Application of ANN to Predict Height of Weld Bead under Magnetic Field R.P. Singh 1, R.C. Gupta 2, S.C. Sarkar 3, K.G. Sharma 4, 5 P.K.S. Rathore 1 Mechanical Engineering Depart, I.E.T., G.L.A. University Mathura, (U.P. 2 Mechanical Engineering Depart, I.E.T., Lucknow, (U.P. 3 Mechanical Engineering Depart, Kumaon Engineering College, Dwarahat, (Uttarakhand 4 Computer Science Depart, I.E.T., G.L.A. University, Mathura, (U.P. 5 Mechanical Engineering Depart, I.E.T., G.L.A. University Mathura, (U.P. * Corresponding author: R.P. Singh; Abstract: t is a significant physical characteristic of a weld which determines the quality of weld. Several welding parameters such as welding speed, welding current, volt, and external magnetic field affect the reinf t of weld. Traditionally, an expert welder from his experience of trial and error selects a set of parameters that may yield fairly good results. However, the trial and error can be avoided, if a suitable automation tool can be developed, which could forecast the output from a set of desired parameters. Artificial neural networks (ANN were applied to predict the reinf t of weld obtained in shielded metal arc welding of mild steel in an external magnetic field, produced by a bar magnet. Back-propagation neural networks algorithm has been followed to associate the welding process parameters with the reinf t. Some basic concepts relating to neural networks were explained as well as how they could be used to model reinf t in terms of the equip parameters selected to produce the weld. Approaches to utilization of neural networks in process control were discussed as well. The performance of neural networks for modeling was presented and evaluated using actual welding data. It is concluded that the accuracy of neural networks modeling is fully comparable with the accuracy achieved by more traditional modeling schemes. Keywords: Artificial Neural Networks, Back Propagation, Welding Parameter, Shielded metal arc welding. 1. INTRODUCTION An Artificial Neural Network (ANN is an information paradigm that works in the same way as the biological nervous systems work. It is composed of a large number of highly interconnected processing eles (neurons working together to solve the given problems. In this ANN system a set of inputs are applied, each representing the output of another neuron. Each neuron input X (n is multiplied by a corresponding weight W (n analogous to a synaptic strength, and all the weighted inputs are then summed to determine the activation level of the neuron. These products are simply added producing the result which is then fed through a transfer function producing the final output. This transfer function is often a sigmoid. Feed forward ANNs allow signals to travel one way only, from input to output but feedback networks can have signals travelling in both directions by introducing loops in the network. A trained neural network can be considered to be as an expert in the category of information for which it has been assigned to work. It can be used to provide projections given new situations of interest and answers what if questions. A major advant of ANN approach is that the domain knowledge is distributed in the neurons and information processing is carried out in parallel distributed manner [1]. ANNs are highly parallel data processing tools capable of learning functional dependencies of data [2]. Being adaptive units they are capable to learn these complex relationships in any condition. This provides the capability to do "Black Box Modeling" with little or no prior knowledge of the function itself. ANNs can be used for nonlinear static-dynamic systems. In the shielded metal arc welding process a minor variation in the arc length i.e. even a few mm is sufficient to produce a very large fluctuation in arc volt which may be beyond the allowable range of power source. The welding system is highly complicated and non linear system because of the non linear relationship between the arc volts, current, welding speed, external magnetic field and reinf t. It is thus not easy to be modeled by conventional mathematical framework based approach. Artificial Neural Network may be beneficial for this. Numerous attempts have been recommended to develop Volume 1, Issue 4 November - December 2012 P 70
mathematical models relating input process parameters and weld bead geometry. Recently, Artificial Intelligence (AI methods such as fuzzy logic, Artificial Neural Networks (ANN and some expert system have been used as key techniques for monitoring and controlling welding processes. An ANN model was developed by Abdullah Alfaruk et al. to predict weld bead geometry and penetration by considering electrode diameter, current, volt, travel speed; electrode feed rate, arc length and arc spread as influential factors for electric arc welding successfully [3]. ANN modeling has been chosen by its capability to solve complex and difficult problems. Kim et al. used multiple regression analysis and back propagation neural network in modeling bead t in metal arc welding [4]. They comprehended that the back propagation neural network is considerably more accurate than multiple regression. Nsh and Datta reported that artificial neural networks are powerful tools for analysis and modeling. They applied back propagation neural network to predict weld bead geometry and penetration in shielded metal arc welding [5]. In present scenario, ANN models have been used by many researchers to understand and predict their targeted information This paper presents the develop of neural network model to predict reinf t for various input process parameters in mild steel butt welding deposited by SMAW. Multilayer feed forward neural network was developed and it was trained using back propagation algorithm. The proposed learning algorithm for this system is the backpropagation learning algorithm. Back-propagation learning is a supervised learning where it needs to know the inputs and the desired outputs in advance. It later compares the actual output computed from the given inputs to the desired output and to calculate the error. The error is then propagated backwards through the network and weights are changed based on the back-propagation. neural networks depends upon, the number of hidden layers and number of neurons in the hidden layers. Hence, optimum structure is obtained by changing number of hidden layers and neurons by making many attempts. The appropriate neural networks structure was chosen by the trial and error method [4]. Feed forward artificial neural network structure was established by keeping four neurons in the input layer, two hidden layers having five neurons in each and one neuron in output layer using C++. It was trained with help of back propagation (BP algorithm. BP is essentially stochastic approximation to nonlinear regression. Several researchers used BP to model welding processes and predicted welding parameters using NN. Twenty five sets of current, volt, speed of welding and external magnetic field are used to find out the corresponding reinf t. Eighteen sets are used to train the ANN and remaining seven sets are used for testing purpose. The flow chart for the back, propagation algorithm is shown in figure 1. 2. METHODOLOGY OF ARTIFICIAL NEURAL NETWORK MODELING Several industrial processes are non-linear, complex and many input variables are involved in processes. The mathematical models cannot give closer approach to describe the behavior of the processes. ANNs are easy to understand, cost effective and have the capability of learning from examples and are used in many industrial application. ANN model has been developed for general application consisting of the following steps: (i Database collection, (ii pre-processing of input/output data, (iii design and training of neural network, (iv testing of trained network, (v post processing and (vi use trained network for prediction [6]. The arrange of neurons into layer and the connection pattern within and between the layers are called as network architecture. The architecture is consisted of three parts: (i Input layer receives the welding parameters, (ii Hidden layers considered as block boxes and (iii Output layer obtaining the values of bead geometry. The performance of the Figure-1 Flow chart for the back propagation neural network 3. FORMULATION USING BACK PROPAGATION ALGORITHM A neural network can be utilized to perform a particular work by using certain procedures. The back propagation (BP algorithm is one of the supervised training algorithms for multilayered feed forward neural networks [7] and [8]. The used algorithm for back propagation is given below: STEP 1: Normalize the inputs and outputs with respect to their maximum values. For each training pair, assume that there are l inputs and n outputs in normalized forms. STEP 2: Assume the number of neurons in the hidden layer to be in between l and 2l. X i represents the neuron to input layer, y i represents that of output of input layer, y is represents the sigmoidal output of input layer which is Volume 1, Issue 4 November - December 2012 P 71
also input to the first hidden layer, Y i represents the neurons of output of first hidden layer and Y is represents that of sigmoidal output of first hidden layer which is also input to the second hidden layer. Z i represents the neurons of output layer and Z is represents that of sigmoidal output of output layer. STEP 3: [W], [V] and [U] represent the weights of synapses connecting input neurons to first hidden neurons, first hidden neurons to second hidden neurons and second hidden neurons to output neurons respectively. Sigmoidal gain λ is assumed as 1 and threshold value θ is taken as zero. Moum coefficient α is assumed to be zero. STEP4: To train data, present one set of inputs and outputs. Present the inputs to the input layer. The output of the input layer may be evaluated as STEP 5: Compute the inputs to the first hidden layer by multiplying corresponding weights of synapses as STEP 6: Write the output of the first hidden layer as a sigmoidal function as This is the input for second hidden layer. STEP 7: Compute the outputs to the second hidden layer by multiplying corresponding weights of synapses as STEP 8: Write the output of the second hidden layer as a sigmoidal function as This is the input for output layer. STEP 9: Compute the outputs to the output layer by multiplying corresponding weights of synapses as STEP 10: Let the output layer units evaluate the output using sigmoidal function as This is the network output STEP11: Calculate the error and the difference between the network output and the desired output as for the ith training set as and here β is known as learning rate coefficient. STEP 15: calculate new values of weights as STEP 16: Check whether the required number of iterations have been completed. If yes, then this indicates that the modified weights are obtained and the training of data is completed. If no, then follow step 5. In training, it is essential to balance the importance of each parameter; hence the data must be normalized. Since, neural networks work better in the range of 0 to 1 [9], the input and output vector values are converted in the range of 0 to 1 using the following equation. Where Xn = normalized value, X= actual input (or output value, Xmax = Maximum value of the inputs (or outputs, Xmin=Minimum value of the inputs (or outputs. The designed neural networks structure was 4-5- 5-1 (3 neurons in input layer, 5 neurons in both hidden layers and 1 neuron in output layer. Proposed feed forward neural network architecture is shown in figure 2 [10]. Non-linearity and input-output mapping are the useful comple in neural networks. Hence, it has been adapted to model the input-output relation of nonlinearity and interconnected system. Table 1: Data for Training and Prediction Data for Seri al Num ber Curr ent (A Volt (V Welding Speed (mm/min Magnetic Field (Gauss Trai ning 1 90 24 40 0 t in second hidden layer. STEP 12: Calculate the error for output and second hidden layer for the ith training set as in second hidden layer. STEP 13: Calculate the error for second hidden layer and input layerfor the ith training set as in first hidden layer. STEP 14: Calculate small changes in weight values as 2 90 24 40 20 3 90 24 40 40 4 90 24 40 60 5 90 24 40 80 6 95 20 60 60 7 95 21 60 60 8 95 22 60 60 9 95 23 60 60 10 95 24 60 60 11 100 22 40 40 1.06 1.07 1.12 1.17 Volume 1, Issue 4 November - December 2012 P 72
12 100 22 60 40 13 100 22 80 40 14 90 20 80 20 15 95 20 80 20 16 100 20 80 20 17 105 20 80 20 18 110 20 80 20 1.15 1.06 1.08 Data for Predi ction 1 90 23 40 0 2 95 22 60 40 3 95 21 80 60 4 100 24 40 40 5 105 21 60 40 1.04 1.16 Figure 2 Feed-forward neural network (4-5-5-1 architecture 4. RESULTS Table-2 depicted the measured reinf t from the experi and predicted output values using artificial neural feed forward network. The measured and predicted output values are close to each other as the maximum percent error in prediction is 3.54, which is very less and even it can be reduced by increasing the number of iterations and hidden layers. The aim of this paper shows the possibility of the use of neural network to predict the weld bead geometry. S. N. 6 105 22 60 20 7 110 21 60 20 Table 2: Measured and Predicted Values with percent Error Cur ren t (A Volt (V Wel ding Spee d (mm /min Magn etic Field (Gaus s 1 90 23 40 0 2 95 22 60 40 3 95 21 80 60 4 100 24 40 40 5 105 21 60 40 6 105 22 60 20 7 110 21 60 20 t(mm Meas ured 1.04 1.16 1.10 t (mm Predi cted 1.10 t % 1.10-3.51 1.08-2.70 1.06 +1.92-1.72-2.63-3.54 1.08-1.82 5. DISCUSSION / ANALYSIS ts of all the joints were evaluated and they were presented in table 1 [11]. The magnetic field had almost no effect on reinf t if it was changed in between 0 and 40 gauss, and after this the reinf t decreased if magnetic field was increased upto 80 gauss which was our investigation range. If the magnetic field was increased from 40 gauss to 60 gauss the reinf t decreased from mm to mm and if it was increased from 60 gauss to 80 gauss the reinf t decreased from mm to mm. If the speed of welding was increased from 40 mm/ min to 80 mm/min the reinf t continuously decreased. Incre in volt from 20 to 24V, increased the reinf t from 1.06 mm to1.12 mm. if the incre in current was from 90 A to 110 A, the reinf t of weld generally. 6. CONCLUSION The experial analysis confirms that, artificial neural networks are power tools for analysis and modeling. Results revealed that an artificial neural network is one of the alternatives methods to predict the weld-bead geometry. Hence it can be proposed for real time work environ. Based on the experial work and the Volume 1, Issue 4 November - December 2012 P 73
neural network modeling the following conclusions are drawn: (1 A strong joint of mild steel is found to be produced in this work by using the SMAW technique. (2 If amper is increased reinf t generally increases. (3 If volt of the arc is increased reinf t generally increases. (4 If travel speed is increased reinf t of weld generally decreases. (5 If magnetic field is increased reinf t of weld generally decreases. (6 Artificial neural networks based approaches can be used successfully for predicting the output parameters like weld width, reinf t and depth of penetration of weld. However the error is rather high as in some cases in predicting reinf t it is more than 3 percent. Increasing the number of hidden layers and iterations can minimize this error. REFERENCES [1] A. Narendranath Udapa et al., An ANN based Approach for volt stability Assess. International conf. on Computer Applications in electrical engineering recent advances, pp: 666-670, 1997. [2] Task f 38-06-06 of study committee 38, Artificial Neural Networks for Power Systems. Electra No.159, pp: 78-101, 1995. [3] Al-faruk Abdullah et al, Prediction of Weld Bead Geometry and Penetration in Electric Arc Welding using Artificial Neural Networks. Int. Jour. of Mech. & Mechatronics Engg 10 No: 04. [4] L.S. Kim et al., Comparison of multiple regression and back propagation neural network approaches in modeling top bead t of multi-pass gas metal arc welds. Sci. & Tech.of welding and joining, 8(5, pp. 347-352, 2003.. [5] D.S. Nsh and G.L. Datta,. Prediction of weld bead geometry and penetration in shielded arc welding using artificial neural networks. Journal of Matr. Proc. Tech, 123(2 pp. 303-312, 2002. [6] Y. K. Yousif et al.,. Prediction of Friction Stir Weldingm Characteristic Using Neural Network. jourdan journal of. mech & Indus. Engg. 2, pp. 151-155, 2008. [7] Valluru Rao and Hayagriva Rao, C++ Neural Networks and Fuzzy Logic BPB Publications, First Indian Edition, 1996. [8] S. Y. Kung, Digital Neural Networks. Englewood Cliffs, NJ, Prentice-Hall, 1993. [9] Rajasekaran and G.A. Vijayalakshmi,. Neural Networks, Fuzzy Logic and Genetic Algorithms Synthesis and Applications. Prentice Hall of India, 2003. [10] R.P. Singh et al., Application of Artificial Neural Network to Analyze and Predict the Mechanical Properties of Shielded Metal Arc Welded Joints under the Influence of External Magnetic Field, International Journal of Engineering Research & Technology (IJERT, pp. 1-12, Vol. 8, Issue 1, October, 2012. [11] R.P. Singh et al., Prediction of Weld Bead Geometry in Shielded Metal Arc Welding under External Magnetic Field using Artificial Neural Networks, International Journal of Manufacturing Technology and Research, Vol. 8 number 1, pp. 9-15, 2012. AUTHOR Rudra pratap singh received the B.E. degree in Mechanical Engineering from MMMEC Gorakhpur in 1992 and the M.Tech. degree in mechanical Engineering in 2009 from UPTU Lucknow. During 1992 to 1999 he worked in Jindal Group as a quality control engineer, from 1999 to till date he is working in GLA group (now GLA University Mathura as a faculty in Mechanical Engineering Depart. He is pursuing Ph. D. (Registered in, March, 2010 from Uttarakhand Technical University, Dehradun. He has published three papers in international journals and three papers in national conferences. Volume 1, Issue 4 November - December 2012 P 74