节点文献

基于神经网络的滑坡预测及其控制研究

Research on Landslide Prediction and Controlling Based on Artificial Neural Networks

【作者】 陈煌琼

【导师】 曾志刚;

【作者基本信息】 华中科技大学 , 控制科学与工程, 2013, 博士

【摘要】 随着社会经济的飞速发展,人类城市化进程不断加快,植被绿化面积逐年减少,尤其是由于工业化的需要在滑坡灾害高发区所进行的切坡、地下采矿等活动,在很大程度上提高了稳定状态或准稳定状态的斜坡岩土体向不稳定状态转变的可能,使得斜坡岩土体沿着贯通的剪切破坏面发生滑移,从而产生滑坡这种地质灾害。滑坡具有自然、社会和经济等多种属性,威胁人类生命财产安全的同时,阻碍国家和地区经济的健康发展。为了积极做好滑坡灾害应急等方面的工作,尽可能减少滑坡造成的损失,对滑坡灾害的实时准确预测及控制是必要的。本文分析了滑坡研究的发展现状和技术趋势,在对滑坡稳定性做出评价的基础上,进行滑坡累积位移预测以及滑坡的控制。滑坡稳定性分析是滑坡预测与控制研究的首要问题之一。针对传统稳定性评价方法中参数选择过度简化所导致的稳定系数计算过程过于主观、结果可信度不高的局限性,考虑到实际项目中的小样本、模型非线性及高维的特征,本文提出了一种基于支持向量机(SVM)的滑坡稳定性评价方法。以影响滑坡稳定的六个主要因素作为SVM的判别因子,选择径向基函数作为核函数,采用交叉验证的方法寻找最优的惩罚系数和方差。与传统方法相比,该方法提高了滑坡稳定性判别的精度和速度。不稳定状态的斜坡岩土体需要预测其即将发生的形变。考虑到滑坡的力学行为和变形趋势呈现出确定性与随机性共存的非线性特征,为了避免建立复杂的力学系统模型,本文针对滑坡预测预报这一复杂性问题提出一种基于前向反馈传播(BP)神经网络的预测模型,描述了滑坡系统输入输出之间的非线性关系。为克服BP网络后期收敛速度慢、易陷入局部极小点等缺点,以遗传算法(GA)和模拟退火算法(SA)相结合的方式即遗传-模拟退火(GSA)算法对网络权值进行优化。用该方法预测滑坡位移获得了较好的效果。进一步考虑降雨量这一外界因素对滑坡位移产生的影响。据“中国地质灾害数据库”中的数据显示,大部分滑坡的发生与降雨有密切的关系。目前关于降雨量与滑坡位移两者间关系的研究多数基于静态神经网络或其他静态模型,难以反映这种关系的动态特性。对滑坡位移的预测可以看作是对一个非线性动态系统的辨识过程。基于此,本文采用Elman递归神经网络建立系统预测模型,并用GA优化网络的初始权值。该方法具有较高的收敛速度和准确度,并且具有较强的适应性和灵活性。论文最后引入控制论的观点,探讨滑坡的控制策略。由于滑坡系统的复杂性,很难获得精确的数学模型,采用预测控制方法对滑坡系统进行控制则避开了这一难点。本文根据滑坡动力学原理选择滑坡推力作为系统控制量,利用GA-Elman神经网络对滑坡系统进行辨识,经过滚动优化和反馈校正,得到控制滑坡发生的最优控制量时间序列。本文系统地探讨了滑坡系统中的三个主要问题:稳定性分析、位移预测及位移控制,对工程实践具有一定的理论指导意义。

【Abstract】 With rapid economic development and urbanization process, green area has been reducedyear after year. Especially some human activities such as slope cut and underground mining inhigh incidence area of slope geological disasters make slope rock mass less stable to a largeextent, so the possibility of slippage of slope rock mass along connecting shear failure surfacewill be larger. Finally the failure of slope will happen. Besides natural attribute, Slope also hassocial and economic attributes. It directly threatens people’s safety of life and wealth, andhurts development of national and regional economy. In order to implement disasteremergency management and reduce the loss caused by landslides, prediction and controlshould be in real time and accurate. Based on summarizing overseas and domestic researchstatus and development tendency of landslide prevention and control, evaluation of slopestability is discussed, then prediction and control methods of slope cumulative displacementare proposed in this dissertation.The first step is evaluation of slope stability. As over-simplification of parametersselection in traditional stability evaluation methods leads to subjective computational processof stability coefficient and reduces credibility of results, the evaluation method based onsupport vector machine (SVM for short) is proposed, considering the characteristics of smallsample, nonlinearity model and high-dimensional data in actual projects.6major factorsinfluencing slope stability are considered as discriminant factors of SVM. Radial basisfunction is selected as kernel function. Then cross validation method is used to determineoptimal penalty coefficient and variance. Compared with conventional methods, SVMapproach increases accuracy and speed of evaluation.It’s necessary to predict imminent deformation of unstable slope rock masses.Considering nonlinearity characteristic (determinacy and stochasticity coexist) presented inmechanical behavior and deformation tendency of slope, the prediction model based on BPneural network is proposed to describe non-linear relationship between input and output ofslope system. Modeling complicated mechanical system can be avoided by use of BP neuralnetwork. Genetic algorithm (GA) and simulated annealing (SA) algorithmare combined,which is genetic-simulated annealing (GSA) algorithm, to optimize network weights, so thatthe shortcomings of BP neural network such as slow rate of convergence and local minimumtrouble in later stage can be overcome. As an external factor, rainfall capacity has an effect on slope displacement. According todata in Chinese geological disaster database, the occurrences of most landslides are caused byrainfall. Currently the research on relationship between rainfall capacity and slopedisplacement is mainly based on static neural network or other static models. These staticmodels have trouble in reflecting dynamic characteristics of this relationship. The predictionof slope displacement can be treated as identification process of a nonlinear dynamic system.Elman recurrent neural network is used to build systematic prediction model, and initialnetwork weight values are optimized by genetic algorithm. This method is adaptive andflexible, it has high rate of convergence and accurate result.Finally the opinions of cybernetics are introduced to discuss control strategy of slope.Due to the complexity of slope system, it’s hard to obtain accurate mathematical model. Theusage of predictive control methods for slope system can avoid this trouble. Landslide thrustforce is selected as systems control variable according to principle of landslide dynamics, thenGA-Elman neural network is used to identify slope system. By rolling optimization andfeedback compensation, we can find time series of optimal control variables.This dissertation systematically discusses3main problems of slope system which arestability analysis, displacement prediction and displacement control. It provides theoreticaldirection for further slope system research.

  • 【分类号】P642.22;TP183
  • 【被引频次】2
  • 【下载频次】359
  • 攻读期成果
节点文献中: 

本文链接的文献网络图示:

本文的引文网络