A NEURAL NETWORK FOR WELD PENETRATION CONTROL IN GAS TUNGSTEN ARC WELDING A NEURAL NETWORK FOR WELD PENETRATION CONTROL IN GAS TUNGSTEN ARC WELDING

A NEURAL NETWORK FOR WELD PENETRATION CONTROL IN GAS TUNGSTEN ARC WELDING

  • 期刊名字:金属学报(英文版)
  • 文件大小:751kb
  • 论文作者:C.S. Wu,J.Q. Gao,Y.H. Zhao
  • 作者单位:Institute of Materials Joining
  • 更新时间:2020-09-15
  • 下载次数:
论文简介

Availableonlineatwww.sciencedirect.comACTAMETALLURGICA SINICABCINCDIRECT(ENGLISH LETTERS)Acta Metall. Sin. (Engl. Lett. )Vol 19 No. 1 pp27-33 Feb. 2006www.ams.org.cnA NEURAL NETWORK FOR WELD PENETRATION CONTROL INGAS TUNGSTEN ARC WELDINGC.S. Wu, J.e. Gao and Y H. zhaoInstitute of Materials Joining, Shandong University, Jinan 250061, ChinaManuscript received 5 February 2005Realizing of weld penetration control in gas tungsten arc welding requires establishment of aonship between the front-side geometrical parameters of weld pool andthe back-side weld width with sufficient accuracy. A neural network model is developed to attainthisaim. Welding experiments are conducted to obtain the training data set(including 973 groupsofers of the weld pool(108 groups ) Two data sets are used for training and verifying the neural network, respectivelyThe testing results show that the model has sufficient accuracy and can meet the requirements ofveld penetration control.KEY WORDS neural network, weld penetration control, back-side weld widthgas tungsten arc welding1. IntroductionGas tungsten arc(GTA)welding is widely used for precise joining of metals in manufacturing industry l Since there are many disturbances in welding process(such as heat dissipation, variation of work-piece, minor changes of groove sizes, thermal distortion of workpiece), the automatic control of the gtawelding process, especially weld penetration control must be achieved in order to guarantee weld qualityl. For the case of full penetration, the state of the weld penetration can be specified by the back-sideweld width. The back-side weld width has a close relationship with the weld penetration s. If a back-sidesensor is used, the back-side weld width can be measured. However, the coordination movement of boththe welding torch at the front-side of weldment and the sensor at the back-side is very difficult to realizeOn the other hand, application of a back-side sensor cannot be allowed when welding some productssuch as pipes and vessels. Thus, it is required that the sensor be attached to and move with the torch toform a so-called front-side sensor. Vision sensors such中国煤化工 CCD)camera,areemployed to observe the weld pool from the front-side of wCNMH Gon of some specialCoresponding author. Tel :+86 531 88392711E-mailaddress:wucs@sdu.edu.cn(CS.Wu)measures of reducing the strong disturbances from the arc light, such a sensor is able to measure theweld pool geometry at the front-side, but is unable to provide directly with any information on the back-side weld width. Hence, the measured weld pool geometry at the front-side must be correlated with theback-side weld width. It is of great significance to establish a model describing the relationship betweenthe front-side geometrical parameters of the weld pool and the back-side weld width. Such a model is aprerequisite for vision-based control of weld penetration because it could predict the information on theback-side weld bead based on the weld pool sizes measured from the front-side. In this study, a neuralnetwork model is developed for predicting the back-side weld width according to the front-side weldpool geometry. Based on the model a strategy is made to control welding process parameters(weldingcurrent and travel speed), thus achieving automatic control of weld penetration in gta welding. Thispaper focuses on the development of the neural network model2. Measurement TestsIn order to develop a neural network model with satisfactory precision, sufficient experimental datacovering all specimen space must be provided for training and verifying. The experimental system consists of the following major functional elements: a host computer, a welding control unit, an image cap-turing board, a CCD camera with narrow-band composite light-filter a GTa welding supply, a worktable, a welding speed controller and a monitor. The information on the weld pool image captured by theCCD camera with narrow-band composite light filter is sent to the host computer via the image capturing board, and then processed to get geometrical parameters of the weld poolThe welding experiments are conducted on mild steel Q235 plates with a size of 250mm x60mmx2mm. Bead-on-plate welding is conducted. In order to acquire enough experimental data covering allspecimen space, 16 kinds of experiments are conducted. Those experiments demonstrate the change oftwo main welding parameters, i.e., welding current and travel speed They are classified into three cate-(1) The welding current varies while other conditions are kept unchanged(2) The travel speed varies while other conditions are kept unchanged(3) Both welding current and travel speed vary while other conditions are kept unchangedig. 1 shows GTA weld pools made using different welding currents. Fig. 2 shows weld pools underdifferent travel speeds. Once the image of the weld pool captured by the vision sensor is digitized throughframe-grabber, it is stored in computer as a matrix in which one element(pixel)represents a dot of im-age. A series of image processing are carried out, such as eliminating noise, enhancing contrast, and ex-tracting the edges of weld pool. The image process algorithm includes two steps: 1)the methods of noiseelimination and image enhancement are combined with each other to sharpen the points between thetrough and peak in grey level; (2) heuristic edge tracking method is used to search for edges. Then theweld pool geometry can be obtained by determining the coordinates of the left edge and right edgepoints. The detailed algorithm and formulae may be referred to Ref[5Through setting up the model for image capturing of vision sensor, the real size corresponding to apixel in the image of weld pool is determined. The calibrat中国煤化工 e transforming co-efficients: 0.043mm/ pixel in the direction perpendicular tNMHGd0.0752mm/pixelalong the welding direction. Therefore, the real dimension uIgeometry at front-side is deter-minedFig. I Gta weld pools made from different welding currents(travel speed 180mm/min, arclength 6mm): (a)100A; (b)105A; (c)110A; (d)115AFig 2 GTA weld pools under different travel speeds(welding cur-rent 100A, arc length 6mm):(a)YH中国煤化工CNMHGFig 3 illustrates the definition of the front-sideof weld pool(L), the maximum width of weld poolW), the area of weld pool(S), the rear angle of weldpool(a)and the ratio of weld pool width to length(W/L). Gta welding tests are carried out under dif-ferent conditions. The front-side geometrical parameters of weld pool are determined. The correspondingback-side weld width is detected off line with enoughaccuracy using a photo- measurement method. Total-ly, 1081 groups of geometrical parameters of the weldpool are acquired and they are divided into two group!Fig 3 The definition of front-side geometricali.e., one is a training data set (including 973 groupsparameters of GTA weld pool.of geometrical parameters of the weld pool and back-side weld width and the other is a verifying data set (108 groups ) two data sets are used for training andverifying the neural network, respectively.3. Construction of the Neural NetworkA three-layer back-propagation type of artificial neural network is set up Its input is the data setthe front-side geometrical parameters of weld pool while its output is the data set of the back-side weldwidth. This kind of neural network allows representation of the relations between input and output values. It is trained with the help of a supervised learning method, i.e., input and output values are specifiedand the relation between them are learnt. The three layers of the neural network are the input layer, thehidden layer and the output layer. The function of a neuron is described via its input function, its activefunction and its output function. The weight sum I of all signals which are active at the input connectionis employed as the input functionl1=∑wxwhere w components represent the connecting weights between neuron j and neuron i, and the x; com-ponents represent the signals at the connection concerned. The sum is transformed by the activation fundtion (sigmoid functionfza direct transfer of the activation of a neuron to its output is employed as the output function. Theoverall transfer function of a neuron is thus structured as follows0=a-=f∑vV凵中国煤化工CNMHGwhere o, is the output of the neuron, a; is its activation, x; is identical to the output of the preceding neuron with index j of the observed element.The learning process is carried out by means of the following stepsAn input vector is presented to the network, i.e., its elements are applied to the neurons of the inputlayer. This information is propagated through the network in forward direction, thereby generating anoutput vector o. Comparison of this network output with the target output values provides an error foreach output neuron. The aim of the training process is to minimize the global network error∑(a-owhere z components are the target output values. This is achieved by always altering the connectionweights in the direction of the steepest gradient of the error function. Adaptation of the weights is effect-d according to the equation△w=%(k+D-w()=dEdwhere a is defined as the learning rate, wy is the connection weights. This results in△w=a8where the local error of a hidden element is calculated via8=f()∑8wThe 8, components represent the errors of the elements in the following layer, while wy representsthe connection weights for these elements. It can be seen that, in order to calculate the error of a neuronin layer k, the errors of all neurons in layer k+l( the next layer in the direction of the output layer)arerequired. The error of a neuron of the output layer is obtained via61=∫()(z-0)(8)This error is first of all calculated and then back-propagated into the hidden layer located before theoutput layer. This procedure is continued until the input layer is reached. By the time the input layer isreached, an error has been calculated for each element. The connection weights can then be modified ac-cording to the calculated Awy in the concluding stage of this processIn order to overcome lower convergence and extinguish the danger of"getting stuck",a momentumterm and time-dependent learning rate are introduced in the training algorithm. The algorithm is as fol-)8x+n△w2(k-1)where Aw(k) is the kth connection weight variation in the algorithm, Aw,(k-1)is the(k-1)th connectionweight variation, n is momentum,(0Sn<1), and the learniThere is only one neuron at the output layer because or中国煤化工he back-side weldwidth)is considered. Since GTa welding process is charactCNMHlation and thermalag, the present values of the front-side geometrical parameters of weld pool can not precisely reflect theback-side weld width. Therefore, the historic values of these parameters must be considered in the mod-el. The different combinations of the front-side geometrical parameters of weld pool are used to get opti-mum results so that the number of neurons at the input layer can be determined. Similarly, differentnumbers of neurons at the hidden layer are tried to get most satisfaction output. The training of the neurnetwork is performed with aid of the commercial software DATE ENGINEI6. The training results showthat the speed of convergence is fastest and the training error(RMS error(RMS: root mean square)issmallest when a(0)=0. 1 and m=0.14. Results and VerificationDifferent combinations of front-side geometrical parameters of weld pool are taken as the inputvector. The number of such parameters is equal to the number of neurons at the input layer. In order todetermine the relationship between the numbers values of front-side geometrical parameters of weldpool and the back-side weld width, 10 kinds of neural network construction are established with different input parameters, as shown in Table l. w(r) and w(t-1)represent the values of weld pool width atthe present moment(t)and the preceding moment(1-1), respectively. Other parameters are representedin the same wayAt present, there is no good way to determine the number of neurons in hidden layer. In this studyten neural network models are established for each kind of neural network input parameters and thenumber of neurons in hidden layer is from 2 to 20 with step variation of 2. For each kind of neural network, the model with the smallest training error is the optimum and the number of neurons in the hiddenlayer is also the optimumFrom Table l, it can be seen that the training error of the model is least when the model takes thepresent values and the latest historic values of W, L, S, a and W/L as input parameters(No 10).Thenumbers of neurons at the input layer and the hidden layer are 9 and 4, respectively. Fig 4 is the structureof the developed neural network modelThe non-training data sets are used to verify the accuracy of the neural network model Fig 5 is theTable I Input parameters, the optimum number of nodes in hidden layer and least training error of each kinds ofneural networkBest number of nodes in Minimumhidden layerRMS. mmWL0.338L s0.411W.L. S0330W.L.S0.351W.L. S a. WIL0.315WD,L(),S(t),Wt-1),L(t-1),S(t-1W(t),L(t),S(U),Wt-1),L(t-1),S(t-1),WL中国煤化工0.303W),L(1),S(n),a(m),W(t-1,L(t-1),(t-1),a(-1CNMHG10W(n,L(m),(t,a(t),W(-1),L(-1),S(-1),a(t-1),0.292ote: RMs is root mean square.FBWEE毛圣卫多器6080100120Sample numberFig 4 The structure of the neural network(BW is the Fig. 5 Testing the validity of the neural network modelback-side weld width)(FBW is the predicted results based on the model)comparison of the measured values of back-side weld width with the output of the neural network model. The maximum error is 0.766mm and RMS error is 0.285mm. this indicates that the model is withgood accuracy.5. Conclusions(1)Aneural network model is developed to describe the relationship between the front-side geometrical parameters of weld pool and the back-side weld width Since the back-side weld width can reflectthe weld penetration, the rnodel lays a solid basis for weld penetration control by using the front-side information of weld pool in GTA welding2)The numbers of neurons in the input and hidden layers are determined according to the RMS error of the neural network, thus avoiding the arbitrary selection ofinput layer parameters. Using this methoda model with 9 neurons in the input layer and 4 neurons in the hidden layer is establishedare, (3)973 samples are used to train the neural network model. 108 samples in a non-training data setsused to verify the model. The maximum error between non-training data sets and the correspondingoutputs of the model is 0.766mm and the rms error is 0.285mm. This indicates that the model is withgood accuracyAcknowledgements-This work was supported by the Shandong Provincial Natural Science Foundation of chinaNVo.Z2003F05)REFERENCES1 Y M. Zhang, R Kovacevic and L Li, IEEE Transactions on Control Systems Technology 4(4)(1996)3942 K A Pietrzak and S M Packer, Journal of Engineering for Industry 116(2)(1994)863 R Kovacevic, Y M. Zhang and L Li, Welding Journal 75(中国煤化工4J Q. Gao and C S. Wu, Acta Metallurgical Sinica 36(2000)1CNMHG5 J Q. Gao and C.S. Wu, Science and Technology of Welding and Joining 6(5)(2001)2886 MIT gmbH, Data Engine Overview and User Manual(Adchen, Germany, 1997). 177.

论文截图
版权:如无特殊注明,文章转载自网络,侵权请联系cnmhg168#163.com删除!文件均为网友上传,仅供研究和学习使用,务必24小时内删除。