Available online at w. sciencedirect coMININGScience DirectSCIENCE ANDTECHNOLOGYELSEVIERMining Science and Technology 19(2009)0199-0204www.elsevier.com/locateljcumtImproved ultrasonic differentiation model for structuralcoal types based on neural networkTIAN Zi-jian, WANG Fu-zhong2, LI Tao, BAI Shan-shanSchool of Electromechanica and information Engineering, China University of Mining Technology, Beijing 100083, ChinaSchool of Electrical Engineering Automation, Henan Polytechnic University, Jiaozuo, Henan 454000, ChinaAbstract: In order to solve the difficulty of detailed recognition of subdivisions of structural coal types, a differentiation model thatcombines BP neural network with an ultrasonic reflection method is proposed. Structural coal types are recognized based on a suible consideration of ultrasonic speed, an ultrasonic attenuation coefficient, characteristics of ultrasonic transmission and otherparameters relating to structural coal types. We have focused on a computational model of ultrasonic speed, attenuation coefficientin coal and differentiation algorithm of structural coal types based on a BP neural network Experiments demonstrate that the modelcan distinguish structural coal types effectively. It is important for the improved ultrasonic differentiation model to predict coal andgas outburstsKeywords: ultrasonic; structural coal types; BP neural network; coal ultrasonic attenuation coefficient; coal ultrasonic speed1 Introductionacteristics, rules of ultrasonic transmission, discrep-ancies of ultrasonic wave speeds and attenuation coStudies have shown that fault spots of coal and gas efficients of different coal structures, a detailed difoutbursts mostly exist in coal seams or coal beds, ferentiation model for structural coal types, based onmade up of a particular kind of deep tectonic coal. the combination of a neural network and an ultrasonicTherefore, effective recognition methods about method has been proposed by us in order to differen-structural coal types should be studied at length. Coal tiate exactly among structural coal typesand gas outburst disasters would be predicted effec-tively and in time by recognizing precisely the struc- 2 Differentiation model of structural coaltural coal types. Currently, there are many studies onlarge-scale detection of tectonic coal, but hardly any When ultrasonic waves are diffused in coal, theirresearch has been carried out in the field of the small speed and attenuation coefficient vary with coalefront areas of the working face. Accurate and detailed structures. Given the requirements of a differentiationmeans of detection are lacking in the small areas of model for structural coal types, a three-layered dif-he working face. What is more, conventional detec- ferentiation model, based on a BP neural network, hastion operations at the working place are too extensive been established by simulation with due considera-to be carried out in this confined space. So it is diffi-ion of the characteristics of ultrasonic transmissioncult to differentiate between the types of structural in coal. The model is shown in Fig.1coal at the refined level required. Up until now, thereare no reports about the recognition of subdivisionswithin structural coal types. Domestic and foreignscientists are largely engaged in the field of structuralcoal detection methods, such as geological radar, racoefficientCoal typesdio waves and Hertzian waves. Detailed recognitionof subdivisions within structural coal type is difficult,Ultrasonic frequencyespecially when differentiating overlapping parts indifferent types of structural coal, where wrong deter-minations can easily occur Given transmission char什H中國煤化工 n model of coal typesCNMHGReceived 12 August 2008: accepted 30 October 2008CorrespondingauthorTel:+86-13269565519;E-mailaddresstianzj0126@126.comVol 19 No. 2This model is made up of a BP neural network, direction; I, is the back and forth time of ultrasonicathematical formulas to calculate ultrasonic speed speed between the surface of the medium and thin coal vo, a coal attenuation coefficient a and an ul reflection surface of the nth layer in the vertical direc-trasonic frequency f, which are accepted as input parameters. In a bP neural network differentiationmodel of coal types, the only output layer of the2.2 Formula of the attenuation coefficient in coalmodel is the coal type. It works on the followingStudies have revealed that for multilayered mediaprinciple: information regarding ultrasonic speed in such as coal seams, the attenuation coefficient of eachcoal Vp, together with the coal attenuation coefficient layer can be calculated with Eq- (2).a and the ultrasonic frequency f, is fed into the BPneural network for synthetically analyzing and proc-2.3.bsing, so that an estimate of the structural coal type20△scan be made. When making use of the analyticalwhere Asn is the distance margin in the nth layer andfunctions of this BP network, misjudgment due to Ab, is the amplitude margin in the nth layer.overlapping parts of various coal structures could beeffectively avoided, thus improving the accuracy of 2.3 Ultrasonic detection frequency fthe differentiationTo increase the accuracy of network learming, frequency and coal structure. In the same coal strucproblems such as local minima, robustness, low er- ture, the higher the frequency, the faster the ultrasonicror-tolerance and overtime learning should bespeed attenuates and the greater the attenuation coef-avoided. Therefore, both stepwise prolonging and ficient becomes: in contrast, the lower the frequency,stepwise trimming methods are adopted to decidethe number of hidden layer nodes. Given the relathe slower ultrasonic speed attenuates and the smallerthe attenuation coefficient becomes[).Thereforetionship: j=vm+n+!(where j is the number of the ultrasonic frequency is mainly used to compen-hidden layer nerve cells, m the number of input layer sate for the effect of the ultrasonic attenuation coeffinerve cells, n the number of output layer nerve cells cient in coaland I a constant(ranging from I to 10), the initialnumber of hidden layer nodes is set as four. Next, we2.4 Algorithm of BP neural network differentiincreased or pruned stepwise the number of nodes ontion modeleach layer by experimentation. The experiments in-We defined the transmission function of each nodedicated the network would not be convergent if nodeswere too few, while too many nodes would cause theas the S function: f(r)increase of the training time and the decWhen a BP network adjusts its connection weightfault-tolerant capability by reason of the hidden layer and nodal threshold, a typical error correction metho2.1 Formula of ultrasonic speed vp in coalW=W4+7(b2-W·x4)·X4In response to the request of trenchless detection,we chose an ultrasonic reflection method. On account△W4=7a4·Xkof the adoption of a trenchless detection technology,it is impossible to know the depth of the reflection where 5=(b, -wr.X, is the difference betweeninterface and difficult to calculate the ultrasonic the expected output b and the actual network output nspeed. In order to overcome this difficultyhe is called the learning factormethod of a travel time-distance curve is used in theThe learning process of a BP Network is composedprocess of calculating ultrasonic speed. A coal seam of positive spreading and reverse spreading. Duringcan be regarded as a kind of multilayered medium, the process of positive spreading, data from the inputwhere the actual ultrasonic speed vn in each layer is as layer are sent to the output layer after being processedfollowsvia the hidden layer. The output of each layer only同∑-過affects the state of the next layer. The output data arecompared with the expectation at the output layer; if1) there are differences between them, the differencesnal nath. The differenceswhere v, is the weighted average square root of ul- strenj中國煤化工 ging the linkingtil the differencestrasonic speed in the medium; vi(i=l, 2, 3,..., n) is areCNMHlolerance providedrasonic speedin the ith layer; t; is the one-way in advance.time of ultrasonic speeding each layer in the verticalTIANImproved ultrasonic differentiation model for structural coal types3 Experimentalformulas. Secondly, a comparison between the calcu-lated value and the measured value was made for theTo verify the accuracy of the ultrasonic differentia- purpose of validating the suitability of the formulas.tion model for structural coal types based on a neural Finally, on the basis of calculating the wave speed vpnetwork, coal structure was divided into type I(unre- and the attenuation coefficient a, the BP neural net(smashed coal) and type I(mylonitic coal). These ing and check samples g these results as the learnformed coal ), type I(fracture of coal ), type Il work was simulated usinstructural coal samples were bituminous coal from 3.1 Experiments of ultrasonic speed in multilaythe Pingdingshan coal mining area, anthracite coal ofered coalAnyang, from the Jiaozuo mining area and so on. TheUltrasonic velocity and coal ultrasonic attenuationUltrasonic speeds were obtained using a transmis-coefficient were measured, and then the records and sion method via the tests of coal samples such as typeclassification were accomplished. The actual super- I and type IV collected from coal mines. The datasonic speeds and attenuation coefficients were ob- are shown as comparative data of the reflection extained and we kept recording each category of coal periment in Table 1. The ultrasonic speeds weresamples. The coal composition of the actual mine was tested in the multi-media coal samples which simusimulated, after which experimental samples were lated the actual ingredients of the coal mines. Theobtained. In order to improve efficiency the most data are shown in Table 2. There was only one grouprepresentative type I and type IV were selected. of data in the lst layer(coal type I ) so the weightedDuring the process of obtaining samples, type I coal average square root of the speed equaled the averagesamples were placed on the surface of type IV sam- ultrasonic speed, i.e., Vi=2ples. In order to simulate the structural characteristics weighted average square root speeds under differentof coal in underground mines, the two different types angles of incidence in the 2nd and 3rd layer(coalof coal samples were bound compactly by a thin layer type IV)were, respectively: V,=1590. 25 m/s andof glycerin on the central contact surface. The coal v=1402.9 m/s. The ultrasonic speed in the 1st andsamples were dried and the experiments were carriedout at room temperature under atmospheric condi- 2nd layer were calculated with the ultrasonic speedtionsformula and were, respectively: V2=669.8 m/s andDuring the experiments, supersonic signals wereV3=659 m/s. Comparing these data with the ultrasonprojected at5°,10°,15°and20° angles of incidencspeed data in Table l, we find that the results shorto the experimental coal samples, by adjusting the that the error is within the allowable range of engiangle of the transmission probe, while the receivingneering precision for the purpose of detecting theprobes were used on the surface of the top layer to types of coal, which indicates that the calculationreceive the reflecting signals sequentially. Theformula of ultrasonic speed is effective and correct.rameters from our experiments are shown in Tables 1Table 1 Data of ultrasonic speed experiments onto 4. The supersonic sending and receiving devicescoal samplesesigned by us. The Dual-trace oscillo-Coal SamplesTravelYB4320A. was used for observationstypes length L(cm) time t(us) speed V(m/s)Afterwards these data were entered into the ultra-2322155472analysis. At first, the ultrasonic speeds and attenua-tion coefficients were obtained from their respectiveTable 2 Data of ultrasonic speed model experiments in coalAngles of incidence e()Data types15Travel time t (um)3726376.93843Probes distance x(cm)Weighted average squarey.(m/s)2143.9215L.121475Travel time t(μm)750.l754.77628Test data in the15.22nd layerProbes distance x(cm)Weighted average square root speed y(m/s)中國煤化工5872Travel time r (um)CNMHG1038.73rd layerProbes distance x(cm)Weighted average square root speed v.(m/s)1403.7l4003.2 Experiment of ultrasonic attenuation coeffTable 3 Ultrasonic attenuation coefficient a in differentcient in coalcoal types under 20 kHz.The attenuation coefficients in Table 3 were ob-ained by testing coal samples of types I and IV. The CoalHorizontal Vertical Attenuationdifference△r(m)experiments showed that the ultrasonic attenuationcoefficient varied only with the coal structure. So it86was illustrated with the data of 20 incidence angle.IV coalThe parameters received at the same ultrasonic fre-quency and under 20 angle of incidence were en-The results in Table 3 were compared with those intered into the formula for the ultrasonic attenuatioTable 4. It is clear that the data are basically concoefficient in coal. The attenuation coefficient ob-tent with each othertained is shown in Table 4Table 4 Data of ultrasonic attenuation coefficient model experiments in coalMedia objectTravel difference inReceived signal amplitudeAttenuation coefficient inevery layer△s(cm)difference△B1dB)every layer△b(dB)every layer a(dB/m)Ist media(I coal02395nd media(V coal)9.l573rd media(IV coal)3.3 Simulation experiment of differentiation and a were obtained by the experimental methods,model based on BP neural networkintroduced in the sections 3. I and 3. 2. Those experiOn the basis of the wave velocity vp and the at- mental sample groups were assigned as training sam-tenuation coefficient a experiments, simulation exples of the BP neural network. Similarly, we selectedperiments of training and checking of the BP neural 10 coal samples at random to form four groups ofnetwork were carried through In order to make deci- coal samples in the same way as illustrated earlier insions correctly and promptly during the training oforder to obtain the wave velocity vp and the attenua-the BP network, 19 coal samples from different coaltion coefficient a. These test data will be regarded asmines were divided into eight groups of coal samples check samples of the BP network. Both the learningaccording to their real coal layer structure. Then vpsamples and check samples are shown in Table 5Table 5 BP neural network leaming and check samplLayer numberYYCoal typesa(dB/m)f(Hz)02520000384Ⅳ2200003.25215125000906eanngl5781.5671122N10876ⅣI“ Layer1.362中國煤化工117233.134YHSCNMHG33.65TIAN Zi-jian et alImproved ultrasonic differentiation model for structural coal types2152902395409200003 layer8.972200004.21721630324.153Check08452017I"layer24.75695100000383lⅣThe parameters must be normalized according to ,A coal seam is equivalent to a multilayered me-their characteristics, while the bp neural network is in dium and the method of a traveltime-distance curve isthe process of training and learning. The learning rate adopted to develop the formulas of speed and the at-was an important factor while the model was being tenuation coefficient for ultrasonic transmission inestablished. For our purpose, the algorithm of our BP coal. There are many factors involved in this trans-neural network was a gradient descent algorithm. It' mission, such as the weighted average square root ofshowed that too low a learning rate contributed to a ultrasonic speed in the medium, the ultrasonic speedvery slow rate of convergence and too high a learning in each layer, the one-way time of ultrasonic in eachrate resulted in a violent surge of the iterative solution. layer in the vertical direction, the back and forth timeTherefore, according to the actual conditions, the of the ultrasonic transmission between the surface oflearning rate should normally be set at 0.3. At the the medium and the reflection surface of the nth layersame time, the minimum mean square error is set at in the vertical direction, the distance margin in the nth0.001, the impulse coefficient at 0.8 and the maxi- layer and the amplitude margin in the nth layer. Ex-mum number of training set at 15000. After entering periments show that the speed and attenuation coeffithe learning samples, the network converged after cient can be calculated for ultrasonic transmission in5960 iterationsThe four groups of examination samples were used In order to ensure that the network makes judg-to verify the accuracy of our decision. The neural ments correctly and promptly, morenetwork output Y and the actual structural coal types samples from different coal mines were needed dur-are shown in Table 5. It is clear that the output of the ing the training of the BP network. These coal sam-neural network is consistent with the actual condi- ples are arranged according to a real coal layer structions of the mines. We conclude that the differentiaThe sampleist be normalizedtion model, based on the BP neural network, is valid their characteristics. During the process of networkand the results are credible. The structural coal types training and learning, the choice of learning rate,could be identified with reasonable accuracy.minimum mean square error and impulse coefficientshould be taken into account4 ConclusionsAcknowledgementsAccording to the characteristics of ultrasonictransmission in coal, we have presented a differentiaThe present research work has been supported bytion model for coal structures based on a bp neural the National Natural Science Foundation of Chinanetwork and an ultrasonic reflection method. For the (No 50674093)and the doctoral Foundation of thepurpose of judging the type of coal structure, the Chinese Education Ministry(20050290010), the au-model makes use of this BP neural network to ana- thors gratefully acknowledge the support of theselyze the ultrasonic parameters of coal samples such as institutionspeed vp, attenuation coefficient a and frequency f.The model effectively decreases the incidence of er-Referencesror judgments, which often occur at overlapping areasof various types about coal; the accuracy of judgment [1] Guo D Y, Han D X Stick-slip mechanism of coal andhas been improved. The validity of the proposed中國煤化工l Sociery, 2003, 28(6model has been confirmed with experiments, whichdemonstrate that the model is capable of judging coal [2CNMHGSof" coal seams instructures. It is important for the improved ultrasonicgliding structural areas. Journal of China Universiry ofdifferentiation model to forecast coal and gas outburstMining Technology, 2007, 17 (4): 484488disasters[3] Mohamed MM, Shaat L T, MaA N. PropagationMining Science and TechnologVoL 19. No.2of ultrasonic waves through demineralized cancellous(4):236-240.( n Chinese)bone. IEEE Transactions on Ultrasonics, Ferroelectrics, [9] Clement G T, Hynynen, K. Forward planarand Frequency ControL, 2003, 50(3): 279-288through layered media. IEEE Transactions on[4] Washer G Fuchs P, Graybeal B A, Hartmann J L. Ultra-ics, Ferroelectrics, and Frequency ControL, 20如onic testing of reactive powder concrete. IEEE Trans-689-1698actions on Ultrasonics, Ferroelectrics and Frequency [10] Bossy e, Talmant M, Defontaine M, Patat F, Laugier P.Control,2004,5l(2):193-201Bidirectional axial transmission can improve accuracy[5] Haunt M A, Jones D L, O'Brien W D. Adaptive focusingand precision of ultrasonic velocity measurement in cor-through layered media using the geophysical"time mitical bone: a validation on test materials. IEEE Transac.gration"concept. IEEE Ultrasonics Symposm, 2002(2)tions on Ultrasonics, Ferroelectrics, and frequency635-1638control,2004,5l(1):71-796] Williams R, Cherin E, Tavakkoli J, Foster FS, Tavakkoli [11] Guo Y S, Meng Z P, Yang R Z, Zhang L H, Liu Y C, SunJ. TheorX Y, Zhao G P. Seismic attributions analysis and it's ap-ear ultrasound propagation through multilayered media.IEEE International Ultrasonic, Ferroelectrics, and FreChina Universiry of Mining& Technology, 2004, 33(5)uency Control Joint 50th Anniversary Conference, 2004557-562.(In Chinese)(3):2193-2196[12] Wu M, Zhang H Y, Sun Z, Liu X. Application of wavelet[7 LiQ Z. The Precise Path toward Exploration. Beijing:analysis and artificial neural network pattern recognitionOil Industry Publishing House, 1993. (In Chineseto flaw classification in ultrasonic testing. Journal of[8] Hui L, Tao L, Wang F Z Study of ultrasonic reflectChina University of Mining Technology, 2000, 29(3):method system on distinguishing coal structure239-243. (In Chinese)Progress in Mining Science and Safery Technolog中國煤化工CNMHG