Study On The Risk Analysis Model Of Seismic Liquefaction Based On Bayesian Network

Further researching the issues about seismic liquefaction is very significant to earthquake resistance and hazards reduction. The existing research results almost are bout prediction of seismic liquefaction, and the predictive methods had so great limitations in practical application that their predictive accuracy is not very high. In addition, there are few researches about integrating prediction of hazards induced by seismic liquefaction, especially researches about decision-making for mitigating hazards induced by seismic liquefaction. In this paper, several significant factors of seismic liquefaction were chosen using the statistic computation method to develop a Bayesian network (BN) model for predicting seismic liquefaction base on Bayesian network method. After that, adding variables of hazards induced by seismic liquefaction, anti-liquefaction measures and losses into the previous BN model, to separately develop a BN model for predicting hazards induced by seismic liquefaction and a BN model for making decisions of mitigating hazards. In these two new BN models, they not only can rapidly and accurately predict seismic liquefaction and liquefaction-induced hazards, but also can provide the best decision-making scheme for different site and its hazards, the results can provide a scientific basis for earthquake resistance and hazard mitigation. The major research issues of this paper are as follows:At first, twelve significant factors were identified from various factors of the seismic liquefaction base on the statistic computation method and selection principles of important factors, which are epicentral distance, magnitude of the earthquake, peak ground acceleration, duration of the earthquake, soil type, fines or clay content, particle composition, relative density, thickness of liquefiable soil layer, depth of liquefiable soil layer, covered effective stress and groundwater table. Their hierarchy structural network was developed using interpretive structural modeling method, and the hierarchical relationships were analyzed subsequently. The result can provide guidance for further developing Bayesian network model of seismic liquefaction.Secondly, magnitude of the earthquake, epicentral distance, SPT (standard penetration test) blow count, groundwater table, and depth of sand layer were selected to construct a subjective Bayesian network model, an objective Bayesian network model and a hybrid Bayesian network model for predicting seismic sand liquefaction in the free field base on interpretive structural modeling method and K2 algorithm, and the three models were compared with other common methods of predicting seismic liquefaction using five performances assessment indexes, overall accuracy (OA), prediction, recall, F1 score and area under the ROC (receiver operating characteristic) curve (AUC), to prove their effectiveness and accuracy, among the three Bayesian network models, the performances of the hybrid Bayesian network model are the best. After that the effects of class imbalance and sampling bias of training data on performances of probability models for seismic liquefaction were discussed, the results showed that the worse class imbalance of training data is. or the worse sampling bias is, the better the performances of the models are in return classification, whereas the worse the predictive performances of the models are, and provided the best ranges of class distribution ratio for common probabilistic methods to reduce their predictive errors. In addition, the application of oversampling technology in improving predictive performances of the models under condition of worse class imbalance or sampling bias.Thirdly, two ways were used to separately develop Bayesian network model of seismic liquefaction based on the twelve significant factors. One is using interpretive structural modeling method and the causal mapping method to construct a subjective Bayesian network model under the condition of incomplete data, another way is combining interpretive structural modeling method and K2 algorithm to construct a hybrid Bayesian network model under the condition of complete data. After that, based on K-fold (K=5) test method, the five performance assessment indexes were used to comprehensively assess the performances of the two Bayesian network models, and prove the effectiveness and robustness of the two models according to compare with other probability methods, and analyze the sensitivities of factors of seismic liquefaction. The new Bayesian network models expanded and improved scope and predictive accuracy of the previous Bayesian network models, so that they can be applied on accurately predicting seismic liquefaction of different soils containing different fines content in the free field and the field containing upper structure.Finally, after adding some indexes about hazards induced by seismic liquefaction, such as liquefaction potential index, sand boils, cracks of the ground, settlement and lateral spreading, a new Bayesian network model for integrally predicting severity of liquefaction-induced hazards were constructed based on the previous hybrid Bayesian network model. Its effectiveness and robustness were proved by comparing with artificial neural network using the five performance assessment indexes. Subsequently, the new Bayesian network model was expanded by adding decision nodes and utility nodes, such as anti-liquefaction measures, losses and costs, to develop a Bayesian decision network, which can not only predict whether seismic liquefaction occur and how much hazards were induced by liquefaction, but also can make the best decision for mitigating the hazards induced by seismic liquefaction. The Bayesian decision network was used in anti-liquefaction of artificial islands, and effectiveness of the model was proved by comparing with the results of numerical simulation, so that the model can provide a scientific basis for preventing and mitigating disasters of seismic liquefaction.