【作者】 任文杰；

【导师】 窦远明； 苏经宇；

【作者基本信息】 河北工业大学 ， 结构工程， 2002， 硕士

【摘要】 地基土液化是地震工程中的重要问题，其发生、发展是一个复杂过程。液化预测包括可能性和危害程度预测两个方面。液化的影响因素很多，随机性大，且各因素呈高度的非线性。传统的判别和等级评价方法多是在宏观震害现象和室内试验基础上总结、分析、统计得到的，有一定的实用性，但也存在着一些局限性，结论可靠度尚需提高。本文在评析传统方法基础上，提出建立一种人为影响因素小、能容定量与定性指标于一体的液化判别及评价模型是非常必要的。 本文阐述了人工神经网络和遗传算法的基本原理及实现技术，并在此基础上利用Matlab5.3编写了人工神经网络程序： (1)编写了BP人工神经网络程序，采用Vogl“批处理”快速算法，学习速率、动量参数在误差修正过程中自适应调节，提高了训练的速度。 (2)编写了遗传神经网络GA-BP程序，采用二级遗传算法与BP算法相结合，同时优化网络隐层节点数和权值、阈值，既克服了寻优过程的盲目性，又避免陷入局部极小，提高了网络的学习精度。 (3)与工具箱的L-M算法结合编写了改进遗传神经网络GA-LM程序，综合考虑了部分预测样本对适应度函数的影响，提高了网络的学习、泛化能力及运行速度。 (4)与工具箱的RBF网络相比较，说明网络的学习算法是改善网络性能的关键。 本文依据地震液化及其危害程度实测资料，综合考虑多方面因素，建立了液化判别及等级评价的人工神经网络模型。并与实测结果及传统方法相比较，得出以下结论： (1)在数据合理情况下，神经网络方法可以快速达到比传统方法更高的预测精度。说明建立的液化判别及等级评价的网络模型是科学的、有效的。 (2)网络模型可以揭示结构和参数与运行之间的内在关系，将输入、输出关系量化公式。根据单个因素贡献率进行主成份分析，不仅验证了传统方法的合理性，且对规范方法提出了建议。更多还原

【Abstract】 Foundation soil liquefaction is an important problem in earthquake engineering, which comes through a complicate process. Predictions of liquefaction include prediction of probability and damage quantity. Liquefaction involves in a great deal of influencing factors, which have strong randomness and reciprocal nonlinearity. Based on macroscopical earthquake calamities and laboratory tests, traditional methods about estimation and grade evaluation of liquefaction are inducted by means of generalization, analyses and statistics, which have some practicability and some limitation. This thesis analyses and assesses traditional methods, and brings forward the necessity of establishing models of estimation and grade evaluation of liquefaction, which concern minor factitious influencing factors and embrace aquantitative and qualitative indexes.This thesis expounds fundamental principle and realization technique of Artificial Neural Network and Genetic Algorithm, and redacts Artificial Neural Network procedures.-(l)Adopting batch processing high-speed algorithm, the thesis redacts Back-Propagation Network procedure to enchance training velocity, in which learning rate and momentum parameters are modulated self-adaptably during error correction.(2)Combining secondary Genetic Algorithm with Back-Propagation Network, the thesis redacts Genetic Neural Network procedure, which optimizes number of hidden node and weight value and threshold value simultaneously. The procedure overcomes blindness during search, avoids falling into localminimum and increases learning accuracy.(3)Combining secondary Genetic Algorithm with L-M algorithm in toolbox and thinking about effects of partial forecasting sample in adaptive value, the thesis redacts improved Genetic Neural Network procedure in order to pick up faculty of learning and generalization and running velocity.(4)Compared with RBF Network, it is concluded that learning algorithm is master key to improve properties of Artificial Neural Network.On the basis of measured data of earthquake liquefaction and damage quantity, this thesis set forth Artificial Neural Network models of estimation and grade evaluation of liquefaction, which take versatile factors into account. Compared with measured data and traditional means, it is concluded:(l)Artificial Neural Network can forecast more accurately and quickly than traditional means with reasonable data. The results show that the Artificial Neural Network models of estimation and grade evaluation of liquefaction is scientific and effective.(2)The Artificial Neural Network models can reveal internal relations between structural parameters and operation, and formularize the maping of input-output information. By means of computing relative contribution rate of single factor, the thesis not only test and verify rationality of conventional means, but also put forth proposition to norm.更多还原