论文部分内容阅读
基于解释结构模型和因果图法,选取12个具有代表性的定性和定量因素,在大量数据不完备的情况下提出了建立贝叶斯网络液化模型的方法。以2011年日本东北地区太平洋近海地震液化不完备数据为例,采用总体精度、ROC曲线下面积、准确率、召回率和F_1值5项指标对模型进行综合评估,并与径向基神经网络模型进行对比。结果表明:贝叶斯网络液化模型的回判和预测效果都优于径向基神经网络模型,且对于数据缺失的样本的预测效果也较理想。此外,该模型对于不同土质的液化评估均有较好的适用性。分类不均衡和抽样偏差会对模型的学习和预测效果产生很大影响,建议应同时采用上述5项评估指标进行综合评估模型的优劣。
Based on the explanatory structural model and the causality diagram method, 12 representative qualitative and quantitative factors are selected, and a method for establishing a Bayesian network liquefaction model is proposed when a large amount of data is incomplete. Taking the incomplete data of liquefaction in the coastal area of the Pacific Ocean in northeast Japan in 2011 as an example, the overall accuracy, the area under the ROC curve, the accuracy, the recall rate and the F_1 value were used to evaluate the model comprehensively. The results were compared with the RBF neural network model comparing. The results show that the Bayesian network liquefaction model is better than the RBF neural network model in predicting and predicting the liquefaction model, and the forecasting result of the data missing sample is also ideal. In addition, the model has good applicability to liquefaction evaluation of different soil types. The unbalanced classification and sampling deviations will have a great influence on the model learning and prediction effects. It is suggested that the above five evaluation indexes should be used together to evaluate the advantages and disadvantages of the comprehensive evaluation model.