Research on Detection Device for Broken Wires of Coal Mine-Hoist Cable

Joumal of China University of Mining TechnologyVoL 17 No. 3Availableonlineatwww.sciencedirect.com8CIENCEDIRECT·J China Univ Mining& Technol 2007, 17(3): 0376-0381Research on detection Device forBroken wires of coal mine-Hoist CableWANG Hong-yao, HUA Gang, TIAN JieSchool of Information and Electrical Engineering. China University of Mining Technology, Xuzhou, Jiangsu 221008, ChinSchool of Mechanical Electronic and information Engineering, China University of Mining Technology, Beijing 100083, ChinaAbstract: In order to overcome the flaws of present domestic devices for detecting faulty wires such as low precisionlow sensitivity and instability, a new instrument for detecting and processing the signal of flux leakage caused by brken wires of coal mine-hoist cables is investigated. The principle of strong magnetic detection was adopted in thequipment. wires were magnetized by a pre-magnetic head to reach magnetization saturation. Our special feature is thatne number of flux-gates installed along the circle direction on the wall of sensors is twice as large as the number ofrands in the wire cable. Neighboring components are connected in series and the interference on the surface of theire cable, produced by leakage from the flux field of the wire strands, is efficiently filtered. The sampled signal sequence produced by broken wires, which is characterized by a three-dimensional distribution of the flux-leakage fieldon the surface of the wire cable, can be dimensionally condensed and characteristically extracted. A model of a BP neu-ral network is built and the algorithm of the BP neural network is then used to identify the number of broken wirequantitatively In our research, we used a 6x37+FC, 24 mm wire cabletest object. Randomly severalwere artificially broken and damaged to different degrees. The experiments were carried out 100 times to obtain data for100 groups from our samples. The data were then entered into the BP neural network and trained. The network was thenused to identify a totalbroken at five different locations. The test data proves that our new device can enhancethe precision in detectingKey words: wire cable;wire; signal processing; detection deviceCLC number: TB 421 Introductionfor this paper. With the special structure of a detectiontransducer, the interfering signal from the leakageIt is well-known that coal mine-hoist cables are an field of wire twists can be filtered efficiently. Afterimportant part in coal mine-hoists or transportation the extraction of dimensional contraction and charac-systems. Wires are, in fact, subjected to breakage due teristic values of multi-ways signals, a quantitativeto wear, corrosion and fatigue. The extent of damage BP neural network recognition for broken wires inand the carrying capacity of wires are directly related steel cables was realized. The test results are pre-to the safety of equipment and staff. At present, there sentedare many detection devices for broken steel cablesmanufactured in China, but most devices do not meet 2 Basic Structural Principle of the On-Linethe conditions ideally required in practice. The rea-Detection Instrument for Coal Mine.sons are largely the complex structure of wires, badHoist Cableworking conditions, the multiplicity and uncertaintyof broken wires. It is therefore quite difficult to detectThe structural principle of the on-line detection de-signs of broken wires as well as to analyze and proc.cables studied bv us is shown in Fig. 1ess detected signal of broken wires in cablesThe中国煤化工 ed of two semicir-A new instrument for broken wires detection andprocession of coal mine-hoist cables was investigated closCNMHGan be opened orthorTel+86-516-83885993;e-Mailaddresshongyaowang2004@163.comWANG Hong-yao et alResearch on Detection Device for Broken Wires of Coal Mine- Hoist Cablemade of a single magnetic core and is single-windingAfter Ce, a,Ps,Ap are assured, Be is a constant.Some magnetic sensing units are evenly arrangeAfter the wire cables are deeplyaround the inner wall of the transducer, the numbernumerical value of u is very small. As a result, thehich is twice as many as the number of the wire value of z is larger and there is no need to mag-strands in the inspected cable. As well, two neighbor-ing units are connected in series to a detection chan-nify and process the detection signal again.When the sensor is operating along wire cables at anel. Consequently, the number of detection channelsspecified speed, the signals detected by each of theof the detection instrument is equal to the number of magnetic fluxgate units can effectively show thewire strands in the cablethree-dimensional distribution status of magnetic fluxleakage, generated at the surface of wire cables2-43 Filtration of the Wavelike Oscillation Interference Signal Plroduceby CableWire Twistsetecting transducerI#detecting channelwire ropeThe signal of broken wires from wire cables ob-:#detecting channeL-Signal iained by a single fluxgate detection unit of the trans-ducer(formula (1)contains all kinds of interferingsignals. The effect of the wavelike oscillation magakage B due to the speciBP neural network recognitionthe steel cables is largest, which directly affects theOutput of the testing resultsdetection of broken or damaged wires, especially incoal mine-hoist cables. We should consider the possiFig. 1 Structural principle of detection instrument forbility of filtering the interference signalsbroken wires in coal mine-hoist cablesIn formula(1), the interference signal Zr. causedAfter being filtered and reshaped, the detection sig-by a wavelike oscillation shows up as periodic varianal from each channel is sent to the signal processingtion. This kind of wavelike oscillation interferenceunit. The analog detection signal is converted into a signal can be regarded approximately as a sine wave,discrete dimensional sequence of sampling values by as shown in Fig. 2multi-channel A/D conversion, followed by a charac-teristic extraction, a BP neural network recognitionand the output of the result.When viewed separately, the leakage field signaldetected by each single fluxgate unit is the leakagefield intensity in the steel cable where the core-sponding fluxgate units are located. That is, the output signal ze, of any jth test unit is:Fig 2 Wavelike oscillation interference signal producedby the cable twistB.:+BAs PDOver the length direction of wire cables, its varia月(B+B)=x+乙tion period T is a Lay length of cable wire strands. Atthe circle direction of the wire cable, its variation pewhere Cp is the structural parameter of the fluxgate, riod is the reciprocal of the number of outer wirea the width of the drive square-wave, A, the satu- strands of the circle length of the wire cable. There-rated magneto-conductivity rate, B, the magnetic fore, the wavelike oscillation interference signal zr. jinduction intensity of the leakage field produced by of the jth detection channel can be expressed asbroken wires, B, the magnetic induction intensityof the leakage field produced by wire cable twists,Zr. j=a+mcosz .i the signal value of broken wires andvalue of the interference signal produced by wire ca-中国煤化工ble twistsCNMHGof theB==Sinawavelike oscillation signal, rm the Alternating Cur-rent Component magnitude of the wavelike oscillaJoumal of China University of Mining TechnolVol 17 No. 3tion signal, T represents the value of periods, yz/=(x(1),x(2),…x(K)the position of the detection unit, starting from theinitial spot, the initial phase of the wavelike osThe N-channel signal sequence will make up acillation signal, N the number of wire strands of the N-dimensional series vector group of broken wiresteel cable andn is the number of detection units, signalsObviously when N =2N, i.e., when the number ofZ=(x1,z,…,z)detection units doubles the number of outer strands of At this moment, Z is a Nx K characteristic matrixthe wire cable, the wavelike oscillation signal con-tained in the leakage magnetic field signal inspected of broken wires and it contains all the information ony any two neighboring detection units is in a revere status of the broken wiressal phase. Therefore, when the neighboring detectionGiven the analysis of repeated experiments, theunits along the inner wall of the cylinder of then the surface of wire cables created by broken wireschannel in series two by two, it is equivalent to add- is not larger than 20 mm. When the speed of the in-ing the (+l)th test channel signal to thechannel signal. Thus the strand peak value of the is 1.2 mm, the number of samples k is 16 at mostwavelike oscillation signal compensates for the strand When the number of inspection channels is N=4, Zvalue for the moment. That is. at this moment. the should be a 4x16 matrix. If the analysis of the char-only remaining wavelike oscillation signal is the Diacteristic matrix of broken or damaged wires Z wererect Current Componentdirectly carried out, the analytical process would bex=x+x1,+=22very complex and would need to be carried out ascomparison and judgment of the sequential value oft this moment, the magnetic field signal of leak- each line. So instead, we carried out a reduction in theage from any of the inspection channels made up of order processing of formula(6), i.e., we carried out athe fluxgate array should be:dimensional contraction. According to a lemma oftheoretical linear algebra Z can also be expressedx=2+乙,+=A(B+B)+乙=B,+x=+(4)2=(,4)((0,如(),…NM)=thZ, of this formula can be eliminated when the zerowheret,,t2,"", Iw are arbitrary, independent basedetection position is adjusted. Therefore we consid- vectors. h is the characteristic vector of one-dimen-ered that the wavelike oscillation interference signal sional broken wires expected to be obtained afterof cable wires is filtered by formula(4). After this mensional contraction. So long as the appropriate t ispretreatment, each leakage from broken wires, shown found, h can be derivedby magnetic field signals from the transducer, be-comes a channel sample value by A/D conversion, ash)=tz1=zrt1(=-,2,…,N)(8)shown in Fig 3According to the L-k transformation principle,when the value of t is the latent vector of the covari-ance matrix P. of Z, the transformation error is aminimum,i.e, t satisfies the characteristic equation2#channel3#channel(P2-4)1=0(=1,4channelwhere a is the characteristic value of p and I isx---Circle direction of wire ropean identity matrix. Represented by formula( 8), they-Length direction of wire ropeexpected characteristic vector h of the broken wirescould be obtained via the dimensional contractionFig 3 Multi-channel sampling value of broken wireThe process of transformation of the dimensionalsignals from wire cablescontraction is, in fact, a conversion from a N-dimensional characteristic vector to a one-dimensional vec.Extraction of Characteristic Value oftorSignals from Broken wiresThe average of the one-dimensional h sequence isregarded as an eigenvector which represents eachAs is shown in Fig. 3, the N-channel inspecticstate of the N-channel brokenignals from the transducer becomes its sampling中国煤化工equence by A/D conversion. If the number of sam-CNMHG(10)ples of the signals of broken wires is K, the sequenceof broken wire sample signals of the jth channel canbe expressed as a row vector with k elementsThe characteristic quantity h expressed in formula(10) contains all the characteristics of brokenWANG Hong-yao et alResearch on Detection Device for Broken Wires of Coal Mine-Hoist Cablewires can be set arbitrarily. If we set the maximumnumber of broken wires in a cross sectionthe5 Quantitative Recognition of Broken unit number of the output layer is n+1Wires in Steel wire Cables4)Determination of the number of hidden layerunitsIn the detecting devices for broken wires, a tradiTo make the network accurately recognize thenumber of broken wires, an appropriate number ofthe recognition of broken wires. Because of the com- hidden layer unit nodes must be chosen. If the num-plexity of the structure of wire ropes, bad working ber of nodes is too small, the pattern space partition isconditions, as well as the limitation of the inspecting inaccurate, possibly causing a failureof the networkmethod based on leakages in the magnetic field, a to study the characteristics of the training sample.Ifuniform model and quantitative recognition rules are the number of nodes is too large, the pattern spacedifficult to obtain. Neural networks are a kind of partition is too precise, possibly causing a failure ofauto-adapted pattern recognition technology, which the network to retain the main characteristics. In ourprovides an effective method for the quantitative research, we have simulated different numbers ofrecognition of broken wires.hidden unit nodes We constructed a network whichTherefore, we adopted the neural network recogniinitially, had few hidden unit nodes and then gradution method and established a nn model for the rec- ally increased the number to an optimum number ofognition of broken wires in wire cables by a BP neu. hidden layer units'ral network, as shown in Fig 4.5)Training of BP neural networkThe core algorithm of the bp neural network is theCharactePositioin modification of the weight value and its threshold bytraining the known samples in order to reduce theroot-mean-square error between the network outputand the target output until a pre-requisite is achievedGiven that there are three layers in a bP neuralThe relative movement speed betwecnnetwork anf if we define the number of input layernodes as 6, the number of output layer nodes as 7, theFig 4 Network structure of quantitative recognitionnumber of connotative lahfor broken wires(√6+8+1≤h≤v6+8+10, the input vector as1) Design of network layer numbersXP=(ro, r, x2,I3, r4, xs), then the expIn the reverse communication network, excessive output vector is D=(do, d,d2,",dumbers of hidden layers not only slow down calcu-lation speeds, but also enlarge the value of propa-For the input layer, the input vector is x, i=0, 1,gated error. Because a three layer neural network can 2,3, 4,5realize discretionary complex function mapping, theFor the connotative layer, the jth node input isthree-layer BP network is used to establish the NNmodelnet1=∑wx-日(l1)2)Determination of input parametersiven our analysis, we conclude that the dependentfactors which carry out the recognition of brokenFor equation (11),W is the network connectionwires are largely the following: the characteristicweight value between the nodes of the input layer andquantity of the broken wires signal, the size of the the nodes of the connotative layer. 0, is the conno-gap between the detection device and the surface of tative layer thresholdwire cables, the type of wire cables, the specificationIts jth node output isof wire cables, the running rate and the fracture size.Therefore, the bP network to be established has sixy=f(ne),=0,1,2,…,h-1(12)3)Determination of output parametersFor the output layer, the kth node input isy(i represents the number of broken wires in aner,=>v(13cross section of steel cables; if y(=l, we consideredthat broken wires do exist and the number of brokenwires is 1. If y(i=0, we considered that there are no中国煤化工 ork connectionbroken wires In our investigation, we set the maxi- weightCNMHGlayer nodes andmum number of broken wires in a cross section to 7 the outpin the network, so that the unit number of the output threshold.layer is 7+1=8. To be sure, the number of brokenThe kth node output is380Journal of China University of Mining Technologyz=∫(net),k=0,1,2,…,7d=f(ner)∑We defined the Energy Function(Error Function)of bp network asThe neural activation function presents in the formof an unipolar function, i.e.,Epf(net,=1/(1+exp(-net, )By modifying the weight values and the thresholden there arethe network converges to a stable state and its learn- f,(net: )=f(net (1-f(net: ))=4 (1-z)(23)ing course ends when the energy is minimized. Thesolution to this optimization equation without restric-f(mr)=r(ne)(1-f(n)=x(-x)(24)tions are found in Newtonian iteration methodsgradient-down methods, etc. The first two methods c Under equations(20), (21),(22 ),(23)and(24), thequasi-Newtonion iteration methods and optimizedquation of modifying weight values can be develinvolve the solution of inverse matrices and require opedlarge numbers of calculation. So we selected the last△y=n(a4-x1)f(ne)ymethod to modify the weight values.(25)For the weight values between the output layer and7(4-x)x41(1-x)ythe connotative layer to be modified, every modified△w=8x=x(1-x)∑v(26△=nE=n0E.The threshold 6 is a variable, with the same the(16 ory of weight value modification, 8 needs modification as the weight values are modifiedIn equation( 16), n refers to the number of steps inThe most apparent shortcoming of the BP networkthe learning process 0