Neural Network Identification Model for Technology Selection of Fully-Mechanized Top-Coal Caving Min

Dec.2001Journal of China University of Mining & TechnologVol 11 No. 2Neural network dentification modelfor Technology Selectionof Fully-Mechanized Top-CoalCaving miningMENG Xian-rui(孟宪锐), XU Yong-ymug(徐永勇), WANG Jin(汪进)Department of Resource Exploitation Engineering, CUMT, Beijing 100083, P.R. ChinaAbstract: This paper mainly discusses the selection of the technical parameters of fully-mechanized top-coal caving mining using the neural network technique. The comparison between computing results andexperiment data shows that the set-up neural network model has high accuracy and decision-makingKey words top-coal caving mining; artificial neural network; reformative back propagation neural net-CLC number: TD 823. 9 Document code:a Article ID: 1006-1266(2002)02-0199-051 IntroductionThe recovery ratio of fully-mechanized top-coalFully-mechanized top-coal caving mining will caving mining is always influenced by the selectiondefinitely become the dominant technology in theof mining technology parameters. These technologydevelopment trend of intensive production of mineralrs to decide are as follows drawinresources, because of its technical characteristics and mining height, drawing interval, and drawing se-advantages and its continuing maturity and perfecIntion in production practice. However, marked im- decided through industrial tests. However, thisprovement on its technical and economic benefits ismethod costs not only much time, but also muchgreatly dependent on the experience of coal face depersonal and material resources, thus leads to muchsigners and producers, this is because of the diversiwaste and affects normal production activities. Atty of coal mining technology caused by the complicathe same time. due to the effects of all kinds of faction of coal production environment and the indetertors,calculation is so difficult and complicated thatminism of coal preserving state. Computer techthe scientificalness and precision of computing re-nique development makes it possible to deal with a sults is very low. However, if neural network techgreat deal of data because neural network computa- nique is applied to the identification of the selectiontion uses the juxtaposing computation and the distri- of technology parameters of top-coal caving miningbution data disposal, so it is widely used in projectscan be done by a computer on thesuch as engineering structure, engineering analysis, ground. This method is convenient and timesavingengineering designing, engineering optimizing, geo-metric modeling, structural failure diagnosis, modelTYH魏dentification, and system reliability analysis.Received date: 2001-04-23ational Natural Science Foundation of China(59734090)an-rui(1951-), male, from Beijing, associate professor, engaged in the research and teaching of coal mining engineer-Journal of China University of mining & TechnologyVol. 11 No. 22 Construction and Implementation of Im- 2.1 Deciding of input and output vectorsproved Neural Network IdentificationThe nature factors affecting the technology andThe BP network, proposed by D E Rumelhart and Jthe recovery ratio are as follows average coal thickL Mcclelland et al in 1986, is a typical multi-playeress M, and top coal hardness factor F, parting andnetwork. Its topological structure is shownfracture efficient P, mining depth H, roof lithologyFiand thickness factor Q, inclination angle 6, and topIgcoal drawing height S. From technological requirements, the technology parameters are as followsdrawing sequence U, mining and drawing ratio Vdrawing interval W. Therefore, the input vectorsLatent hAveTfor a coal seam are: average thickness X,, miningdepth X2, top coal hardness factor X3, parting andfracture efficient X,, roof lithology and depth factorX,, inclination angle X, and the output vectorsig. 1 Typical triple-layer topologicalare: drawing sequence Y, mining and drawing rastructure of a BP networktio Y2, and drawing interval yIt consists of an input layer, a latent layer and 2.2 Numerical study stylebookan output layer. The fully-connected type is adoptSince study stylebooks are conceptual depic-ed between layers, and connections do not exist betions,only when a numerical mode is converted,atween the same layer unit. The character of the BP neural network can identify and deal with it. for annetwork isS function, its output values are only between 0 and1) The weight values between every layer con- 1. For parting and fracture efficient, roof lithologynections can be adjusted through studying.and thickness, drawing sequence, mining and2)Its dealing units, except for the input unit, drawing ratio, and drawing interval, they are dealtare with a nonlinear input-output relationshipwith respectively as Table3)Input and outputthe network can1) Parting and fracture. If the thickness ofge in series, in which output values are randparting is respectively more than 0. 5 m, betweennumbers between 0 and0.2 and 0.5 m, and less than 0.2 m, and fractures4)When an input mode is given, it transmits develop, the parting under the three conditions isfrom an input unit to a latent unit. From the latent regarded as classes 1, 2 and 3. And the correspondunit which has been deposed to an output unit layer ing weight values of 0-1/3. 1/3-2/3, and 2/3an output mode is produced after bI are given respectively for the three conditions.deposed by the output unit. If an error between th2)Roof lithology. For roof of grades I,Ioutput response and the expected output mode exⅢ,andⅣ, the values3/4~4/4,2/4~3/4,1/4sts and the requirement is unsatisfied. And the er2/4, and 0/4-1/4 are given respectivelyshifted, transmitted along3) Drawing sequence. There are four mainconnection access layer by layer, which rectifies the drawing sequences: multipleluence drawingconnection weight values between layers. For a giv- multiple interval drawing, single sequence drawingen group of practice modes, every mode is triedand forward propagation and back propagation are 3H中国煤化工their weight values areCNMHG2/4,0-1/4,resperepeated until the requirement is satisfied for every tivelyBased on successful field experiences, a tytechnology study stylebook is listed in Table 2.MENG XIan-rui et alResearch of Neural Network Identification modelTable 1 The disposing of some technology parametersCriteriongradeValue rangeParting thickness is larger than 0. 5 m: fracture develop-is not obion0~0.33Parting and fracture Parting thickness is between 0. 2-0. 5 m: fracture dev0.34~0.66Parting thickness is less than0.67~1.0Firsthting is unobvious, with a weighting span of lessthan 25 m: thickness of immediate roof is less than 10 m0.75~1.0First weighting is obvious, with a weighting span of 25-Roof lithology Xsa weighting span of 2550 m: thickness of immediate roof is less than 5 m0.25~0.49First weighting is very intensive, with a weighting span ofabove 50 m: thickness of immediate roof is larger than3 mMultiple sequence drawing0.76~1.00.51~0.75Single sequence drawing0,26~0,50Single interval drawing0~0.25Drawing ratio YDrawing height/ mining height1/5 of mining and drawing raDrawing interval yMovement distance of support for each drawing1/2 of drawing intervTable 2 The study stylebook of the technology selectionNumberCoal face’ s nameH/M/1 Yangquan 186052.60,320.61oMicun west No. 2(1)Wulan 311.50.2430Micun15036 Fangezhuang2180(4)175.961.210.510.62110.71.32Weijiadi Xj-11012.480.750.2520760.2Yizhouyao 8902Tanshan 2429N1.210.25140.780.410.53Zhuxianzhuang 8413-235012 Chaohua1 1051Nanshan 18260.880.822.3 Determination of the neural network struc- step length. The more the convergence factor nture and the study parametersthe more the rectification of weight value and the1) The design of the network structureaster the convergence velocity, but may be the conFrom the practical conditions, the unit num- cussion of system error will occur. On the contrarybers of input and output layers are 6 and 3 respec- if the velocity is slow, the study process will be sta-tively, and the model is the BP network which con- ble. Ordinarily the value of n is between 0 and 1tains a latent layer. By repeated testing and compar- Here, the initial value of n is 0. 5. Study parameter aing, the model topological structure is determined中国煤化工, smooth the curve andthe number of units in the latent layer is 18,iCNMHGonCuSSIon of system er-which the rectified SIGMOID function is chosen asHis equal to the filter wave of high-orthe excitation functionder components of an error cur2) The selection of convergence factor n and weight value space. The effect of the previoustudy paray数据weight value is added to the current weight valueConvergence facmeans study velocity or which utilizes the inertia of error rectificationJournal of China University of mining & TechnologyVol. 11 No. 2Therefore, in order to ensure the current rectifica- immediate roof thickness 5-6 m, which is sandition value as the main part, a should be less thanness shale and easily falls, main roof, sandstoneor equal to 7, here we suppose a=0thickness more than 5 m, coal faces length3)Determination of value convergence factor 5 150 m, mining height 2.6 m, drawing height 2. 23and error convergence factor Pm, and mining and drawing ratio is 1:0.86.Generally, the values of s and P are less thanThe disposed input vector are: X: M=4.830. 01 and 0. 05, respectively. Thus we take 5<0. , H=450, X3: F=1.05,X: P=0.42,X:Q005,B≤0.010. 56,Xs: 6=13. And the output vectors are:Y2.4 The practice course of the modified BP net- 0.27, Y=0. 16,Y3=0. 32. Analysis results shothat A. the optimum drawing sequence is single1) Give the initial weight value and the thresh- interval drawing in coal face 1121(3)in XieqiaoId value randomly. For the BP network, when the Mine, and the next is single sequence drawing; Binitial weight values are equal, the network can not the optimum mining and drawing ratio is 1:0.8study normallynamely when the mining height is 2. 68 m and the2) Forward compute the practical outputs of drawing height is 2. 15 m, the drawing effect isnodes in the output and latent layers layer by layerOpp=f, (netp) )=/,(wgOp-1+0,)In order to validate the result, different kindsf(x)=p/(1+e)+yof technology are used in the fields, the results are(u, d, y are random numberslisted in Table 3. According to the practical condlayer 3) Compute the error of nodes in the output tions of fully-mechanized top-coal caving face 1121(3)in XMine, the optimum technology in themine is single interval drawing, drawing interval isE=∑(Tn-O,)0. 6 m. When coal thickness >5.8 m, two-min-4) Backward compute theped signal of ing-and-one-drawing is reasonablenodes in the network laer by layerThis is in coincidence with the result of modelf'(netpi)(tpi-Ooi) (output layer)identification. If the industrial test is carried out. itf'(netpi)(>opWki)(latent layer)will take nearly three months and cost much person-al and material resources. However, if the neural5) Rectify the connection weight values andnetwork technologyolied, it will only takethe threshold values of nodes using the modified BPmore than thirty hours, the economic and technicalbenefits b△W(n+1)=70nO+a△W2(n)+△(n)Table 3 Comparison of recovery ratios ofdifferent kinds of technology6)Whether AW