应用灰色关系集群和CGNN分析矿井深部巷道围岩的稳定性控制 毕业论文外文翻译

应用灰色关系集群和CGNN分析矿井深部巷道围岩的稳定性控制 毕业论文外文翻译 英文原文 Application of Grey Relational Clustering and CGNN in Analyzing Stability Control of Surrounding Rocks in Deep Entry of Coal Mine 12Wanbin YANG , Zhiming QU (1.Beijing University of Science and Technology, Beijing, 100083; 2. Hebei University of Engineering, Handan, 056038) Abstract—With combination of grey neural network (CGNN) and grey relational clustering, the models are constructed, which are used to solve the prediction and coMParison of surrounding rocks stability controlling parameters in deep entry of coal mine.The results show that grey relational clustering is an effective way and CGNN has perfect ability to be studied in a short-term prediction. Combined grey neural network has the features of trend and fluctuation while combining with the time-dependent sequence prediction. It is concluded that great improvements coMPared with any methods of trend prediction and simple factor in combined grey neural network is stated and described in stably controlling the surrounding rocks in deep entry. I. INTRODUCTION GREY system technology states the uncertainty of small sample and poor information. With the development and generation of the unknown information, the real world will be discovered and the system operation behavior will be mastered properly. Through original stability with the pre-processing, the grey system law will be described. Though the real world is expressed complicatedly and the satisfied irregularly, the integrated functions will be appeared as a certain inner regular pattern [1]. The studying of grey system technology is based on the poor information which is generated by parts of the known information to extract valuable stability and to properly recognize and effectively control the system behavior. The neural network is dependent on its inner relations to model, which is well self-organized and self-adapted. The neural network can conquer the difficulties of traditionally quantitative prediction and avoid the disturbance of man’s mind. The grey relational analysis is based on the similarity of geometric parameters curve to determine the relation degree.The closer the curve shape similarity is, the greater the corresponding sequence correlation is. The similarity is described with correlation coefficient and correlation degree which describes the effect on the results by various factors.The greater the correlation is, the greater the iMPact extent is.While analyzing a practical system, the data series with the behavioral characteristics are identified. Additionally, it is necessary to ascertain the effective factors influencing system [1, 2]behavior characteristics, namely, sub-factors . Though the objective system are expressed complicatedly,the development and change are still of logic laws and the different functions are coordinated and unified. Therefore,how to find its inner developing regularities from the dispersed stability seems to be important. In the light of the description above, it can be found that the combination among grey relational clustering will take great effect on stability control of surrounding rocks in deep entry in coal mine. The combined grey neural network model will be built in solving and analyzing this problem. II. GREY RELATIONAL CLUSTERING A. Grey Relational Clustering [3-6]As the general system of grey trend relation, D. J. CHEN, etal has done a lot of work. In order to apply it into the practice, the basic idea about his study is introduced. The similarity and approximation in the dynamic system behavior can be expressed using grey trend relation (GTR) With the aid of GTR, the implicit system operation laws maybe stated aptly.Generally, the general system theory is applied in the general GTR system, and combining with GTR, the systemized models of GTR analysis will be deduced. It is assumed that U is the referred factor set and W the coMPared factor set, uy R is the set of GTR in (B, W) The matrixR,(,) is called the GTR matrix for uyqip，k(0)(0)，， and , while ,,b(k),,(k,1,2,？？,n)(q,1,2,？？,p)w(k)(i,1,2,？？,h)qin1the set (B, W) is finite. Where , the trend relation and ,,,(k),qiqi,n1,2k,(k),the trend relation function. qi (0)(0)(0)., And Q is the general ,z(k),b(k),w(k),,,,,,0,1(k,2,3,？？,n)qiqi ,,Q,(B,W),RGTR system, . uy In order to illustrate and serve application in this paper,some definitions are introduced here. V is the evaluating space of system Q and H is the evaluating functions. Thus, the relation between Q and evaluating space are described as H:B，W,V.Therefore, the general GTR system model is defined as H:B，W,V,,Q,(B,W),R;.The general GTR system is generalized, which uw includes the problems using the GTR analysis. In order to solve different problems, the GTR should be not alike in the light of B, W and H. On the basis of GTR matrix, the GTR clustering method is to assemble the observed index or objects into many definable classifications. The clustering can be seen as the observed object set of the same classification. Actually, any observed objects have many characteristic indexes which are not accurately classified. Through GTR clustering, the factors of the same classification are collected and the complicated system will be simplified [3-6].Z, the factor set of GTR system, has h factors. Each one represents a sequence, ,,,,Z,z(k),k,1,2,？？,n;i,1,2,？？,m.is the specific relational mapping, the iji trend relation of on the referred factor.;,; z,,,(z,z)z,z,Zzi,j,Mjijijiji ,,.Composed of Z and , the GTR system ,,,,i,1,2？？,m;j,1,2,？？,mij is called self-relational system of GTR. is GTR matrix, H the evaluation rule and , V the evaluation space.As to, is the threshold of clustering ,,,,0,1Q,(Z,,)c analysis and the evaluation rule is defined as . and are the similar terms Z,,,Zjci of characteristics while ,. ,,,i,jc At the classification of threshold ,the system characteristic variable is the ,c trend relational clustering. The system output is the clustering, which is expressed ,,as.Where, is the set including a group of V,,V,,,V,,,,,1,2,？？,, ,,mcharacteristic variables, the same as above and. , B. Grey Relational Clustering Prediction (0)Assuming that is the GTR time-dependent sequence Z(k), ,,and is the known model set. Each model set, in the light of GTR ,,fj,1,2,？？,mj sequence, can be supported by a group of prediction data,. f(k)(j,1,2,？？,n)j(0)(0),,is the GTR ofand.The meaning of andis f(k),F,f(k),,Z(k)Z,Z(k),jji,i(0)the same with that and.Z and F is the prediction set, f(k)Z(k)(j,1,2,？？,n)j, ,where k =1,2,……,n,i =1,2,……,h, j=1,2,……,m . Q = ((Z,F), ) is the GTR prediction system if H : ,,,is the prediction and evaluation rule of GTR and V opt is the evaluation space of system prediction effect. The system is mapped as H:Z，F,V. III. CGNN Using GM (1, 1) to predict sequence is one of the most frequently applicable fields. Because the grey model is in the light of stability to acquire the regularities, some predictable errors maybe appeared and many differently independent models will be setup to many related sequences, which can not consider the relations among stability sequences sufficiently. Generally, the shortcoming can be made up through setting up the combining models such as A. Combined grey neural network (CGNN) Tprediction model. is the input sample, and y, the single X:(x,x,？？,x)12nTToutput,the implicit node output,the U,(u,u,？？,u)W,(w,w,？？,w)1212nn weight connecting with implicit and output nodes.The connect weight value is 1 between input and implicit nodes because the signal is transmitted to the implicit layer ,,u,RX,C,by the input node. The output of NO. i implicit node is.Where i is iii Rthe number of implicit nodes, i,1,2,？？,m.is the radial function which is XCexpressed by Gaussian kernel function. is the input sample. is the center of ithi,radial basis function of neuron. is the width parameter of radial basis function i,22,thiX,Cof neuron. is the Euclidean norm.The activation function of implicit node y,wexp,0.5X,C,iiii,i02,R(x),exp,0.5xhas different expressions. The Gaussian kernel function, ,is always used, and the output of RBF neural network is. Two stages are included in RBF network. and of all implicit nodes are C,ii calculated by k-average clustering algorithm and all the input samples in the first stage. Then, according to training samples and least square method, is solved after wi the implicit layer parameters are calculated. In the light of the reasonable input parameters and the prediction principles, the input and output stability based on radial basis function can be calculated, trained and predicted by the functions in MATLAB tool. Combining with neural network, the GM (1, 1) is used to setup the grey neural network prediction model. A series of prediction values can be acquired to the raw series stability while GM (1, 1) is setup to many series. But a certain deviation still existed, which is related to the raw unintuitive series. Thus, the relationship between series and the deviation of prediction and original stability should be taken into account. The prediction value is considered as the input samples of neural network, and the original stability as the output sample. Using a certain stability structure, the network will be trained and series of well-trained weight and threshold values can be acquired. The prediction in one or more different time of different GM (1, 1) is as the well-trained input of network from which the final prediction in the next time or next different time will be carried out. As to the algorithm, the CGNN prediction is [1, 7]introduced in detail in reference . In stability control of surrounding rocks in deep entry in coal mine, it is very complicated that the variables inside the stability system are produced at the beginning of the model setup. The variables explained in the model should be selected correctly, which, on one hand, relies on the further study and cognition by the model builder to the system and on the other hand, on the quantitative analysis. To solve this problem, the grey relational principle will bring active action on it. z,z,？？,zLet y be the system variable, are the positive or negative 12n ,zcorrelated factorsis the relation on the basis of to y. Given the lower threshold ii ,z,,,value, ,can be deleted while,in which parts of explaining variables 0ii0 relating to the weak relation can be deleted in the stability system. To the network and using the method above, the input variables of network are selected, which can simplify the input samples greatly. Let,be grey prediction value, ,the prediction 12 ,value by neural network, prediction value by optimal combined model. The c ,,,,prediction errors areand respectively.The corresponding weighted c12 w,,w,，w,w,ww，w,1coefficients areand ,and ,.Thus, the errors and cc11221212 ,,w,，w,variations are as. c112222Var(,),Var(w,，w,),wVar(,)，wVar(,)，2wwCov(,,,). c112211221212 wAs to, in order to determine the functional minimum value, let 1 22 Var(,),wVar(,)，(1,w)Var(,)，2w(1,w)Cov(,,,)c111211122,Var,()cand. ,Var(,),w,0,Var(,)，Var(,),2Cov(,,,)c1121222,w1 22Obviously, ,then ,,,Var(,)(,w),0c1,,,Var(),Cov(,)212w,1Var(,)，Var(,),2Cov(,,,)1212and . w,1,w21 Because ,let,, then the weighted Cov(,,,),0Var(,),,Var(,),,12111222 coefficients of combined prediction are ,,,,,,. 22112211,，w,w,,,,12c12,，,,，,，，,,,,11221122 11221122 IV. CASE STUDY In the process of low stability control of surrounding rocks in deep entry in coal mine, the stability control parameters of surrounding rocks will cause serious accidents in coal mine production safety. How to forecast the stability control of surrounding rocks and control the ultra-limit of stability of surrounding rocks has been the focus of disasters and difficulties. In the recovery process, stability control of surrounding rocks is influenced by many factors and constraints such as Tensile strength, Elastic modulus, Possion ratio, Appearance density, Inner friction angle, Cohesion, Residual inner friction angle, Residual cohesion, Tensile strength and so on. Therefore, the stability control system of surrounding rocks is a multi-variable system whose characteristic equation is of generally high-order, which is difficult to use the same analytical style to quantitatively describe of the stability control changes of surrounding rocks and the complex function relations among the factors. No matter what means are often unable to obtain all the information. All decisions are made between some pieces of known information and partial unknown information. Therefore, the stability control system of surrounding rocks is grey. Through the study, it is found that, using the grey control system theory, the modeling and forecasting techniques are applied to analyze the stability control changes of surrounding rocks. Based on grey relation clustering models, the dynamic models are created to solve the practical problems in order to avoid the difficulties is solving high-order differential equations. At the same time, the application of the dynamic prediction model can better predict the stability control changes of surrounding rocks so as to forecast and control the stability control changes of surrounding rocks to prevent accidents. A. Prediction of stability control of surrounding rocks at upper corner of working face In a coal mine, the working face is at the level of -736 meters underground. Using the grey relational analysis, some main variables influencing the stability control of surrounding rocks are selected [2] and the measured data is shown in TABLE I. In TABLE I, data group 1-7 are used to establish GM (1, 3) prediction model to forecast the stability control of surrounding rocks of upper corner at the level of -736m of A1 working face, which is coMPared and analyzed with the 8th measured data. Then, the prediction model is analyzed by the grey errors and accuracy. According to the prediction model, the original sequence is coMPared with the measured values shown in TABLE II. From the results, it can be seen that the prediction model residuals and relative errors meet accuracy requirements. TABLE I MEASURED DATA 21 July 26 July 31 July 5 Aug 10 Aug 15 Aug 20 Aug 25 Aug Group 1 Group 2 Group 3 Group 4 Group 5 Group 6 Group 7 Group 8 Stress of surrounding rock 0.55 0.54 0.61 0.62 0.67 0.66 0.69 0.67 Pressure of surrounding rock 633 589 583 603 645 684 721 722 Strength 0.93 0.95 0.94 1.01 1.11 1.20 1.22 1.20 TABLE II ORIGINAL DATA RESIDUAL CHECKING 000kˆDate ,x(k)x(k),(k)/%111 20 July 1 0.55 0.5503 0.0003 0.055 25 July 2 0.56 0.5503 -0.0097 -1.73 30 July 3 0.62 0.6175 -0.0025 -0.40 5 Aug 4 0.61 0.5908 -0.0192 -3.15 10 Aug 5 0.67 0.7012 0.0312 4.66 15 Aug 6 0.68 0.6988 0.0188 2.76 20 Aug 7 0.72 0.7214 0.0014 0.19 With the elapsed time and the increasing information, a group of forecasting models can be created. Only 4 forecasting models are simulated so the models are changed with time. The models above are checked over the grey accuracy errors, which meets the precision. In the whole process, the coMPared curve of predictive values and actual values are shown in Figure 1. Fig. 1 CoMParison between prediction and practical data B. Training and checking of CGNN The quantification and normalization of input and output functions of neural network are that the function values are normalized to the interval [0, 1]. In accordance with the characteristic and normalized parameters, the input and output functions are quantified and normalized. Through site measurement and relevant research information, 22 groups of data in a typical mine are selected. All the data are normalized to be shown in TABLE III. 6 random samples of data are used for checking the samples, and the remaining 16 sets of data are used for training. During training, the selected -6square error is 10. Ahead of each training cycle, a new round of samples is randomly sorted. In checking, the output function of each checking sample meets the accuracy requirements. Using CGNN, the stress of surrounding rock at -328m is analyzed and predicted. After the data collation and normalization, ththe 9 group of data in TABLE III is formed. Inputting the data into neural network, it can be seen that the average stress of surrounding rock is 2.01%. Rockbolt is used to eliminate stress accumulation of surrounding rock. Practically, the local supporting is used to eliminate the local stress accumulation, which achieves good results. And the stress accumulation is down to permitted scope, which ensures the normal coal deep entry recovery. Therefore, the results by CGNN are in line with the field stress value of surrounding rock, which is the same as that in deep entry in the working face and guides the production practice. TABLE III TRAINING AND CHECKING DATA BY CGNN Tensile Elastic Possio Appearance Inner Residual Residual Tensile inner Data strength modulus ratio density friction Cohesion cohesion strength friction ID Cp,,,t,tE angle angle Cr (MPa) 003,p(),r()(MPa) (GPa) (KN/m) (MPa) (MPa) 1 0.39 0.45 0.45 0.56 0.78 0.56 0.55 0.13 0.02 2 0.51 0.56 0.55 0.5 0.78 0.5 0.9 0.05 0.03 3 0.39 0.45 0.55 0.56 0.78 0.6 0.55 0.40 0.03 4 0.39 0.56 0.65 0.5 0.78 0.35 0.9 0.13 0.02 5 0.51 0.45 0.65 0.6 0.78 0.33 0.66 0.05 0.03 6 0.39 0.45 0.65 0.56 0.78 0.56 0.56 0.40 0.03 7 0.51 0.56 0.48 0.56 0.78 0.5 0.55 0.07 0.04 8 0.32 0.45 0.51 0.5 0.78 0.6 0.9 0.13 0.02 9 0.42 0.51 0.55 0.6 0.78 0.65 0.66 0.05 0.02 V. CONCLUSIONS Using the grey control theory, the grey correlation prediction model is established. The practical application and analysis show that it is fully capable of reflecting the changes of stability control of surrounding rocks at the upper corner. The dynamic forecasting model can not only predict the stress changes of surrounding rock and trends, but also the iMPact of the major factor change on stress of surrounding rock. However, the original data by dynamic forecasting model is less than required. Generally, the 4-dimensional data can meet the model establishment, which is very conducive to the scene and the existing monitoring system. Only considering the iMPact of the stress at working face and tensile strength can be through the modeling analysis. The stress distribution at surrounding rock and the choice of stress accumulation treatment are the results by a variety of factors. CGNN model, combined with the influential factors, the stress accumulation of surrounding rock is set up, which can map out the complex non-linear relationship. The improved CGNN algorithm can speed up the convergence, avoid oscillation computing and overcome local minimum value. After training and checking, the generated neural network is correct and practical. And in the subsequent applications, the modification and improvement are necessary. Successful application examples show that the neural network and grey relation clustering are the effective methods to analyze the stress accumulation. 中文译文 应用灰色关系集群和CGNN分析矿井深部巷道围岩的稳定性 控制 12 杨万斌 曲志明 (1. 北京理工大学，北京，100083;2. 河北工程大学，邯郸，056038) 摘要——应用神经网络及灰色关系集群，建立模型，用以解决预测比较矿井深部入口围岩稳定性控制参数，结果显示灰色关系群集是一种有效地方法，同时神经网络能够很好对短期效应做出预测，结合灰色神经网络对时间依赖的波动趋势特点得出改善的方法，应用简单的要素结合神经网络预测灰色波动趋势，描述在深部入口的围岩稳定控制。 一、前言 美国灰色系统技术的不确定性和贫乏的小样本资料。与随着发展而产生的未知信息，真实的世界将被发现，系统运行的实际行为将会被正确掌握。通过与原有稳定信息，结合灰色系统的图像预处理法，尽管现实世界是表达复杂和不断的 [1]变幻地，但能能够揭露某些内在规律。这个研究是基于灰色系统技术的基于已知信息而产生部分未知信息中获取可靠有价值的规律，正确认识和有效的控制系统的行为。 神经网络依赖它的内在关系模型，该模型是自发的，自觉适应的。神经网络能解决变量的困难预测和避免传统的干扰，一如人的思想系统。灰色关联分析的几何参数的关系曲线，确定程度越靠近曲线形状相似度越大，相应的序列相关性越好。相似的相关系数和关联度的影响结果受多种因素的制约。更大的影响关系是同时分析了实用的系统，数据系列的行为特征。此外，它还可以确定有效的影 [1,2]响因素，即选择系统各自的行为特征。 虽然这个目标体系的发展和表达，其复杂的逻辑规律变化仍然是以不同功能实现的，但实质是协调统一的。因此，如何找到它的内在发展规律的分散稳定似乎都是很重要的。针对上述描述，可以发现，结合灰色关联聚类将对煤矿深部入口围岩稳定性控制产生重大影响。结合灰色神经网络模型的建立将分析解决这一问。 二、灰色关系群集 A 灰色关系群集 作为通用的灰色趋势关系系统，D. J. CHEN等人做了大量的工作。为了将其应用到实践中，最基本的理念对他的研究作了简要介绍。相似的,在这个动态系统的行为才能用灰色趋势的关系(GTR)。借助于GTR、隐性的系统运行规律也许可以更贴切地描述。一般而言，一般系统理论应用于一般GTR体系的基础上，结合GTR、体系化的分析模型，并推导出GTR。假定U是所规定的因素，W是对 R,(,)比因素，uy R是GTR在(B, W)中的参数。这个矩阵即是GTR，矩uyqip，kn(0)(0)1,,b(k),,阵在(k,1,2,？？,n)(q,1,2,？？,p)w(k)(i,1,2,？？,h)，，和,，然qi,,,(k),qiqi,n1,2k而对 这种趋势关系及这种趋势关系函数，(B, W)是被限定的。,(k)qi(0)(0)(0)，，及Q都属于一般的GTR,z(k),b(k),w(k),,,,,,0,1(k,2,3,？？,n)qiqi 体系，其中。 ,,Q,(B,W),Ruy 本文为了并使其适合应用，先将对一些定义在此介绍。V是对系统空间的评价，Q 和 H是评价功能。因此，在Q和评价空间之间的关系被描述成H:B，W,V因此，一般的GTR体系模型是广义的，包括利用GTR进行问题的分析。为了解决不同的问题，GTR应该对B, W 和 H是不一样的，GTR矩阵的基础上，GTR聚类方法，观察指标或物体分成许多可定义、分类。这个类可以被看作是观察对象的不同的种类。事实上,任何观察对象都有许多特性指标不准确的 [3-6]分类。通过聚类、GTR相同的分类因素和复杂的系统将会被简化。 Z，GTR系统这个因素,有h因素。每一个代表了一个序列 ,。为特定的关联映射，这种趋势在z,,,Z,z(k),k,1,2,？？,n;i,1,2,？？,mijji 提到的关系因素，;，。z,z,Z,,,(z,z)i,j,Mijijij ,,,,,,i,1,2？？,m;j,1,2,？？,m由Z和组成的GTR系统被称为GTR自身相ij ,关系统。是一个GTR矩阵，H是评价尺度，而V是评价空间。至于，Q,(Z,,) 是聚类分析的极限，同时这评价尺度被限定为。 和 是特征,,Z,,0,1,,,Zjcci同类项，，。在极限中的分类，系统特征的可变因素为趋势关系群,,,i,j,cc,,V集，这个系统输出的聚类,主要表现。而是一套,,,,V,V,,,,,1,2,？？,, ,,m包括一组的特征变量，同上，同时。 , B.灰色关联度预测 (0),,,,假定fj,1,2,？？,m是GTR对时间依赖的序列，而是已知的模型Z(k)j, f(k)(j,1,2,？？,n),组。每个模型针对GTR序列可以由一群预测数据，。为j,j(0)(0)(0),,f(k)GTR属于和。的含义，F,f(k)与和,,Z(k)Z,Z(k)Z(k)ji,i,f(k)相同的。Z 和F诗预测组，k =1,2,……,n,i =1,2,……,h, (j,1,2,？？,n)j ,,,,j=1,2,……,m . Q = ((Z,F), )是GTR预测体系，当H : 为GTR预测opt H:Z，F,V和评价尺度，且V是系统预测效果的评价空间，体系被映射成。 三、CGNN 利用GM(1,1)预测序列是一种最常见的应用。因为这个灰色模型的稳定性，获得了一些预测误差规律和许多不同的可能出现的独立模型将安装到许多相关的序列，它是可以不考虑之间关系的稳定序列。一般来说，这个缺陷可以通过设置结合模式来解决，如:应用灰色神经网络组合预测模型(CGNN)。 TTX:(x,x,？？,x)为输入样本，而y为单独输入U,(u,u,？？,u)，这隐含的1212nnT节点输入W,(w,w,？？,w)，重量结合隐性和输出节点这个连接权重值为1,12n 隐含节点之间的输入信号传送给隐含节点的输入。即输入第i隐含节点为 ,,u,RX,C,，i为隐含节点的数目i,1,2,？？,m。R为径向函数也是一类高iiithiC,斯核函数。X为输入样本。为中心的径向基函数的神经元。是宽度参数ii,2th2,iX,C的径向基函数的神经元。是欧氏分，隐节点的活化作用有不同的y,wexp,0.5X,C,iiii,i0, 2表达。高斯核函数，常被使用，并输出RBF神经网络为， R(x),exp,0.5x 两个阶段包括在径向基函数神经网络。在所有的隐节点中和 按开平均聚类C,ii算法计算，所有的输入样本在第一阶段。然后，根据神经网络尝试样本和最小二乘法拟合方法，解决了隐层后的参数计算。依据合理的参数和预测原理，输入和输出稳定性基于径向基函数可以计算，预测功能可在MATLAB工具实现。结合神经网络，GM(1，1)灰色神经网络建立了预测模型。一系列的预测值能获得稳定的原序列，同时GM(1,1)应用到许多系列。但是仍然存在一定的偏差，与原系列存在直观差距。因此，基于之间的关系，提出了一系列带偏差的预测，并对原来的稳定性加以考虑。预测价值作为神经网络输入样本，与原来的稳定性和输出样本。用某种稳定结构，网络将接受演算，演算的重量和系列可获得的极限值。作为对算法的预测，CGNN详细介绍，参考预测在一个或多个不同的时间，不同的GM(1,1)是演算的输入的网络，最后预测在接下来的时间或下个不同的 [1, 7] 时间将会被执行。 对煤矿深部围岩稳定控制项目，这是一个很复杂的内部变量产生稳定系统之初的模型建立问题。这个变量的模型的解释，应该选择正确。一方面依靠进一步研究和认知的模型，对系统的建造;另一方面，进行了定量分析。为了解决这个 是正问题，灰色关联原则将带来积极的。将y作为系统变量，z,z,？？,z12n面或负面的相关因素表现于和y的关系坐标中。给出了较低的极限值，当,zii ，，可以被忽略，在解释变量与弱性关系可以删除的稳定系统中，网,,,,zi00i 络和使用方法，对网络的输入及变量的选择，均可简化输入样本。让灰色预测,1价值，由神经网络预测模型的优化组合预测价值。预测误差为与均,,,,,,cc212为独立的。 相应的加权系数为和w，同时，,,w,，w,，因此，w,ww，w,1cc11221212 误差和变量描述为,,w,，w,。 c112222 Var(,),Var(w,，w,),wVar(,)，wVar(,)，2wwCov(,,,)c112211221212 对，为了确定功能的最小值，需 w122 Var(,),wVar(,)，(1,w)Var(,)，2w(1,w)Cov(,,,)c111211122,,Var()c,Var(,)，Var(,),2Cov(,,,),Var(,),w,0且 12122c12,w1 ,,,Var(),Cov(,)22212w,显然，,,，此时 ,Var(,)(,w),01c1Var(,)，Var(,),2Cov(,,,)1212 w,1,wCov(,,,),0Var(,),,Var(,),,同时，由于，让，然后加2112111222,,,,22112211,，w,w,,,,权系数的组合预测为，， 12c12,，,,，,，，,,,,1122112211221122四、个案研究 在这个过程中对煤矿深部入口围岩稳定性控制，在煤矿安全生产中围岩稳定控制参数不准确也会引起严重的事故。如何预测控制围岩的稳定性和控制围岩的 稳定性在临界条件下的灾害和困难。在恢复过程中对围岩稳定性控制，受多种因素的影响和制约因素，如抗拉强度、弹性模量、跳跃-扩散率、外观、内部摩擦角、密度、内聚力、剩余内部摩擦角、残留的内聚力、抗拉强度等。因此，对围岩稳定性控制系统需要一种多变量系统的特征方程的描述，围岩控制的稳定变化和复变函数关系的因素，这是很难用同一解析式定量描述的。无论什么样的手段往往不能获得所有的信息。所有的方案决定于一些已知信息和部分未知的信息的可靠性。因此，对围岩稳定性控制系统是灰色的。通过研究，结果表明，应用灰色系统理论，建模和预测技术的应用，分析了控制围岩稳定性的变化。基于灰色关联聚类模型、动态模型建立过程中存在的问题，并对解决实际困难是为了避免高阶微分方程的求解。同时，应用动态预测模型能更好地预测控制围岩稳定性等的变化，对预测和控制围岩稳定性控制的变化来防止事故发生。 A. 工作面上隅角围岩稳定性预测控制 在煤矿工作面为水平-736米的地下。利用灰色关联度分析，在一些主要影响 [2]因子的稳定控制围岩的基础上，选择的测量数据，即是显示在表1中，由表1可见，1-7群集建立GM(1、3)预测模型，预测控制围岩稳定性的A1工作面-736m的上隅角，比较和分析了在第八个测点测量数据。然后，该预测模型进行了灰色的误差和精度演算。据预测模型，预测了原序列对比测量值显示在表2。从结果可以看出，预测模型的剩余误差和相对误差满足精度要求。 表1 测量数据 7.21 7.26 7.31 8.5 8.10 8.15 8.20 8.25 1 2 3 4 5 6 7 8 围岩应力 0.55 0.54 0.61 0.62 0.67 0.66 0.69 0.67 围岩压力 633 589 583 603 645 684 721 722 强度 0.93 0.95 0.94 1.01 1.11 1.20 1.22 1.20 表2 原始数据的残留 000kˆ, x(k)x(k),(k)/%111日期 7.20 1 0.55 0.5503 0.0003 0.055 7.25 2 0.56 0.5503 -0.0097 -1.73 7.30 3 0.62 0.6175 -0.0025 -0.40 8.5 4 0.61 0.5908 -0.0192 -3.15 8.10 5 0.67 0.7012 0.0312 4.66 8.15 6 0.68 0.6988 0.0188 2.76 8.20 7 0.72 0.7214 0.0014 0.19 依据时间和预测信息，一组预测模型被建立。只有4组预测模型的模型模拟随时间不断变化。上述模型的精度误差检查，能够满足灰色的精度。在整个过程中，曲线的预测值，比较实际值如图1所示。 图1 预测和实际数据进行比较 B.CGNN的培训及检验 量化和规范化的输入/输出函数的神经网络，都归一化于函数值的区间[0， 1]。按照规范的特点和参数，输入和输出功能进行量化和规范化。获得的现场实 测和相关研究资料，及在一个典型的矿井的22个组的数据。所有的数据都显示在 表3。 6个随机抽样的样本数据用于检查样本，其余的16套数据用于培训。在演算 后，选定方块10-6误差。每个演算周期之前，新一轮的样本是随机排序。检查输 出功能并检查样本是否满足精度要求。使用CGNN对-328m的围岩应力进行了分 析和预测。这个数据的第九组数据整理、标准化后见表3。 输入的数据到神经网络，可以看出，对围岩平均应力2.01%是用来消除锚喷 支护巷道围岩压力累积。实际上，这个地方的支持是用来消除了局部应力积累， 取得了良好的效果。压力积累下来的安全许可范围，是用以确保正常的煤炭深入 口应力回复。因此，CGNN演算结果均符合围岩应力值，这是一样的，可以指导 在深部开采工作面的生产实践。 表3 培训和检验CGNN数据 拉伸 弹性 泊松内部 剩余 剩余 拉伸 数可视密度 比强度 系数 摩擦 内聚力 内摩擦 内聚力 强度 据 E,, 角 Cp(MPa) 角 Cr ,t,tID 00(MPa) (GPa) (KN/m3) (MPa) (MPa) ,p(),r() 1 0.39 0.45 0.45 0.56 0.78 0.56 0.55 0.13 0.02 2 0.51 0.56 0.55 0.5 0.78 0.5 0.9 0.05 0.03 3 0.39 0.45 0.55 0.56 0.78 0.6 0.55 0.40 0.03 4 0.39 0.56 0.65 0.5 0.78 0.35 0.9 0.13 0.02 5 0.51 0.45 0.65 0.6 0.78 0.33 0.66 0.05 0.03 6 0.39 0.45 0.65 0.56 0.78 0.56 0.56 0.40 0.03 7 0.51 0.56 0.48 0.56 0.78 0.5 0.55 0.07 0.04 8 0.32 0.45 0.51 0.5 0.78 0.6 0.9 0.13 0.02 9 0.42 0.51 0.55 0.6 0.78 0.65 0.66 0.05 0.02 五、结论 利用灰色的控制理论，建立了灰色关联度预测模型。实际应用的分析表明， 它是完全有能力反映围岩稳定及控制上隅角的变化。动态预测模型不仅可以预测 的围岩应力变化趋势，而且可检测影响围岩应力变动的主要因素。然而，原始数 据的动态预测模型，低于所必需的要求。一般来说，能够满足了四维数据模型的 建立，非常有利于现场和现有的监控系统。只考虑影响采煤工作面，其压力和抗 拉强度可以通过建模分析。 处理各种影响围岩应力分布和应力累积因素的结果。以CGNN模型为基础， 结合影响因素、围岩压力积累，可以在坐标上标出了复杂的非线性关系。CGNN 改进算法，可以加快收敛速度，避免振动计算和克服局部最小值。培训和检查后， 生成的神经网络的正确性和实用性，在随后的修改和完善应用，是必要的。成功 的应用实例表明，神经网络和灰色关联聚类是分析压力累积的有效方法。 参考文献 [1] PENG Ying, “The Applications of Stock Analysis Based on Grey System Theory of DM,” Ph.D. dissertation, Dept. Management Eng., Changsha University of Science & Technology, Changsha, China, 2006 [2] LV Pin, MA Yunge, and ZHOU Xinquan, Research and application on dynamic forecasting model of gas consistence in top corner, Journal of China Coal Society, vol. 31, pp. 461-465, Aug. 2006. [3] Chen D.J., Xiong H.J., Chen M.Y. (2004) Grey trend relational clustering and its application in data mining, Systems engineering and electronics, 26, 599-601. [4] Yang J.Q., Xing Z.L., Li Q., et al. (2004) Study on seasonal regular trend value and data digging method about beer market, Liquor-making science and technology, 6, 104-107. [5] Chen D.J. (2003), Sheng Y.Z., Chen M.Y. (2003) The general grey trend relation system and its analysis, Journal of central China university of science and technology, 31, 82-84. [6] Wang R.X., Liu W.C. (2004) The core algorithm in business data mining, Journal of Beijing university of technology, 30, 510-514. 内部 资料， 请勿 外传～ 序规格型名称 单位 数量 备注 号 号 一 制冷系统 1 压缩机组 4AV10 台 4 2 冷凝器 LN-70 台 1 3 贮氨器 ZA-1.5 台 1 ZWB-1.4 桶泵组合 台 1 5 5 氨液分离器 AF-65 台 1 6 集油器 JY-219 台 1 7 空气分离器 KF-32 台 1 8 紧急泄氨器 JX-108 台 1 KLL-259 冷风机 台 8 0 1KLD-15冷风机 台 4 0 0 1KLD-10冷风机 台 2 1 0 1阀门 套 86 2 1电磁阀 套 6 3 1管道及支架 吨 18.6 4 1管道及设备3 m22 5 保温 1管道保温包镀锌板 吨 1.6 6 扎 1附件 套 1 7 二 气调系统 中空纤维制1 CA-30B 台 1 氮机 二氧化碳洗2 GA-15 台 1 涤器 3 气动电磁阀 D100 台 14 电脑控制系CNJK-44 台 1 统 06 5 信号转换器 8线 台 1 果心温度探6 台 7 头 37 库气平衡袋 5 m 个 7 8 库气安全阀 液封式 个 7 小活塞空压9 0.05/7 台 1 机 1PVC管 套 1 0 1附件 套 1 1 三 水冷系统 DBNL-31 冷却塔 台 2 100 SBL80-2 水泵 台 2 160I SBL50-3 水泵 台 2 160I 4 阀门 套 30 5 管道及支架 吨 2.8 6 附件 套 1 四 电仪控系统 1 电器控制柜 套 1 2 照明系统 套 1 3 电线电缆 套 1 4 桥架管线 套 1 5 附件 套 1