矿山边坡变形监测数据的高斯过程智能分析与预测

发布时间：2018-03-25 22:29

本文选题：矿山边坡　切入点：变形监测　出处：《太原理工大学》2016年博士论文

【摘要】：矿产资源开采引发的地表塌陷、崩塌、地裂缝和地面沉降等矿山灾害给人类的生命财产安全造成严重的威胁,集成多传感器的自动化、智能化监测系统是矿山地面灾害监测的发展方向。以中煤平朔井工二矿边坡(简称为二号井边坡)自动化监测系统作为研究案例,应用高斯过程(Gaussian Process,简称GP)理论研究变形数据智能分析方法和预测模型,据之对矿山地面灾害进行防治提供科学依据。将这种研究成果结合自动化、智能化监测技术应用于矿山地面灾害监测具有广阔的应用前景。监测数据的可靠性是变形监测分析和预测的基础,针对原始观测数据可能存在的异常值,提出的完整搜索估计法(Full Search Estimation,简称FSE)能够实现多维异常数据定位、估值和修正:根据异常数据影响验后方差这一基本思想,设计了多维异常数据定位搜索算法,在算法执行的过程中能够自动生成包含异常数据位置的定位矩阵,同时给出了动态阈值计算公式用于判断搜索是否结束;应用可靠性理论结合最小二乘方法推证了异常数据的估值和修正方程。分别以测量机器人异常数据探测和矿山坐标转换参数可靠性求解为例对FSE进行实证分析,结果表明FSE具有较好的抗差能力。边坡变监测过程中受外界环境及施工作业等因素的影响有时造成数据缺失,需要应用时空插值方法对缺失数据加以插补形成完整的时空序列数据。通过研究高斯过程回归(Gaussian Process Regression,简称GPR)在时间域上插值的样本数量,给出GPR在时间域上的一维插值方法和步骤,实验证明GPR在时间域上可以适应线性和非线性插值;按照空间插值样本数据选择的一般原则,进一步研究了基于gpr的空间插值方法;顾及监测数据的时空关联性,利用gpr在时间域和空间域插值输出的验后方差作为定权因子,给出了基于gpr的时空插值的计算公式,并用交叉验证法证明了gpr时空插值的可行性。对变形区域进行时空位移特征分析是变形数据分析的一项主要内容,就描述监测点三维位移特征常用的绝对指标进行了论述,但仅使用绝对指标凸显不出监测点相对稳定状态。将短期位移速率和累积位移速率的比值定义为累积位移速率比作为一种相对指标,使用累积位移速率比的大小和符号可以简单直观的分析监测点的相对稳定状态。通过计算分析一段时间内的累积位移速率比,依据3σ准则将监测点的稳定状态分为四个级别,即稳定、较稳定、不稳定和极不稳定。单独分析监测点的变形特征难以从整体上掌握变形区域的时空演化趋势和变形规律。为此,在gpr时空插值的基础上研究了基于gpr变形趋势面模型建模方法和流程,以三维累积位移量作为分析对象,构建了二号井边坡的变形趋势面模型,以此来分析其变形的时空演化过程;应用fse提取累积位移速率比的离群值,并将提取结果赋予高斯过程分类(gaussianprocessclassification,简称gpc)标志,进而给出基于gpc变形区域局部稳定性分析方法和流程,以累积位移速率比作为分析对象,对二号井边坡的局部稳定性进行整体分析。监测点在发生变形的过程中经常表现出明显的非线性特征,利用gpr超参数自适应求解、输出结果具有概率意义的优点研究了变形智能预测模型。鉴于gpr的核函数对预测性能有很大影响,应用核函数相加方式得到与变形曲线特点相吻合的组合式核函数“matern32+se”;考虑到监测数据的不断更新和累积,为保持超参数与训练样本集的一致性,研究了“递进~截尾式”超参数动态更新模式和gpr最佳训练样本数量确定方法;在此基础上建立了以时间作为输入项的GPR监测点时间驱动智能预测模型(GPR-TIPM)和以历史数据作为输入项的GPR监测点数据驱动智能预测模型(GPR-DIPM)。分别将两种模型应用于二号井边坡进行中短期变形预测,实验结果表明两种预测模型均取得了较为理想的效果,GPR-TIPM的预测性能总体上优于GPR-DIPM。通过GPR-TIPM模型与经典的AR(p)和GM(1,1)模型的实验对比分析,结果表明GPR-TIPM的预测精度明显提高。最后部分设计了GP变形监测数据处理软件原型系统架构,并以此架构为导向,根据文中提到的模型和算法用Matlab和C#语言分别实现了服务端近实时数据处理系统和GIS客户端可视化在线分析系统。
[Abstract]:The surface subsidence caused by exploitation of mineral resources, collapse, ground fissure and ground subsidence of mine disaster caused a serious threat to human life and property safety, multi sensor integrated automation, intelligent monitoring system is the development direction of mining ground disaster monitoring. The slope coal underground mine two Ping Shuo (referred to as No. two well slope the automatic monitoring system) as a case study, the application of Gauss (Gaussian Process, referred to as GP) intelligent data analysis method and prediction model of deformation theory, according to the control of mining ground disaster to provide a scientific basis. The research achievements of the combination of automation, intelligent monitoring technology application in mining ground disaster monitoring has broad application prospects. The reliability of the monitoring data is the basis of the analysis and forecasting of deformation, the abnormal value of raw data may exist, proposed a complete search. Meter method (Full Search Estimation, referred to as FSE) can realize multidimensional anomaly positioning, estimation and correction: according to the basic idea of abnormal data influence posterior variance, design a multidimensional anomaly positioning search algorithm, can automatically generate the location matrix including abnormal data in the process of implementation of the location algorithm, and dynamic threshold formula to determine whether the end of the search is given; the application of the reliability theory combined with the least squares method to deduce abnormal data estimation and correction equation. By measuring robot abnormal data detection and mine coordinate conversion parameters reliability for the empirical analysis of FSE, the results show that FSE has good robust ability. The influence of outside environment and construction work the factors such as slope deformation monitoring in the process sometimes resulting in missing data, requires the application of spatio-temporal interpolation method for missing data to be inserted Fill space complete sequence data. Through the study of Gauss (Gaussian Process Regression, the regression process referred to as GPR) the number of sample interpolation in time domain, GPR is given in the time domain of the one-dimensional interpolation method and steps, experiments show that GPR can adapt to the linear and nonlinear interpolation in time domain; according to the general principles of spatial interpolation samples data selection, further study of the spatial interpolation method based on GPR; take into account the spatio-temporal correlation of monitoring data, the use of GPR in time domain and space domain interpolation output posterior variance as weighting factor, calculation formula of spatio-temporal interpolation based on GPR is proposed, and the feasibility of GPR temporal interpolation is proved by cross validation method analysis of characteristics of time and space. The displacement deformation area is a major content of deformation data analysis, describe the absolute index of three-dimensional displacement monitoring points of common features were discussed , but only use the absolute index not highlight the relatively stable state monitoring points. The short-term displacement rate and cumulative displacement rate is defined as the ratio of cumulative displacement ratio as a relative index, use the sign and magnitude of the cumulative displacement rate ratio of the monitoring points can be analyzed a simple relatively stable state. Through the analysis of cumulative displacement rate the time ratio calculation, on the basis of the 3 Sigma standards monitoring points in the steady state is divided into four levels, namely, stable, stable, unstable and very unstable. A separate analysis of deformation characteristics of monitoring points to the overall evolution of regional spatial and temporal trends of the palm grip and deformation deformation. Therefore, based on GPR the spatio-temporal interpolation of GPR deformation trend surface modeling method and process based on three-dimensional cumulative displacement as the analysis object, construct the deformation trend of No. two well slope surface model, which is used to divide Analysis of the spatial and temporal evolution of the deformation; FSE is used to extract the outlier cumulative displacement rate ratio of the value, and the result of the extraction process to Gauss classification (gaussianprocessclassification, referred to as GPC), and then gives the method and process of analysis of local stability of regional GPC deformation based on the cumulative displacement speed ratio as the research object, the local stability of slope No. two well the overall analysis. Monitoring points during a deformation process often exhibit obvious nonlinear characteristics, super parameter adaptive solution by GPR, the output and the probabilistic prediction model of intelligent deformation. In view of the kernel function of GPR has great influence on the prediction performance, application of nuclear function additive combined nuclear way function "is consistent with the characteristics of deformation curve of matern32+se; taking into account the monitoring data update and accumulation, in order to maintain the hyper parameters and training The consistency of training sample set, on the "progressive censoring ~" super dynamic parameter update mode and GPR optimal method to determine the number of training samples; set up on the basis of the GPR monitoring point in time as the time of entry driven intelligent prediction model (GPR-TIPM) and GPR based on historical data as input data driven monitoring points intelligent prediction model (GPR-DIPM) respectively. The two models applied to No. two well in the short term the slope deformation prediction, the experimental results show that the two models have achieved satisfactory effect, the prediction of GPR-TIPM overall performance is better than that of GPR-DIPM. through the GPR-TIPM model and the classic AR (P) and GM (1,1) analysis experiment comparison of model, results show that the prediction accuracy of GPR-TIPM is obviously improved. Finally a prototype software system architecture of data processing of deformation monitoring of GP, and this architecture oriented, according to the mentioned in this paper The model and algorithm use Matlab and C# language to implement the server near real-time data processing system and the GIS client visual online analysis system.

【学位授予单位】：太原理工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TD325.4

【相似文献】