STUDY ON THE COAL-ROCK INTER-FACE RECOGNITION METHOD BASED ON MULTI-SENSOR DATA FUSION TECHNIQUE

CHINESE JOURNAL OF MECHANICAL ENGINEERINGvol.16,No.3,2003321·STUDY ON THE COAL-ROCK INTER-FACE RECOGNITION METHODRen FangBASED ON MULTI-SENSOR DATAYang ZhaojianFUSION TECHNIQUEXiong shibTP8,Cz3月Research institute of Mechano-Abstract: The coal-rock interface recogmition method based on multi-sensor data fusion technique isTaiyuan University of Technology.forward because of the localization of single type sensor recognition method. The measuring theoryTaiyuan 030024, Chinabased on multi-sensor data fusion technique is analyzed, and hereby the test platform of recognitionsystem is manufactured. The advantage of data fusion with the fuzzy neural network(FNN)techniqhas been probed. The two-level FNN is constructed and data fusion is camied out. The experimentsshow that in various conditions the method can always acquire a much higher recognition rate thanKey words: Coal-rock interface recognition( CIR) Data fusion(DF) Multi-sensorsource information. The purpose of DF is0 INTRODUCTIONdescription of the object by calculating andthe existedmulti-source signals. Thus a more accuratethe ability to automatically trace the coak) can make the shearer have ment can be made. Multi-sensor may have different features: on-ciple is shown in Fig. 1. It cannot only contribute to mine automation, fuzzy or accurate information; compleary or contradict infor-but also reduce the content of rock and the other mineral that must be mation. Multi-sensor DF can fully take advantage of multi-sensorremoved in the process of coal beneficiation. In the recent years, USA sources, and then syntheses the complementary or redundant infor-and some other countries have developed CiR as an important project mation both in time and in space according to certain algorithms orbecause it is one of the key techniques of mining automation.principles, so identical description to detect object can be availableTop rock Vibration sensor CoalThese DFs can enlarge the range of time and space, enhance confi-dence bounds, reduce fuzzification, improve detecting property, andenhance system reliability greatly. In the multi-sensor system,theinformation given by each sensor is somewhat unreliable. DF proc-ess is in fact an unsurely inductive process. The recognition algo-theory, which constructs state space model of fusion object andestimate with the different estimate method, such as kalman filterFig 1 Detecting method of coal-rock interfacemaximum likelihood estimate, least square fitting. @2 Data fusionmethod based on statistic theory. It is based on estimate theoryAt present, methods worldwidely adopted in this field are as which realizes data fusion with the iteration algorithm, such asfollows:( Sensor method based on NGR(nature gamma radia- Bayesian method, D-S algorithm and markov random field methodtions)o.@ Sensor method based on cutting force response of 3 Artificial intelligence fusion method based on knowledge.Theshearer.@3 Infrared rays detecting device. Vibration sensor method is based on theory of rule-based model. The data fusion ismethod. These methods collect data by the single type sensor only. carried out when the two or some regulations about the same objectBecause every single sensor has its own special certain precision and are combined into a same regulation in the logic interfere process,application range, collecting data by the single type sensor is limited such as blackboard architecture. 4 Fusion method based on in-to a certain area and fault recognition is sometimes unavoidable. formation, the method is based on information processResearching and developing a recognition system using multi-sensor which is developed recent years and possesses the higheris a tendency to provide design base for the height adjustment system. gence such as neural network(NN), entropy method, clusterDue to the above consideration, a new idea has come with collecting sis, fuzzy logic method and vote methodgnals related to cutting status of shearer with multi-sensor, such asNN technology wins favor with its unique advantage amongmotor current signal, oil pressure signal, vibration signal of non rotary the methods. NN regards the known data and the known conditionnal moment and torsional vibration signals of drum shaft. modal as sample, gets the map relation between the data and theIn a word, fusing multi-signal to recognize coal rock interface can modal through studying. In the NN system, signal storage andhance precision and accuracy greatlycess merge into a whole. NN can induce the integrative inforation from the partial and imprecise information, because it has1 MULTI-SENSOR DATA FUSIONlitv. information Drocessing ability and perfectData fusion (DF)is a kind of processing method with multi- develop中国煤化工 The production andfoCNMHG Because the generlogic mode, the inputThis project is supported by Provincial Youth Science Foundation of Shanxi, or output of network belongs to either this kind or that kinds-Butna(No 20011020)and National Natural Science Foundation of China in practice, the classification boundary of data is usually fuzzy or(No. 59975064). Received June2003: accepted April 10, 2003sometimes wan. So if the information is not kept in the network322. Ren Fang, et al: Study on the coal-rock interface recognition method based ondata fusion techniqueduring the training, the recognition ratio of network cannot beTo every xian be determined according to thered. In order to solve the problem, the author is enlightened formula u f(oxhich f= sigmoid(r). a and 6 can beby the fuzzy technique, and combines fuzzy technique with the determined whenInction a is minimized, given(x, u,),nn to identify stateuzzy technique can be combined with nN due to many is1.2....Nimilarities0l.( From the mapping view they posses non-liE=∑(1-)/Nfunction nearly ability. As for the style of data process theyall settle the process parallelly. In the aspect of the inside where Ar--Desire outputmechanism there are many similarities between the membershipu-Current outputfunction of fuzzy theory and output property of nn, between theDetermining o and 6 according to above is equal to deter-max-min operation in the likelihood inference of fuzzy logic and mining the relevant membership function. The values are no lon-be combined with NN and further their application field y canweighted input, operation in every cell. So fuzzy technolo2.2 Design of non-fuzzication2 DF BASED ON FUZZY NEURAL NETWORKA common nn doesn't contain non-fuzzicatlayer is joined to output layer directly, The number of outp(FNN)nodes is equal to the number of classified patterns. while trainiof fNN is shown in Fig. 2. make the output of the relevant pattern be I and the others be O,is the addition of fuzzica- and take the largest output as classification result. But the dataclassification boundary of coal-rock interface is fuzzy, the datasometimes are wan, the trained patten has nonzero value on moreFutzication Hidden Non-fuzzication Outputthan one output. So in the training moment, network is wanted toayerstudy the relationship between input and output.If there are L pattens, accordingly there are L output nodesror of the ith result of samples respectively, so the distanc 6a.Definition 1: 0, and y denote mean value and standard er-tween the training sample and the result is defined ask=1,2,…,Lwhere F-The pth element value of the jth pattern-kkDicast square error is bigger, the weight value is smaller in classi-Definition 2: The membership function of the jth pattern tothe ith classification isFig2 Topology structure of FNN2.1 Design of fuzzication layerThe aim of the layer is to preprocess the data, that is to say, tomake the input value of input cell standardized and then outputFor the CiR question, the paper modifies the above algorithmthe standard value determined by membership function in order to In the algorithm, if the distance between the sample and the centermake it adapt to post process. To fuzz input, there are two ques- of this classification is far, the u, value is smaller. So it cannot fitions that must be settled: One is to determine fuzzy level, and for the samples whose value is bigger than the mean value of rockanother is to determine membership functionor is smaller than the mean value of coal. In order to make thealgorithm fit, order F which is bigger than rock mean value equby the sensors consist of two states, so every input value can be to rock mean and fi, which is smaller than coal mean value equaldivided into two levels. The realizing process of membership to coal mean valuebelongs to the membership degree of coal and An that belongs to 3 MULTI-SENSOR RESEARCH OF CUTTINGthe membership degree of rock, so the process of xp Ha,x* 2 STATEcan be obtained by the nn as Fig 3 show3. 1 Multi-sensor measuring principle of shearer cutting stWhile cutting, the coal cutting resistance will be differentfrom rock cutting because of the different hardness. Thus it makesshearer respond differently to different cutting state. According tothe observation and summarization of industrial test, the changesSuppose the membership function is shown as Fig 4of cutting state can be shown in the following aspects: 0 oilpressure;② arm vibration;,③ motor cutting current;④trinal moment and torsional vibration of drum shaIn an existed system, while it normally runs, the transferfunction keistic and phase-frequency characteristic are fixed accordingly. Toency characteristic. it mainly plays a role of中国煤化工 ith different frequencychanges, the energy ofMembership functionCN MH Gecause the system canhere x1—Me of 100% coal featurex3-Minimal value of 100% rock featureshearer can be judged. So the coal rock interface recognitionmeasuring system of a model shearer is set up with the pressureCHINESE JOURNAL OF MECHANICAL ENGINEERING323sensor,dimension vibration sensor, current sensor, torsionalThe hardware principle figure of the measuring system ismoment and torsional vibration sensorshown in Fig. 5.Torsional vibration sensInstrumentCurrent sensoFig 5 Hardware principle figure of measuring systemAfter setting up the system, thesimulation test, collects the data of coaltput IctiExperiments parameter: drum rotate sper/min, draggingspeed 1 m/min, cutting width 25 mmCurrentheight 8 mm, 10n 15 mm.Torsional vibration3.2 Test results analysisTo the original signals, feature is extracted by waveletpacket technique discussed in another paper. It completes theTorsional monmenttransform from pattern space to feature space, filtrates thelements which cannot open out sample essence, and com-presses the dimensions of feature so as to acquire the featurePressure sensorvector which can clearly open out the samples. After completing the above work, vibration signal feature vector is 12 dipensions, current signal feature vector is 2 dimensions, theother feature vector is 1 dimension respectively. So first thefeature fusion is carried out with the FNN whose training data two-level fusion model makes the system robust, and when there vibration signal data and current signal data respectively. certain bad or failure sensor provides incorrect or contradict data,hen the fusion is carried out with the FNN whose input data the complementary or redundant data acquired by sensors makeare the former FNN outputs, pressure data, torsional moment the system more reliabledata, torsional vibration data. So the process model decided isThe recognition result is shown in Tables 1, 2. From the tablesdistributed multi-sensor structure and two-level fusion model. that the recognition ratio can be increased after fuzzing and theThe fusing principle is shown in Fig. 6recognition ratio with the two-level fusion can be increased greatlyTable 1 Result comparisorCorrect 0.8717coa088270.1267006170.89710.1086Correct 0,863 8Coal0.75770.2431orrect 0.92150035187730.775502336Correct 0.943 0Correct 0.9225CorrectCoal09104Correct 0.892 80.0634Correct0.94140.069Correcta0.5316FaultCorrectcoal0.84030.1640 Correct0.882800284Correct0.96910.04060.909800971Correct0.15080982000361Correct 09172003960.1351I1oalFault 0.7967 0. 1986 Faultcoa0.553104708 Fault0.681303032Faut0933700771 Correct0.1542Correct 0.8157Correct 0.913 4 0.0965 Correctock0.023109847Correct 0.165308269Correct 0.084Correct150.0960Correct0071709320Correctock0.105208959orrect 0.178 80.811500735009940.8676Correct 0.0855ock0.34660.6672fault0.71890,2016ock002310.9847Correct 0. 40.1785Comect 0 1266 0中国煤化工980com02755CNMHG9213 Correct0.1374Correct 0.028 0orrect 0.0768 0.9273 CorrectRok0.04150.9707 Corect0,2050.8710090009145Rock0.010310114coct0.10640,8933orrect 0.088 60.9160324. Ren Fang, et al: Study on the coal-rock interface recognition method based on multi-sensor data fusion techniqueSampleStatetput 2 Retput 1 Output 2 Result000370.917512698683 Correct0.1536Correct 0.0840.920975%coa1.264302846 Fault0.40590.5951Fault0.70350.3146Correct50%coa154215247Fault06982Correct0.62880369175%coa0.080009125Fault0.73490.2567Correct0.72530.2734Correct25% rock25%cal0.02170.9789Faut0.269707638Correct0.30360.5880Correct25%coal0.07890.919000269l.01890.13250.842350%coal0.72900.2709Fault 0.213 205331Fault004820.8540 Fault50% rock2廉自生基于采煤机截割力响应的煤岩界面识别技术研究:[博士学位able 2 Comparison between NN and FNN论文]北京:中国矿业大学,19953许永江无源红外线煤岩界面探测系统煤炭技术,1994,134:10FNN4徐瑛国外煤岩界面传慼器开发动态综述.煤矿自动化,195162)vibration152 Test samples5聂伟荣.基于改进BP网络的地震动信号目标识别,南京理工大学学121Corres130报(自然科学版),2004202341raining samples 236 Training samples 2366 Looney G. Pattern recognition using neural networks. New York: OxCurrent152 Test samplesford University Press, 1997Correct120 Corn7 PoyuTson. Structural damage detection and identificationetworks. AIAA Joumal, 1994, 32(1): 176-186Training samples 2368 CherianR P. A neural network approach for selection of powder metallur-Two-levelgy materials and process parameters. Artificial Intelligence in Engineering,Correct9 Raman H, Sunilkumar N Multivariate modeling of water resources time4 CONCLUSIONSes using artificial neural network. Hydrological Sciences JourmzBased on multi-sensor DF10 Kuo R J, Cohen P H Intelligpresented in the paper. Analyzingartificial neural network and fuzzy modeling. Artificial Intelligence in En-fuzzy technique and NN, andof combination of gineering, 1998, 12(3): 229-242level FNN fusing system is setognition system adopted FNN technique with the two-level strucIts show that rec- Biographical notes: Ren Fang received her master degree from Taiyuan Univ-ture can carry out the CIR and the system adopted the modified ersity of Technology, Shanxi, China, in 1998. Her research interests includealgorithm has the higher recognition ratio. It is both theoretically measurement and fault diagnosis.andpracticallyimportanttotheCIRandtheautomationoftheTel:+86-351-6014551:E-mail:tmmpieco@public.tysx.cnnearer horizon control system.Yan Zhaojian, bom in 1955, is a professor in Taiyuan University of TechnologyReferences1秦剑秋,自然γ射线煤岩界面识别传感器的理论建模及实验验证煤炭Xiong Shibo, bom in 1938, is a professor in Taiyuan University of Technology学报,1996,21(5):513~516中国煤化工CNMHG