasoning framework of dam failure risk was constructed using the structural causal model and the twin network.The framework was applied to an earth dam project,and a structural causal model of the earth dam was established by assigning the probabilities of hidden variables.Considering the operation conditions of the earth dam,four counterfactual hypotheses were proposed and corresponding counterfactual twin networks were constructed.On this basis,the reduction rate of the failure risk of the earth dam under the four risk treatment measures(i.e.,improving slope stability,repairing spillway,solving the problem of both slope and spillway,and preventing piping)was calculated using the estimated posterior importance sampling(EPIS)reasoning algorithm,which were 57.6%,29.2%,74.2% and 12.1%,respectively.The results showed that the counterfactual,as a higher-level reasoning framework,not only help to explain the causal relationships of dam systems,but also analyzes the whole process of risk reasoning of dam failure.Meanwhile,the counterfactual can effectively fit the actual operation conditions of earth dams and compare the effects of different decision-making schemes on earth dam safety.The counterfactual reasoning framework of dam failure risk enriched the reasoning system and improved the interpretability and applicability of risk reasoning of dam failure,which provided a more effective theory support for the risk control and management of earth dams.