A forecasting and forewarning model for methane hazard in working face of coal mine based on LS-SVM

Availableonlineatwww.sciencedirect.comJOURNAL OF CHINA UNIVERSITY OF. ScienceDirectMININGTECHNOLOGYELSEVIERJ China Univ Mining Technol 18(2008)0172-0176www.elsevier.com/locate/jcumtA forecasting and forewarning model for methane hazard inworking face of coal mine based on LS-SvMCAO Shu-gang, LIU Yan-bao, WANG Yan-ping2' Key Laboratory for the Exploitation of Southwest Resources& the Environmental Disaster Control Engineering,Ministry of Education, Chongqing University, Chongqing 400044, ChinaCollege of Petroleum Engineering, Chongqing University of Science and Technology, Chongqing 400042, Chinaimprove the precision and reliability in predicting methane hazard in working face of coal mine, we have proposed ac二d forewarning model for methane hazard based on the least square support vector(LS-SVM)multi-classifier andhine. For the forecasting model, the methane concentration can be considered as a nonlinear time series and the timemethod is adopted to predict the change in methane concentration using LS-SVM regression. For the forewarningis based on the forecasting results, by the multi-classification method of LS-SvM, the methane hazard was identifiedto four grades: normal, attention, warning and danger. According to the forewarning results, corresponding measures are taken. Themodel was used to forecast and forewarn the K9 working face. The results obtained by LS-svM regression show that the forecast-ing have a high precision and forewarning results based on a LS-SVM multi-classifier are credible. Therefore, it is an effectiveodel building method for continuous prediction of methane concentration and hazard forewarning in working faceey words: working face: methane concentration; LS-SVM; forecasting; forewarning1 Introductioninherent disadvantages, such as over-fitting, trainingcourse being affected by the local minima,and selec.Mine methane is one of the main dangerous factors tion of network relying excessively on experiencein coal mining. The quantity of methane emission in Therefore, the prediction precision and reliability ofworking face accounts for about 40%-80% of the neural network is not satisfactory. Recently, a metotal gas emission of coal mine. So the working face thod of LS-SVM was proposed by SuyKens JAK-is the chief source of gas. Meanwhile, it is one of the In this method, the least square linear system wasmain places of gas accidents. So it is very important adopted as the loss function, which significantly acto forecast and forewarn the methane hazard in celerates the solving speed by changing the inequalityworking face in order to maintain the safety of coal constraints into equality constraints. Therefore, thisminingmethod has been widely recognized and accepted inOver-limit methane concentration in working face pattern classification and regressionhas always been the direct cause of gas accidents.The methane concentration in working face is af-Many scholars have studied this problem by using fected by many engineering factors, such as occur-modern mathematical methods and computer tech- rence state of gas, gas permeability of coal seam andnologies. The methods commonly used are grey sys- surrounding rock, gas aspiration of coal seam and gobsem prediction method has a high precision for mining system 2-1Se ow, production capacity andshort-term prediction whereas, it can not fit well to vironment conditions keep unchanged, the methanethe data series which have large random fluctuation, concentration can be considered as a time series in athus has lowertion precision 5. The neural relatively short period of time, which changes overnetwork methodtraining course follows thela information. Theprinciple of minimizing the experience risk has many中国煤化工LSsMCNMHGReceived 09 December 2007; accepted 15 February 2008Project 506741 I I supported by the National Natural Science Foundation of ChinaCorrespondingauthorTel:+8613594167483:E-mailaddresscqulybesina.comCAO Shu-gang et alA forecasting and forewarning model for methane hazard in working face ofsion and classification method to the methane hazard by changing the constrained problem into an uncon-forecasting and forewarning in the working face. In strained problem and introducing the Lagrange mul-e forecasting model, the methane concentration istipliers a, we obtain the objective functiononsidered as a nonlinear time series and then theime series analysis technology is adopted to predictL啊.ba)=1(w)-∑甲x)+b+e-对(2)the change in methane concentration by the LS-SVMregression. According to the prediction result, we can According to the optimal solution of Karush-Kuhntake the advantage of the LS-SVM classification ma- Tucker(KKT)conditions, take the partial derivativeschine to establish the forewarning system which can of (2)with respect to w, b, e and a, respectively,identify different hazard grades and predict the me and let them be zero, we obtain the optimal condithane hazard in futuretions as follows2 Forecasting model=∑αx),∑a=0,a=,Forecasting the methane concentration in workingwP(r)+b+e-y=0face can be regarded as the Ls-SVM regression algoedure is as Fig. 1:o, the following linear equations are obtainedY ZZ+ISupport vectorwhere z=lor)y,.,r)y,)[x)[飘x)[玩,) Nonlinear function=[x,….y],=[1,…,1],a=[x区][x2]KmmnDefine K(,x)=dx).Mx), i=l,.,7which is satisfied with Mercers condition. TheLS-SVMFig. 1 Forecasting model of methane concentrationf(x)=∑aK(x,x)+bAt first, we can obtain a time series of methaneconcentration using methane detectors,Common examples of K(r, r: )arex∈R,y∈RPolynomial kernel functionwhere xi is the ith input vector of time series; n the(xx)=(x·x)+1,q=12dimension of input vectors(n=1); y: the ith output of Gaussian radial base kernel function(RBF)methane concentrationThe nonlinear function f( shows the relation-K(r,x)=exp(-ll r-x,IP/20)ip between the methane concentration and time. and Sigmoid kernel functionAccording to the LS-SVM theory, the nonlinearK(x,x)=tanh(x·x)+C]unction is defined asf(x)=wo(r)+bAs seen from Eq (4), LS-SVM is less computa-tionally complex compared to SVM, in which unwhere w is the l-dimensional weight vector; oo equal constraints is replaced by equal constrains.Thisthe mapping function that maps r into theleads to solve a set of linear equations instead ofL-dimensional feature vector and b the bias term.quadratic programming program. Substituting theHaving comprehensively considered the complexity of function and fitting error, we can express the predicting vectors (r+5 +21", Ji+) into trainedregression problem as the constrained optimization LS-SVM, we obtained the predicting value of me-problem according to the structural risk minimization thane concentration.principleTy(1). Forewarning modelCombining the hazard grades and LS-SVM classi-which is subjected to the constraintsfication machines, we established the forewarningy=wq(x)+b+e(i=12,…,l)中国煤化工Based on the me-where y is the margin parameter; and e the slack thandResults in workingCNMHGpotential methanevariable for rhazard, so corresponding emergency measures can beIn order to solve the above optimization problems, adopted in time to avoid accidentsJoumal of China University of Mining& TechnologyVol 18Methane consistency data hTaking the partial derivatives of Eq (9)with respectLS-SVM learniLS-SVM regressito wh,bk, e and a k, respectively, and let them bezero, the decision function of multi-category LsILS-SVM classificationSVM is obtained as follows:Forewarning resultsg(y)=sgn∑ax"“K4(y,y2)+b](10)Practically, The MATLAB LS-SVM toolboxNormal Attention Waming DaSearch reason Take meaproblemsFig 2 Forewarning model of methane hazard4 PretreatmentIn LS-SVM classification, for a given set of da- 4.1 Selection of tr{(y,v),i=1,2,…,l,y∈R,v∈{L+],另For the forecasting model, the selection of trainingis the methane concentration and vi the hazard class. samples is a dynamic process, namely, the monitoringThe optimum hyper-plane can also be shown as data within a shorter interval of time were chosen asg0)=way)+b According to the structural risk mi- training samples, and updated continuously with thenimization principle, choosing the 2-norm of e as sampling. By this way, the continuous and promptthe slack variable, the problem can be transformed prediction of methane concentration can be impleinto an optimization problem of Eq (1), solved it in mented. For the forewarning model, the training sam-the same way as in Eqs.(2-4), the decision functioncan then be obtained as:namely, with time increasing, the number of trainingsamples should be properly increased in order to sa-g(y)=sign∑aK(y,y)+b(6) tisfy the variable conditions and improve the fore-warning accuracy in the premise of guaranteeing theThe methane hazard forewarning is a multi-classcognition problem, which need to be resolved by 4.2 Sparseness of samplesthe LS-SVM multi-category method. Assuming thatA disadvantage of the LS-SvM is that the Lagranthe learning sample of multi-category is(vi yi) gian multipliers for the LS-SVM tend to be all noni=L,. ,L,k=l,.m, where m is the hazard grade, zero, whereas for the SVM case, most of the multipliers are zero(only support vectors are nonzero)v() is the ith sample of the kth grade. The multi- Therefore, the LS-SVM is lack of sparseness to somecategory problem can be transformed into several extent. To solve this problem, we referred to Referbi-class problems through a proper coding program. ence [19] for the process of sparseness. SupportingAccording to the mine Safery Regulation 4, the me- vectors with small absolute values of the associatedthane hazard is divided into four grades by methane dual variables were pruned and an LS-SVM is re-concentration: normal (CH4<0.5%), attention trained using the reduced set of training data. Thedanger(CH21.5%). In this paper, the laur %)and process is repeated until meeting the requirement of(0.5%CH4<1.0%), warning(1.0%≤CH4<1.5were encoded using MOC method as followin4.3 Selection of kernel functions and optimiza.[-1,-1],[-1,1,[1,-1]and[,1].Then, the mul-tion of parametersti-category LS-SVM can be shownTo apply the previous method, kernel functionsquired to be specified for LS-SVM. Theoretically,2呢+少)址 satisfy Mercer condition canwhere m is the number of binary classification. different kernel functions and parameters, the per()is subjected toformance and training results of LS-SVM model arsignificantly different. In this research, the RBFvIw 9()+b]=l-e,k (=1, m)(8) kernel function is chosen for the forecasting and foIntroducing the Lagrange function, we haverewan中国煤化工 imon experence(w,,b,, ek, a r)=J(w,be,ek)CNMHGa{w"w(y)+b]-1+e(9) The values of the two parameters greatly affect theCAO Shu-gang et alA forecasting and forewarning model for methane hazard in working face of.training and generalization capability of LS-SVM, lyzed and 200 groups of data selected as trainingwhich can be obtained by cross footing tests referring samples, which satisfy the premise of multiplicityto Reference [11Parts of the training samples are shown in Table 2.The rbF kernel function was chosen in the same5 Case studyprocedure and parameters y=5.171 an0.5413 were obtained by cross footing tests5. 1 General situation of the working faceSituated near Chongqing, China, the Moxinpo coalMethane concentration (%) Hazard classesmine is a coal-and-gas outburst mine with a compli-cated geological structure. The main minable coalNormal [-1, -1]seams belong to Longtan group of upper Permian0.78[1,1stem. The dip angle of mining coal seam is betweenNormal[-1,-160°and63°. High gas pressure and high methanecontent are presented in coal seams. The K9 workingface is situated in the middle of the -115 m level asarningprimary mining working face of protective layer, of0.58Attention[-1,n1,1inclined length is 120 m and the mining height is 0.9 0121m. The relative amount of gas emission of K9 workUsing the LS-SVM classification model, the pre-ing face is 40.5 m/t, and the amount of absolute gas dictive values in Table I are forewarned and the re-sults given in Table 35.2 Prediction of methane concentrationTable 3 Fororking faceThe gas monitoring data obtained from the K9TimeActual hazardForewarningworking face having a continuous working for 1201.097armingWarninghours were chosen as the training samples to predicthe value of methane concentration in the next 10hours. Limited by the space, the training data werenot listed in this paper. The parameters obtained by0988cross footing test are y=20.113 and 0=0.012.According to the previous models, we predicted theAttentionmethane concentration of the working face, and theAttentionresults were shown in Table 1Table 1 Predicting and actual value of methaneconcentration in K9 working faceAs seen in Table 3, the predicting results agree wellwith the actual hazard grades, showing that the modelActual valuePredicting valueRelative erm0273is satisfactory0.0886 Conclusions3.9226.237In this paper, a forecasting and forewarning model60930of the methane hazard in working face of coal mine is3.871established based on the regression and classificationof LS-SVM. The methane hazard forecasting and0.866forewarning in K9 working face is highly accurate,which can well reflect the change of the methaneconcentration in the working face. The forewarningThe results in Table 1 show that the predicting results are also credible, which can satisfy the devalues agree well with the actual value and the rela- mand for forewarning the methane hazard in thetive error is between 0.088% and 6.915%, which isworking face. The LS-SVM can, therefore, be thewithin the error tolerance. Therefore, the predicting method to build models for continuous forecastingresults can satisfy the demand for safety production the safety monitoring data and forewarn methane ha中国煤化工5.3 Forewarning of methane hazardThe methane concentration of K9 working face and AckCNMHGcorrespondingly hazard grads for 3 months was ana-This work was supported by the National NaturalJoumal of China Uniof Mining TechnologyVoL 18 No. 2Science Foundation of China( Grant No. 50674111). [10] Yan Z G Zhang H R, Du P J. Application of SVM inMeanwhile, the authors would like to thanks theanalyzing the headstream of gushing water in coal mine.leaders of Moxinpo coal mine for their supplying theJournal of China University of Mining andfield data, and the reviewers of this paper for their2006,16(4):433438constructive comments and suggestions, which[Il] Jiang A N, Liang B. Forecast of water inrush from coalgreatly helped us to improve the quality of the paperfloor based on least square support vector machine.Journal of China Coal Society, 2005, 30(5): 613-617References[12] Tao Y Q, Xu J, Li S C. Predict gas emissing quantity ofining coal face with improved Grey Markov model1 Wang Y P. The Application Study on Risk ForewarningJournal of Coal Sociery, 2007, 32(4): 391-395.(In Chinese)stem of Gas Explosion and Gas Discharge Based on (131 Xu J L, Yu B J, Lou J F, et al. Characteristics of gasHazard Theoraster dissertation]. ChongqingChongqing University, 2006.(In Chinese)emission at super-length fully-mechanized top coal cav-[2] Wang Y P, Cao S G Wu X M, et al. Gas emission fore-ing face. Journal of China University of Mining andasting and fore-warning research of a working face (14) State Administration of Work Safety, State Administrabased on improved BP network. In: Proceedings of the2007 International Symposium on Mining Science andtion of Coal Mine Safety, Coal Mine Safery ReguleSafety Technology, Jiaozuo: Science Press, 2007(Part A)Beijing: Coal Industry Publishing House, 2004.(In1226-1231[3] Fu L Y, Hua G The research of coal mine gas concentra-[15] John S T, Nello C. Support Vector Machines and Othertion prediction system based on artificial neural networkKernel-Based Learning Methods. New York: CambridgeColliery Mechanical Electrical Technology, 2006(6)University Press, 200041-43. (In Chinese)[16] Liu ZG Li DR, QinQ Q, et al. An analytical ov[4] Lu P, Ma Y G Zhou X Q. Research and application onof methods for multi-category support vector madynamic forecasting model of gas consistence in topcomer. Journal of Coal Society, 2006. 31(4): 461-46510-13. (In Chinese)(n Chinese[17] GaoG H, Zhang Y Z, Zhu Y. Hybrid support vector ma-1 ZhangG Q, Qin Y J, Jiang W Z, et al. Study status andchines-based multi-fault classification. Journal ofprospect of methods for predicting amount of mine gasUniversity of mining and Technology, 2007, 17(2)emission in china. Safery in Coal Mines, 2007(8): 58-60.(In Chinese)[18] Suykens J A K, Gestel T V, Brabanter J D, et alI7 sm York: Springer, 199 tatis[6] Vapnik V N. The Nature of Statistical Learning TheorySquares Support Vector Machines. SingaporeAScientific Co Pte Ltd. 2002[7] Suykens J A K, Vandewalle J. Recurrent least squares[9] Xu F, Mei wS, Zhang Z H. Least squares support vectorsupport vector machines. IEEE Transactions on Circuitsmachines for aerial images segmentation. Geomatics andand System-I,2000,477):1109-1114Information Science of Wuhan University, 2005, 30(8)694698. (In Chinese)[8] Suykens JA K, Vandewalle J Least square support vec- (20) Guo H, Liu H P, Wang L Method for selecting parame-tor machine classifiers Neural Processing9(3):293-300ters of least squares support vector machines and appli[9] Qi H N. Support vector machines and application re-cation. Journal of System Simulation, 2006, 18(7): 2033-search overview. Computer Engineering, 2004, 30(10):2036.(In Chinese6-9. (In Chinese)中国煤化工CNMHG