杜星, 孙永福, 宋玉鹏, 等.2020.基于MPL神经网络的地震作用下砂土液化评估及预测[J].工程地质学报, 28(6):1458-1465.doi: 10.13544/j.cnki.jeg.2019-321.
    引用本文: 杜星, 孙永福, 宋玉鹏, 等.2020.基于MPL神经网络的地震作用下砂土液化评估及预测[J].工程地质学报, 28(6):1458-1465.doi: 10.13544/j.cnki.jeg.2019-321.
    Du Xing, Sun Yongfu, Song Yupeng, et al.2020.Multilayer perception neural network for assessment and prediction of earthquake-induced sand liquefaction[J].Journal of Engineering Geology, 28(6):1458-1465.doi: 10.13544/j.cnki.jeg.2019-321.
    Citation: Du Xing, Sun Yongfu, Song Yupeng, et al.2020.Multilayer perception neural network for assessment and prediction of earthquake-induced sand liquefaction[J].Journal of Engineering Geology, 28(6):1458-1465.doi: 10.13544/j.cnki.jeg.2019-321.

    基于MPL神经网络的地震作用下砂土液化评估及预测

    MULTILAYER PERCEPTION NEURAL NETWORK FOR ASSESSMENT AND PREDICTION OF EARTHQUAKE-INDUCED SAND LIQUEFACTION

    • 摘要: 砂土在地震的作用下会产生剧烈的液化现象,液化引发的地基失稳会对道路、建筑物、堤坝等各类设施造成严重危害。因此,地震作用下的砂土液化判别预测一直是地质灾害领域研究的热点问题。本文使用过去几十年发生在世界各地的166组地震作用下砂土液化实例数据,通过大量数据训练和参数分析建立了基于机器学习的地震作用下砂土液化判别模型。结果表明,当网络结构为6(输入层)-15(隐藏层)-1(输出层)、训练函数为Levenberg-Marquardt时,对地震液化预测效果较好,最大准确率可达96%。参数分析结果表明不同参数对网络预测准确率影响程度不一:锥端阻力、地表归一化峰值水平加速度影响相对较大;地震震级、总垂向应力、有效垂向应力影响中等;贯入深度对其影响较小。因此在不同网络预测精度要求的条件下,可考虑适当简化输入参数。

       

      Abstract: Sand can be severely liquefied under the action of earthquakes. The instability of foundation caused by sand liquefaction can cause serious damage to various facilities such as roads, buildings and dams. Therefore, the prediction of sand liquefaction under earthquake action has always been a hot topic in the field of geological disasters. This paper uses 166 sets of identification data of earthquake induced sand liquefaction cases that occurred in the world in the past few decades. Through numerous data training and parameter analysis, we establish the sand liquefaction assessment model based on machine learning. The results show that the network structure is optimized when the network structure is 6(input layer)-15(hidden layer)-1(output layer) and the training function is Levenberg-Marquardt. The best prediction accuracy of test data can reach 96%. The results of parameter analysis show that different parameters have different effects on the accuracy of network prediction. Cone end resistance, the surface normalized peak horizontal acceleration has a relatively large impact. Earthquake magnitude, total vertical stress, effective vertical stress have moderate influence; penetration depth has less effect. Therefore, the input parameters can be simplified appropriately under the conditions of different network prediction accuracy requirements.

       

    /

    返回文章
    返回