基于BP神经网络的饱和砂土液化判别方法
|
摘要
基于唐山地震中大量的砂土液化现场实测资料,选取描述地震动特性的烈度、震中距、地面峰值加速度和描述砂土层埋藏环境条件的地下水位、标贯点深度(土层深度)、上覆非液化覆盖土层厚度、有效覆盖压力,以及表示砂土自身属性的标准贯入锤击数、平均粒径、不均匀系数、修正标贯击数共11个指标的不同组合作为输入变量,采用快速BP算法和LM算法构造了饱和砂土液化判别的BP神经网络预测模型。通过所建网络模型的训练、验证和应用,结果表明:(1)所建14个BP神经网络模型都是有效的,液化判别的准确度与模型输入变量的不同组合有关;(2)增加网络模型的节点(考虑因素较多)并不一定能够提高BP神经网络模型的液化判别准确度,反而增加了BP神经网络模型的复杂性和学习时间;(3)两种算法的BP神经网络模型都有很高的液化判别准确度,LM算法的计算速率要比快速BP算法快得多,但在计算过程中需要更多的内存,建议采用LM算法;(4)采用所提BP神经网络模型的权值与阈值进行其它预测样本的液化判别时,判别结果可能偏于保守;(5)从影响砂土液化的主要因素、获取指标难易程度考虑,在与《建筑抗震设计规范》砂土液化判别公式考虑指标一致的情况下,建议采用BP神经网络模型M4或M5a,该模型简单、方便,且其预测准确度远高于《建筑抗震设计规范》
Based on a lot of sandy soil liquefaction data obtained during Tangshan earthquake in 1976, selecting different combinations of 11 indexes as input variations, the faster algorithm of BP and LM algorithm are adopted to construct the BP neural network model for liquefaction potential estimation of saturated sand. The 11 indexes are earthquake intensity, the epicenter distance, peak ground acceleration for describing the earthquake ground motion characteristics, and the depth of underground water, the depth of standard penetration test point(soil layer), the thickness of the covered non-liquefied soil layer, effectively overlaying pressure describing the embedding environment of soil layers, and standard penetration blow-count, mean diameter, nonuniformity coefficient, modified standard penetration blow-counts for denoting the sandy soil feature. Through the practice and verification of the model, it shows that: (1) constructed 14 types of BP neural network model are effective, and evaluating accuracy depends on different combinations of the model input variations; (2) increasing nodes of the neural network model not always improve the evaluating accuracy, instead, increases the neural network model's complexity and learning time; (3) two algorithms of the BP neural network model both have the higher accuracy for liquefaction potential estimation of sand and LM algorithm is faster than the BP algorithm, but needs more EMS memory in the process of calculation, and therefore, LM algorithm is recommended in this paper; (4) the liquefaction potential estimation for other samples under using the weights and thresholds of the BP neural network model presented in this text, the evaluated result maybe conservative; (5) considering the main factors to affect sand liquefaction potential, the difficulty of acquiring the indexes, and under the condition that the indexes selected are consistent with those in Code for seismic design of buildings in China,M4 or M5 a neural network model is recommended in this paper. These models are (very) easy in application, moreover, the accuracy of the models for liquefaction potential estimation of sand is higher than that in Code for seismic design of buildings in China. hmsoftheBPneur LMalgorithmisfa fore,LMalgorithm
Based on a lot of sandy soil liquefaction data obtained during Tangshan earthquake in 1976, selecting different combinations of 11 indexes as input variations, the faster algorithm of BP and LM algorithm are adopted to construct the BP neural network model for liquefaction potential estimation of saturated sand. The 11 indexes are earthquake intensity, the epicenter distance, peak ground acceleration for describing the earthquake ground motion characteristics, and the depth of underground water, the depth of standard penetration test point(soil layer), the thickness of the covered non-liquefied soil layer, effectively overlaying pressure describing the embedding environment of soil layers, and standard penetration blow-count, mean diameter, nonuniformity coefficient, modified standard penetration blow-counts for denoting the sandy soil feature. Through the practice and verification of the model, it shows that: (1) constructed 14 types of BP neural network model are effective, and evaluating accuracy depends on different combinations of the model input variations; (2) increasing nodes of the neural network model not always improve the evaluating accuracy, instead, increases the neural network model's complexity and learning time; (3) two algorithms of the BP neural network model both have the higher accuracy for liquefaction potential estimation of sand and LM algorithm is faster than the BP algorithm, but needs more EMS memory in the process of calculation, and therefore, LM algorithm is recommended in this paper; (4) the liquefaction potential estimation for other samples under using the weights and thresholds of the BP neural network model presented in this text, the evaluated result maybe conservative; (5) considering the main factors to affect sand liquefaction potential, the difficulty of acquiring the indexes, and under the condition that the indexes selected are consistent with those in Code for seismic design of buildings in China,M4 or M5 a neural network model is recommended in this paper. These models are (very) easy in application, moreover, the accuracy of the models for liquefaction potential estimation of sand is higher than that in Code for seismic design of buildings in China. hmsoftheBPneur LMalgorithmisfa fore,LMalgorithm
引文
[1]陈文化,孙巨平,许兵.砂土地震液化的研究现状及发展趋势[J].世界地震工程,1999,15(3):16-24.
[2]陈国兴,胡庆兴,刘雪珠.关于砂土液化判别的若干意见[J].地震工程与工程振动,2002,22(1):141-151.
[3]陈国兴,张克绪,谢君斐.液化判别的可靠性研究[J].地震工程与工程振动,1991,11(2):85-96.
[4]YoudTI,IdrissIMetal.Liquefactionresistanceofsoils:summaryreportfromthe1996NCEERand1998NCEER/NFSworkshopsonevaluation ofliquefactionresistanceofsoils[J].JournalofGeotechnicalandGeoenvironmentalEngineering,ASCE,2001,127(8):817-833
[5]RahmanMS,WangJun.Fuzzyneuralnetworkmodelsforliquefactionprediction[J].SoilDynamicsandEarthquakeEngineering,2002,(5):685-694
[6]郭晶,杨章玉.MATLAB6.5辅助神经网络分析与设计[M].北京:电子工业出版社,2004.64-69.
[7]闻新.MATLAB神经网络应用设计[M].北京:科学出版社,2000.89-109.
[8]唐山地震砂土液化联合研究小组.唐山地震砂土液化现场勘察资料研究报告[R].北京:北京市勘察处,1983.2,26-41.
[9]虞和济.基于神经网络的只能诊断[M].北京:冶金工业出版社,2000.
[10]GarsonGD.Interpretingneuralnetworkconnectionweights[J].AIExpert,1991,6(7):47-51.
[2]陈国兴,胡庆兴,刘雪珠.关于砂土液化判别的若干意见[J].地震工程与工程振动,2002,22(1):141-151.
[3]陈国兴,张克绪,谢君斐.液化判别的可靠性研究[J].地震工程与工程振动,1991,11(2):85-96.
[4]YoudTI,IdrissIMetal.Liquefactionresistanceofsoils:summaryreportfromthe1996NCEERand1998NCEER/NFSworkshopsonevaluation ofliquefactionresistanceofsoils[J].JournalofGeotechnicalandGeoenvironmentalEngineering,ASCE,2001,127(8):817-833
[5]RahmanMS,WangJun.Fuzzyneuralnetworkmodelsforliquefactionprediction[J].SoilDynamicsandEarthquakeEngineering,2002,(5):685-694
[6]郭晶,杨章玉.MATLAB6.5辅助神经网络分析与设计[M].北京:电子工业出版社,2004.64-69.
[7]闻新.MATLAB神经网络应用设计[M].北京:科学出版社,2000.89-109.
[8]唐山地震砂土液化联合研究小组.唐山地震砂土液化现场勘察资料研究报告[R].北京:北京市勘察处,1983.2,26-41.
[9]虞和济.基于神经网络的只能诊断[M].北京:冶金工业出版社,2000.
[10]GarsonGD.Interpretingneuralnetworkconnectionweights[J].AIExpert,1991,6(7):47-51.
版权所有：© 2023 中国地质图书馆 中国地质调查局地学文献中心