Short Paper Proc. of Int. Conf. on Recent Trends in Mechanical, Instrumentation and Thermal Engineering 2012

Comparison of Artificial Neural Network and Response Surface Methodology in the Prediction of Springback and Bend force in Air Bending of Electrogalvanised Steel Sheets Dr. D.Vasudevan1and R.Srinivasan2 1

Department of Mechanical Engineering, PSNA College of Engineering and Technology, Dindigul, India. Email: [email protected] 2 Department of Mechanical Engineering, RVS College of Engineering and Technology, Dindigul, India. Email: [email protected]

Abstract— This paper compares the springback and bend force prediction models based on artificial neural network and response surface methodology. The models were developed based on five level-half factorial central composite rotatable design of experiments conducted on air bending of electrogalvanised steel sheets with strain hardening exponent, coating thickness, die opening, die radius, punch radius, punch travel and punch velocity as the process parameters. The ANN model of springback and bend force was developed using a multilayer feed forward back propagation network, trained with a Levenberg-Marquardt (LM) learning algorithm. The predictive capability of the developed ANN model was compared with the second order RSM models of springback and bend force. The comparison shows that the ANN models provide more accurate prediction than the RSM models.

are springback and bend force. During bending, when the bending stress is removed at the end of the deformation process, elastic energy remains in the bent part causing it to partially recover to its original shape. This elastic recovery is called springback. Bend force is the force needed to deform the sheet metal part to the required shape. The material geometry and properties (thickness, yield strength, Young’s modulus, etc.), tooling geometry (punch radius, die radius, die opening, etc.) and process parameters (punch travel, punch velocity, etc.) have considerable influence on springback and bend force. Springback prediction is important to control the process and die design to achieve the desired shape of the product. The prediction of bend force provides a base to the designer for the die design and press selection. Traditional trial and error methods are time consuming and expensive, while the development of theoretical models for springback/bend force is difficult and cumbersome due to the complexity of sheet bending process. Consequently, an empirical model developed based on experimental research is more useful in industrial applications. The empirical modeling techniques namely response surface methodology (RSM) and artificial neural network (ANN) become promising tools for prediction of springback and bend force because of their robustness and predictability Very few literatures is available for the prediction of springback and bend force in air bending using this prediction tools. Narayanasamy and Padmanabhan [3] developed a prediction model applying RSM, for bend force in air bending operation of interstitial free steel sheet. Forcellese et al [4] developed a neural network based punch displacement control system in air bending process of AA5754 aluminium alloy sheets to control springback Inamdar et al [5] discussed the development of ANN model for predicting springback in air vee bending of metallic sheets. Garcia Romeu and Cirauna [6] described the application of neural networks for the prediction of springback, punch displacement and final bending radius. Narayanasamy and Padmanabhan [7] compared the neural network model with regression model for the prediction of springback of interstitial steel sheets in air bending process. Fu et al [8] developed a back propagation neural network model to predict the springback for air bending of high-

Index Terms— Artificial neural networks, Response surface methodology, air bending, electrogalvanised steel;, springback, bend force.

I. INTRODUCTION Sheet metal forming is one of the major near-net-shape processes for the manufacturing of automobile, aircraft, electrical components etc. Sheet metal bending, being an important sheet metal forming process, employs a press brake with proper tooling to produce different bend components. The ûexibility of the bending process is improved by the air bending technique [1]. Flexibility is achieved in such a way that the different bend angles can be produced by solely controlling the punch travel into the die without the need for changing tool sets. Because of this versatility, the lead time is reduced in air bending technique and hence it is commonly used in sheet metal industries. Steel sheets are widely used to manufacture various parts in sheet metal industry, but the susceptibility to corrosion is their natural weakness. To overcome this, the uncoated steel sheets are substituted with galvanised steel sheets based on the improved corrosion resistance. Since electrogalvanised (EG) steel sheets have better formability [2] and surface quality, they are much preferred than hot dipped, in many of the industrial applications. The two important subjects related to bending process © 2012 AMAE DOI: 03.MES.2012.2. 2

73

Short Paper Proc. of Int. Conf. on Recent Trends in Mechanical, Instrumentation and Thermal Engineering 2012 TABLE I. THE SCHEME O F LEVELS AND VALUES OF PARAMETERS

strength sheet metal. The model is based on orthogonal test with input parameters such as sheet thickness, tool gap, punch radius, ratio of yield strength to Young’s modulus and punch displacement. The model yields satisfactory result in sheet metal air bending of a workpiece used as crane boom. In this paper, development of RSM and ANN models for predicting both springback and bend force has been attempted and the models are compared based on their prediction performance. II. EXPERIMENTAL DETAILS The input-output database required for the development of predictive models is obtained through bending experiments. The process of selecting appropriate set of data points is in such way that it maximises the accuracy of the prediction is known as design of experiments (DoE). In the present study, based on the previous work done on this field by various researchers [9-19], seven important and independently controllable process parameters which strongly influence the springback and bend force have been identified. They are strain hardening exponent, coating thickness, die opening, die radius, punch radius, punch travel and punch velocity. The design matrix considered in this study is a five level central composite rotatable design (CCRD) consists of half replication of 27(128/2 = 64) design points plus 10 center points and 14 (2X7) star points. The rotatability value is calculated by taking the fourth root of number of design points . The upper level of a factor is coded as +2.82 and the lower level as -2.82 for designing the experiments. The intermediate coded values are obtained from the equation as Coded value of the parameter value i

xi xi 0 di

The experimental setup for bending is shown in Fig.1. The punch was mounted in the upper arm of UTM and the die was placed on the lower platform. The sheet blanks were positioned on the die with necessary care. The punch traveled to the required depth for bending the sheet and the digital meter attached to the UTM was used to measure the punch travel. The larger edge of the bent sample was coated with black ink and the impression of the profile was taken carefully on a thick white paper supported by a board. Two impressions were taken before unloading and after unloading. The impressions of the sheet were scanned and converted into digitised images. The digitised images were imported to CAD software and the lines were drawn on the edges of the legs of images using the software. The necessary angles were measured using CAD software [18]. The difference between the bend angles during loading (θi) and after unloading (θf) gives the springback angle ( Δθ). The bend force was measured by the digital display of load cell arrangement. The sheets were bent to different depths by controlling the punch travel.

(1)

where xi is the actual parameter value, xi0 is the parameter value corresponding to zero level and di is the incremental parameter value. The actual and coded values of the process parameters are listed in Table 1. The substrate used in this study is Aluminium Killed Draw Quality (AKDQ) steel sheet. AKDQ steel sheet is a widely used sheet metal in automotive, electrical and other domestic applications because of its excellent formability and low cost. The steel sheet was electro galvanised for various coating thickness such as 4 µm, 7µm, 10µm and 14µm. The blanks from the steel sheets were cut to the required dimensions of 120 mm x 40 mm x 1mm and the edges were cleaned to remove the burrs. As the strain hardening exponent n values (0.211, 0.219, 0.232, 0.206 and 0.227) belong to five different orientations 0o, 22.5o, 45o, 67.5o and 90o respectively, the blanks were prepared in the five directions. The experiments were conducted in a 40T Universal Testing Machine (UTM) and the tooling (dies and punches) was made of hardened steel.

© 2012 AMAE DOI: 03.MES.2012.2. 2

74

Short Paper Proc. of Int. Conf. on Recent Trends in Mechanical, Instrumentation and Thermal Engineering 2012 TABLE II. ANOVA FOR RSM BASED SPRINGBACK AND BEND FORCE MODELS

Rp – Punch Radius ; Wd – Die Opening; θi – Bend Angle (During Loading) θf – Bend Angle (After Unloading) Δθ – Springback Angle

The experimental data were analysed to obtain the coefficients of polynomials using MINITAB 15 software. The experimental data were analysed to obtain the coefficients of polynomials using MINITAB 15 softwar The p value approach is used for testing the significance of coefficients. According to this, the coefficients are tested at 0.05 level of significance (95% confidence The analysis of variance technique (ANOVA) is employed to check the adequacy of the developed models at a 95% confidence level. As per this technique, the calculated value of F ratio for the lack-of-fit must be lesser than its standard value for a desired level of confidence level. Here, both the springback and bend force models are adequate in non-linear form, as the F ratio of both the models is less than the standard value for 95% confidence level, as illustrated in Table II. The final mathematical models for the springback and bend force given in terms of factors, as Springback

Rd – Die Radius; t – Sheet Thickness

Figure 1 Experimental Setup

III. RSM MODELLING Response surface methodology adopts mathematical and statistical techniques to evaluate the relation between a cluster of controlled experimental factors and a response. RSM provides an approximate relationship between the response Y and the independent variables xi which is based on the ob-

Y1 151. 5028 1256 .9472 n 0. 7578t c 0.4065Wd 0.8244 R d 0.1744 R p 0.2558t p

served data from the process. The optimal response can be obtained from this methodology even when there is minimal information about the process.

Y f ( x1 , x 2 , x3 ,...x n )

5. 4963V p 2785 .0397 n 2 0.0071t c2 0.0019Wd2 0.0591R d2 0.0348 R p2 0. 006t 2p 6. 3553V p2 2.5333 nt c 1.0036 nWd 3.2095 nRd 1.4023 nR p 0.9879 nt p 0.005t cW d 0.0093t c Rd 0.0243t c R p 0.0021t c t p 0. 0362t cV p 0.0055Wd R d 0.0023Wd R p 0.003Wd t p 0.0118Wd V p 0.0169 R d R p 0.0088 Rd t p 0.0843 Rd V p 0.0121 R p t p 0. 0597 R pV p

(2)

(5)

where denotes an error component. As air bending is a non-linear process [6], a linear polynomial is unable to predict the response accurately, and therefore the second-order model (quadratic model) is found to adequately model the process [20]. Second order non-linear models have been developed to predict springback and bend force which are of the following form k

k

Bend force Y 51.41 299.629n 0.029t c 0.291Wd 0.363R d 0.059 R p 0.085t p 0.823V p 721.323n 2 0.007t c2 0.002Wd2 0.006 Rd2 0.002 R p2 0.005t 2p 0.310V p2 0.671nt c 0.289nWd 0.664 nRd 0.312 nt p 0.004t c R d 0.026t cV p 0.001Wd Rd 0.001Wd R p 0.003R d t p 0.011R p t p 0.039t pV p

(6) where n is strain hardening exponent, tc is coating thickness, Wd is die opening, Rd is die radius, Rp is punch radius, tp is punch travel and Vp is punch velocity.

k

Y 0 i xi ii x i2 ij xi x j (3) i 1

i 1

i, j

IV. ANN MODELLING

where Y Response, i.e., springback and bend force xi Coded values for parameters

A. ANN Approach An Artificial Neural Network (ANN) is an information processing system that is inspired by information processing of the biological nervous system (brain). ANN has a parallelly distributed architecture with a number of information processing elements called neurons which are interconnected by connection links. A multilayer neural network has an architecture that consists of an input layer, one or more hidden layers and an output layer. The input layer receives input data and after processing it, sends them to the hidden layer. The hidden layer processes the data and sends a response

β Regression coefficients The values of the regression coefficients as defined above are determined as:

b XTX

1

X TY

where b Matrix of parameter estimates X Matrix of model terms evaluated at n data

(4) points

Y Matrix of the measured response © 2012 AMAE DOI: 03.MES.2012.2.2

75

Short Paper Proc. of Int. Conf. on Recent Trends in Mechanical, Instrumentation and Thermal Engineering 2012 to the output layer. The output layer accepts the responses and produces the result. The network which has an information flow from input layer to output layer in forward direction is known as feed-forward network. Since the ‘intelligence’ of the network exists in the values of the weights between neurons, a method of adjusting the weights is needed to solve a particular problem. The process of modifying the weights in the connections between network layers with the objective of achieving the expected output is called ‘learning’ or ‘training a network’. The backpropagation (BP) algorithm is the most widely used learning algorithm. The principal steps in the learning process are given below: Step1: Initialise all the weights of the links to random values. Step2: Present the input - desired output pattern sets for updating the weights and bias. The equation for updating weights and bias is

w zji1 w zji w zji1

B. ANN Architecture As neural network is a non-linear analysis tool, it maps the non-linear relationships with multiple interactions between input variables and springback/bend force in air bending in a better way [4,5,7]. Hence, a multilayer perceptron with back propagation training algorithm, widely used for engineering applications, is considered as a practical choice for this work. Single hidden layer is used for many practical problems since it can approximate any function which contains a continuous mapping from one finite space to another [21]. Hence single hidden layer is selected and the architecture of the ANN in this work has a three layer structure of one input layer, one hidden layer and one output layer. The number of neurons in input and output layers is 7 and 2, corresponding to the number of inputs and outputs. As a heuristic rule, the number of hidden layer neurons can be up to 2n+1 where n is the number of neurons in the input layer [22]. In this study, the number of neurons in the hidden layer is changed and the Mean Square Error (MSE) is evaluated. After several trials, the number of neurons which results least MSE is selected for hidden layers and is 14. The designed architecture becomes 7-14-2. The designed ANN architecture is shown in Fig. 2. Levenberg-Marquardt (LM) learning algorithm, the fastest method for training moderate-sized feed forward neural networks [23] is an efficient method for nonlinear problems [24]. This makes LM algorithm the much preferred one in this work. The transfer function is required to introduce the nonlinearity characteristics into the network. Log-sigmoid function takes any real valued input and returns an output bounded between [0,1]. In backpropagation, it is important to calculate the derivatives of any transfer functions used. As it is differentiable, the function log-sigmoid is commonly used in back propagation networks [25]. Hence, log-sigmoid was chosen as the activation function for the present network.

(7)

where w ji Weight of the link connecting neuron j to i, z is z 1

the learning step and w ji is the incremental change in the weight. The weight change is determined by a learning algorithm. Based on LM learning algorithm, the weight change can be written as

w zji1 J T J I

1

J T e w zji w zji1

(8)

where J is Jacobean matrix that contains first derivatives of the network errors with respect to the weights and biases, μ is the adaptive training parameter, I is the identity matrix, e is the vector of network errors and α is the momentum term. The use of a momentum term in the backpropagation algorithm is primarily to overcome the possible trap to local minima and also to overcome the oscillations during the training of the network. Moreover, the use of a small momentum helps in increasing the network training convergence with the LM algorithm. Step 3: After presenting the patterns, compute the mean square error (MSE) of all outputs as:

MSE

1 QK

Q

K

d

y kq

2

kq

(9)

q 1 k 1

where Q is the number of sets of input-output data K is the total number of outputs

Step 4: If MSE MSE t arg et

then stop, else go to

Figure 2 Architecture of the present ANN

step 2. If the new MSE is smaller than that of the preceding, then reduce the training parameter µ by µ-. If the MSE is increased, then increase µ by µ+. After the training stage, the ANN learning performance is evaluated by running it using the test data set, which is not included in the training set. If the network responds correctly to the inputs, the generalisation of network is confirmed. © 2012 AMAE DOI: 03.MES.2012.2. 2

C. ANN Training The experimental design matrix of 88 bending experiments along with the experimental results of springback and bend force was used as the training set for the network. As the dimension and magnitude of the testing or training data are different, they should be normalised before applying in the network, so that they lie between 0 and 1. The 88 inputoutput pairs were fed into the computer and a computer pro 76

Short Paper Proc. of Int. Conf. on Recent Trends in Mechanical, Instrumentation and Thermal Engineering 2012 gram was performed in the MATLAB 7.0.1 software. The training was performed in batch mode with the following learning factors: Learning rate = 0.1; Momentum constant = 0.5; Target for MSE =0.0000001; Maximum number of epochs = 1000. During training, the network compares its predicted value to the actual output and adjusts all the weights to improve the model. Once the MSE of the raining data reached the target value, the training is terminated and the weights and biases are automatically saved by the program. The MSE target is achieved in 292 epochs and the variation of MSE during the training is shown in Figure 3.

V. COMPARISON OF RSM AND ANN MODELS The developed RSM and ANN predictive models were compared on the basis of their prediction accuracy. The measured values of springback and bend force from the 15 test experiments and the corresponding values of the ANN and RSM models were compared to test the interpolation of the prediction models. The accuracy of the developed models is assessed by means of absolute error function and it has been defined as

Vm Vexp % Absolute error =

Vexp

100

(10)

where Vm is the predicted value of the model and Vexp is the measured experimental value. Figures 4 and 5 illustrate the comparison of the error profiles for springback and bend force respectively, for 15 test data points. The average absolute error for springback and bend force in ANN model is 1.24% and 1.41% while for RSM model it is 3.50 % and 3.92% respectively. Despite the error being within the acceptable level for both the modeling techniques, ANN outperforms the RSM technique. VI. DISCUSSIONS Though the researchers [4-8] developed ANN models in air bending earlier, they concentrated mainly on the prediction of springback and not on bend force, which is another important parameter to be considered in air bending. The important parameters like die opening and die radius which are having major influence on springback and bend force have not been considered by Narayanasamy and Padmanabhan [7]. The main advantage of the present ANN is that it can be used as a computer aided tool for providing easy, quick and precise predictions in process design in air bending of electrogalvanised steel steel. This work is a new attempt on electrogalvanised steel sheet and it can be further extended for other types of coated steel sheets also.

Figure 3. Variation of MSE during training the network

CONCLUSIONS

Figure 4. Comparison of error profiles of ANN and RSM for Springback

This study described a comparative analysis of the two modeling approaches namely artificial neural network (ANN) and response surface methodology (RSM) for the prediction of springback and bend force in air bending of electro galvanised steel. A five level central composite design of experiments was employed to create the input-output data pairs for developing both the models. The ANN model was based on a multilayer feed forward topology and trained with LM back propagation algorithm. The performance of the models was compared based on their prediction accuracy using a new set of 15 input-output pairs. The performance of the ANN model for predicting springback and bend force was found to be better than RSM models.

Figure 5. Comparison of error profiles of ANN and RSM for Bend force

© 2012 AMAE DOI: 03.MES.2012.2. 2

77

Short Paper Proc. of Int. Conf. on Recent Trends in Mechanical, Instrumentation and Thermal Engineering 2012 REFERENCES

[12] C.Bruni, A.Forcellese, F.Gabrielli, and M.Simoncini, “Air bending of AZ31 magnesium alloy in warm and hot forming conditions”, J. Mater. Process Technol., vol.177, No.1-3, pp.373-376, 2006. [13] D.Fei,. and P.Hodgson, “Experimental and numerical studies of springback in air v-bending process for cold rolled TRIP steels”, Nucl. Eng. Des., vol.236, No.18, pp.1847-1851, 2006. [14] M.L.Garcia-Romeu, J.Ciurana, and I.Ferrer, “Springback determination of sheet metals in air bending process based on an experimental work”, J. Mater. Process Technol., vol.191, pp.174-177, 2007. [15] S.Y.Kim, W.J. Choi, and S.Y.Park, “Spring-back characteristics of fiber metal laminate (GLARE) in brake forming process”, Int. J. Adv. Manuf. Tech., vol. 32, No.5-6, pp. 445-451, 2007. [16] R.Narayanasamy, and P. Padmanabhan, “Effects of material orientation on air bending of interstitial free steel sheet”, J. Manu. Eng., vol.3, No.3, pp.163-169, 2008. [17] K.Yilamu, R.Hino, H.Hamasaki, and F.Yoshida, “Air bending and springback of stainless steel clad aluminium sheet”, J. Mater. Process. Technol., vol. 210, pp.272-278, 2010. [18] D.Vasudevan, R.Srinivasan, and P.Padmanabhan, “Effect of process parameters on springback behaviour during air bending of electro galvanised steel sheet”, J. Zhejiang UniversityScience A, vol 12, No.3, pp.183-189, 2011. [19] R.Srinivasan, D.Vasudevan, and P.Padmanabhan, “Experimental investigations of bend Force in air bending of electrogalvanized steel sheets”, J. Mater. Engg. Perform., DOI: 10.1007/s11665-011-0029-7. [20] K.C.Kapoor, and Q.Feng, , “Statistical method for product and process improvement”, in Springer hand book of engineering statistics, Pham, H. Ed. Springer-Verlag, London, 2006. [21] K.Hornik, M.Stinchcombe and H.White, “Multilayer feedforward networks are universal approximators”, Neural Networks, vol. 2, No. 5, pp.359-366, 1989. [22] R.Z.Saen, “The use of artificial neural networks for technology selection in the presence of both continuous and categorical data”, World Appl. Sci. J., vol. 6, No. 9. pp.1177-1189, 2009. [23] M.T.Hagan, and M.B.Menhaj, “Training feedforward networks with the Marquardt algorithm”, IEEE Trans Neural networks, vol 5, No.6, pp.989-993, 1994. [24] M.Sanjari, A.K.Taheri, and M.R.Movahedi, “An optimization method for radial forging process using ANN and Taguchi method”, Int. J. Adv. Manuf. Technol., vol.40, No. 7-8, pp. 76–784, 2009. [25] M.T.Hagan, H.B. Demuth, and M.H.Beale, Neural network design, PWS Publishing company, Boston, 1996.

[1] L.J. de Vin , “Curvature prediction in air bending of metal sheet”, J. Mater. Process Technol., vol. 100, pp. 257-261, 2000. [2] H.M. Jiang, X.P.Chen, C.Wu, and Li, H.H. “Forming characteristics and mechanical parameter sensitivity study on pre-phosphated electrogalvanised sheet steel”, J. Mater. Process Technol., vol. 151, pp. 248-254, 2004. [3] R.Narayanasamy, and P.Padmanabhan, “Application of response surface methodology for predicting bend force during air bending process in interstitial free steel sheet”, Int. J. Adv. Manuf. Technol., vol.144, No.1–2, pp. 38–48, 2009. [4] A.Forcellese, F.Gabrielli, and R.Ruffini, “Effect of the training set size on springback control by neural network in air bending process”, J. Mater. Process. Tech., vol. 80-81, pp. 493-500, 1998. [5] M.Inamdar, P.P.Date, K.Narasimhan, S.K.Maiti, and U.P.Singh, “Development of an artificial neural network to predict springback in air vee bending”, Int. J. Adv. Manuf. Technol., vol. 16, pp. 376-381, 2000. [6] M.L.Garcia-Romeu, and J.Ciurana, “ Springback and geometry prediction – Neural networks applied to the air bending process”, Huang, D.S., Li, K., and Irwin, G.W. Eds. ICIC 2006, LNCS 4113, pp. 470–475, Springer-Verlag, Berlin Heidelberg, 2006. [7] R.Narayanasamy, and P.Padmanabhan, “Comparison of regression and artificial neural network model for the predicting springback during air bending process of interstitial free steel sheet”, J. Intell. Manuf., DOI: 10.1007/s10845-009-0375-6, 2010. [8] Z.M.Fu and J.H.Mo, “Springback prediction of high-strength sheet metal under air bending air bending forming and tool design based on GA-BPNN”, Int. J. Adv. Manuf. Technol., vol. 53, Nos. 5-8, pp. 473-483, 2011. [9] C.Wang, G. Kinzel, and T.Atlan, “Mathematical modeling of plane-strain bending of sheet and plate”, J. Mater. Process. Technol., vol. 39, pp. 279-304, 1993. [10] Y.M.Huang, and D.K.Leu, “Effects of process variables on V-Die bending process of Steel Sheet”, Int. J. Mech. Sci., vol.40, No.7, pp.631–650, 1998. [11] M.V.Inamdar, P.P. Date,. and S.V.Sabnis, “On the effects of geometric parameters on springback in sheets of five materials subjected to air vee bending”, J. Mater. Process Technol., vol.123, No.3, pp.459-463, 2002.

© 2012 AMAE DOI: 03.MES.2012.2. 2

78

Comparison of Artificial Neural Network and Response Surface Methodology in the Prediction of Springback and Bend force in Air Bending of Electrogalvanised Steel Sheets Dr. D.Vasudevan1and R.Srinivasan2 1

Department of Mechanical Engineering, PSNA College of Engineering and Technology, Dindigul, India. Email: [email protected] 2 Department of Mechanical Engineering, RVS College of Engineering and Technology, Dindigul, India. Email: [email protected]

Abstract— This paper compares the springback and bend force prediction models based on artificial neural network and response surface methodology. The models were developed based on five level-half factorial central composite rotatable design of experiments conducted on air bending of electrogalvanised steel sheets with strain hardening exponent, coating thickness, die opening, die radius, punch radius, punch travel and punch velocity as the process parameters. The ANN model of springback and bend force was developed using a multilayer feed forward back propagation network, trained with a Levenberg-Marquardt (LM) learning algorithm. The predictive capability of the developed ANN model was compared with the second order RSM models of springback and bend force. The comparison shows that the ANN models provide more accurate prediction than the RSM models.

are springback and bend force. During bending, when the bending stress is removed at the end of the deformation process, elastic energy remains in the bent part causing it to partially recover to its original shape. This elastic recovery is called springback. Bend force is the force needed to deform the sheet metal part to the required shape. The material geometry and properties (thickness, yield strength, Young’s modulus, etc.), tooling geometry (punch radius, die radius, die opening, etc.) and process parameters (punch travel, punch velocity, etc.) have considerable influence on springback and bend force. Springback prediction is important to control the process and die design to achieve the desired shape of the product. The prediction of bend force provides a base to the designer for the die design and press selection. Traditional trial and error methods are time consuming and expensive, while the development of theoretical models for springback/bend force is difficult and cumbersome due to the complexity of sheet bending process. Consequently, an empirical model developed based on experimental research is more useful in industrial applications. The empirical modeling techniques namely response surface methodology (RSM) and artificial neural network (ANN) become promising tools for prediction of springback and bend force because of their robustness and predictability Very few literatures is available for the prediction of springback and bend force in air bending using this prediction tools. Narayanasamy and Padmanabhan [3] developed a prediction model applying RSM, for bend force in air bending operation of interstitial free steel sheet. Forcellese et al [4] developed a neural network based punch displacement control system in air bending process of AA5754 aluminium alloy sheets to control springback Inamdar et al [5] discussed the development of ANN model for predicting springback in air vee bending of metallic sheets. Garcia Romeu and Cirauna [6] described the application of neural networks for the prediction of springback, punch displacement and final bending radius. Narayanasamy and Padmanabhan [7] compared the neural network model with regression model for the prediction of springback of interstitial steel sheets in air bending process. Fu et al [8] developed a back propagation neural network model to predict the springback for air bending of high-

Index Terms— Artificial neural networks, Response surface methodology, air bending, electrogalvanised steel;, springback, bend force.

I. INTRODUCTION Sheet metal forming is one of the major near-net-shape processes for the manufacturing of automobile, aircraft, electrical components etc. Sheet metal bending, being an important sheet metal forming process, employs a press brake with proper tooling to produce different bend components. The ûexibility of the bending process is improved by the air bending technique [1]. Flexibility is achieved in such a way that the different bend angles can be produced by solely controlling the punch travel into the die without the need for changing tool sets. Because of this versatility, the lead time is reduced in air bending technique and hence it is commonly used in sheet metal industries. Steel sheets are widely used to manufacture various parts in sheet metal industry, but the susceptibility to corrosion is their natural weakness. To overcome this, the uncoated steel sheets are substituted with galvanised steel sheets based on the improved corrosion resistance. Since electrogalvanised (EG) steel sheets have better formability [2] and surface quality, they are much preferred than hot dipped, in many of the industrial applications. The two important subjects related to bending process © 2012 AMAE DOI: 03.MES.2012.2. 2

73

Short Paper Proc. of Int. Conf. on Recent Trends in Mechanical, Instrumentation and Thermal Engineering 2012 TABLE I. THE SCHEME O F LEVELS AND VALUES OF PARAMETERS

strength sheet metal. The model is based on orthogonal test with input parameters such as sheet thickness, tool gap, punch radius, ratio of yield strength to Young’s modulus and punch displacement. The model yields satisfactory result in sheet metal air bending of a workpiece used as crane boom. In this paper, development of RSM and ANN models for predicting both springback and bend force has been attempted and the models are compared based on their prediction performance. II. EXPERIMENTAL DETAILS The input-output database required for the development of predictive models is obtained through bending experiments. The process of selecting appropriate set of data points is in such way that it maximises the accuracy of the prediction is known as design of experiments (DoE). In the present study, based on the previous work done on this field by various researchers [9-19], seven important and independently controllable process parameters which strongly influence the springback and bend force have been identified. They are strain hardening exponent, coating thickness, die opening, die radius, punch radius, punch travel and punch velocity. The design matrix considered in this study is a five level central composite rotatable design (CCRD) consists of half replication of 27(128/2 = 64) design points plus 10 center points and 14 (2X7) star points. The rotatability value is calculated by taking the fourth root of number of design points . The upper level of a factor is coded as +2.82 and the lower level as -2.82 for designing the experiments. The intermediate coded values are obtained from the equation as Coded value of the parameter value i

xi xi 0 di

The experimental setup for bending is shown in Fig.1. The punch was mounted in the upper arm of UTM and the die was placed on the lower platform. The sheet blanks were positioned on the die with necessary care. The punch traveled to the required depth for bending the sheet and the digital meter attached to the UTM was used to measure the punch travel. The larger edge of the bent sample was coated with black ink and the impression of the profile was taken carefully on a thick white paper supported by a board. Two impressions were taken before unloading and after unloading. The impressions of the sheet were scanned and converted into digitised images. The digitised images were imported to CAD software and the lines were drawn on the edges of the legs of images using the software. The necessary angles were measured using CAD software [18]. The difference between the bend angles during loading (θi) and after unloading (θf) gives the springback angle ( Δθ). The bend force was measured by the digital display of load cell arrangement. The sheets were bent to different depths by controlling the punch travel.

(1)

where xi is the actual parameter value, xi0 is the parameter value corresponding to zero level and di is the incremental parameter value. The actual and coded values of the process parameters are listed in Table 1. The substrate used in this study is Aluminium Killed Draw Quality (AKDQ) steel sheet. AKDQ steel sheet is a widely used sheet metal in automotive, electrical and other domestic applications because of its excellent formability and low cost. The steel sheet was electro galvanised for various coating thickness such as 4 µm, 7µm, 10µm and 14µm. The blanks from the steel sheets were cut to the required dimensions of 120 mm x 40 mm x 1mm and the edges were cleaned to remove the burrs. As the strain hardening exponent n values (0.211, 0.219, 0.232, 0.206 and 0.227) belong to five different orientations 0o, 22.5o, 45o, 67.5o and 90o respectively, the blanks were prepared in the five directions. The experiments were conducted in a 40T Universal Testing Machine (UTM) and the tooling (dies and punches) was made of hardened steel.

© 2012 AMAE DOI: 03.MES.2012.2. 2

74

Short Paper Proc. of Int. Conf. on Recent Trends in Mechanical, Instrumentation and Thermal Engineering 2012 TABLE II. ANOVA FOR RSM BASED SPRINGBACK AND BEND FORCE MODELS

Rp – Punch Radius ; Wd – Die Opening; θi – Bend Angle (During Loading) θf – Bend Angle (After Unloading) Δθ – Springback Angle

The experimental data were analysed to obtain the coefficients of polynomials using MINITAB 15 software. The experimental data were analysed to obtain the coefficients of polynomials using MINITAB 15 softwar The p value approach is used for testing the significance of coefficients. According to this, the coefficients are tested at 0.05 level of significance (95% confidence The analysis of variance technique (ANOVA) is employed to check the adequacy of the developed models at a 95% confidence level. As per this technique, the calculated value of F ratio for the lack-of-fit must be lesser than its standard value for a desired level of confidence level. Here, both the springback and bend force models are adequate in non-linear form, as the F ratio of both the models is less than the standard value for 95% confidence level, as illustrated in Table II. The final mathematical models for the springback and bend force given in terms of factors, as Springback

Rd – Die Radius; t – Sheet Thickness

Figure 1 Experimental Setup

III. RSM MODELLING Response surface methodology adopts mathematical and statistical techniques to evaluate the relation between a cluster of controlled experimental factors and a response. RSM provides an approximate relationship between the response Y and the independent variables xi which is based on the ob-

Y1 151. 5028 1256 .9472 n 0. 7578t c 0.4065Wd 0.8244 R d 0.1744 R p 0.2558t p

served data from the process. The optimal response can be obtained from this methodology even when there is minimal information about the process.

Y f ( x1 , x 2 , x3 ,...x n )

5. 4963V p 2785 .0397 n 2 0.0071t c2 0.0019Wd2 0.0591R d2 0.0348 R p2 0. 006t 2p 6. 3553V p2 2.5333 nt c 1.0036 nWd 3.2095 nRd 1.4023 nR p 0.9879 nt p 0.005t cW d 0.0093t c Rd 0.0243t c R p 0.0021t c t p 0. 0362t cV p 0.0055Wd R d 0.0023Wd R p 0.003Wd t p 0.0118Wd V p 0.0169 R d R p 0.0088 Rd t p 0.0843 Rd V p 0.0121 R p t p 0. 0597 R pV p

(2)

(5)

where denotes an error component. As air bending is a non-linear process [6], a linear polynomial is unable to predict the response accurately, and therefore the second-order model (quadratic model) is found to adequately model the process [20]. Second order non-linear models have been developed to predict springback and bend force which are of the following form k

k

Bend force Y 51.41 299.629n 0.029t c 0.291Wd 0.363R d 0.059 R p 0.085t p 0.823V p 721.323n 2 0.007t c2 0.002Wd2 0.006 Rd2 0.002 R p2 0.005t 2p 0.310V p2 0.671nt c 0.289nWd 0.664 nRd 0.312 nt p 0.004t c R d 0.026t cV p 0.001Wd Rd 0.001Wd R p 0.003R d t p 0.011R p t p 0.039t pV p

(6) where n is strain hardening exponent, tc is coating thickness, Wd is die opening, Rd is die radius, Rp is punch radius, tp is punch travel and Vp is punch velocity.

k

Y 0 i xi ii x i2 ij xi x j (3) i 1

i 1

i, j

IV. ANN MODELLING

where Y Response, i.e., springback and bend force xi Coded values for parameters

A. ANN Approach An Artificial Neural Network (ANN) is an information processing system that is inspired by information processing of the biological nervous system (brain). ANN has a parallelly distributed architecture with a number of information processing elements called neurons which are interconnected by connection links. A multilayer neural network has an architecture that consists of an input layer, one or more hidden layers and an output layer. The input layer receives input data and after processing it, sends them to the hidden layer. The hidden layer processes the data and sends a response

β Regression coefficients The values of the regression coefficients as defined above are determined as:

b XTX

1

X TY

where b Matrix of parameter estimates X Matrix of model terms evaluated at n data

(4) points

Y Matrix of the measured response © 2012 AMAE DOI: 03.MES.2012.2.2

75

Short Paper Proc. of Int. Conf. on Recent Trends in Mechanical, Instrumentation and Thermal Engineering 2012 to the output layer. The output layer accepts the responses and produces the result. The network which has an information flow from input layer to output layer in forward direction is known as feed-forward network. Since the ‘intelligence’ of the network exists in the values of the weights between neurons, a method of adjusting the weights is needed to solve a particular problem. The process of modifying the weights in the connections between network layers with the objective of achieving the expected output is called ‘learning’ or ‘training a network’. The backpropagation (BP) algorithm is the most widely used learning algorithm. The principal steps in the learning process are given below: Step1: Initialise all the weights of the links to random values. Step2: Present the input - desired output pattern sets for updating the weights and bias. The equation for updating weights and bias is

w zji1 w zji w zji1

B. ANN Architecture As neural network is a non-linear analysis tool, it maps the non-linear relationships with multiple interactions between input variables and springback/bend force in air bending in a better way [4,5,7]. Hence, a multilayer perceptron with back propagation training algorithm, widely used for engineering applications, is considered as a practical choice for this work. Single hidden layer is used for many practical problems since it can approximate any function which contains a continuous mapping from one finite space to another [21]. Hence single hidden layer is selected and the architecture of the ANN in this work has a three layer structure of one input layer, one hidden layer and one output layer. The number of neurons in input and output layers is 7 and 2, corresponding to the number of inputs and outputs. As a heuristic rule, the number of hidden layer neurons can be up to 2n+1 where n is the number of neurons in the input layer [22]. In this study, the number of neurons in the hidden layer is changed and the Mean Square Error (MSE) is evaluated. After several trials, the number of neurons which results least MSE is selected for hidden layers and is 14. The designed architecture becomes 7-14-2. The designed ANN architecture is shown in Fig. 2. Levenberg-Marquardt (LM) learning algorithm, the fastest method for training moderate-sized feed forward neural networks [23] is an efficient method for nonlinear problems [24]. This makes LM algorithm the much preferred one in this work. The transfer function is required to introduce the nonlinearity characteristics into the network. Log-sigmoid function takes any real valued input and returns an output bounded between [0,1]. In backpropagation, it is important to calculate the derivatives of any transfer functions used. As it is differentiable, the function log-sigmoid is commonly used in back propagation networks [25]. Hence, log-sigmoid was chosen as the activation function for the present network.

(7)

where w ji Weight of the link connecting neuron j to i, z is z 1

the learning step and w ji is the incremental change in the weight. The weight change is determined by a learning algorithm. Based on LM learning algorithm, the weight change can be written as

w zji1 J T J I

1

J T e w zji w zji1

(8)

where J is Jacobean matrix that contains first derivatives of the network errors with respect to the weights and biases, μ is the adaptive training parameter, I is the identity matrix, e is the vector of network errors and α is the momentum term. The use of a momentum term in the backpropagation algorithm is primarily to overcome the possible trap to local minima and also to overcome the oscillations during the training of the network. Moreover, the use of a small momentum helps in increasing the network training convergence with the LM algorithm. Step 3: After presenting the patterns, compute the mean square error (MSE) of all outputs as:

MSE

1 QK

Q

K

d

y kq

2

kq

(9)

q 1 k 1

where Q is the number of sets of input-output data K is the total number of outputs

Step 4: If MSE MSE t arg et

then stop, else go to

Figure 2 Architecture of the present ANN

step 2. If the new MSE is smaller than that of the preceding, then reduce the training parameter µ by µ-. If the MSE is increased, then increase µ by µ+. After the training stage, the ANN learning performance is evaluated by running it using the test data set, which is not included in the training set. If the network responds correctly to the inputs, the generalisation of network is confirmed. © 2012 AMAE DOI: 03.MES.2012.2. 2

C. ANN Training The experimental design matrix of 88 bending experiments along with the experimental results of springback and bend force was used as the training set for the network. As the dimension and magnitude of the testing or training data are different, they should be normalised before applying in the network, so that they lie between 0 and 1. The 88 inputoutput pairs were fed into the computer and a computer pro 76

Short Paper Proc. of Int. Conf. on Recent Trends in Mechanical, Instrumentation and Thermal Engineering 2012 gram was performed in the MATLAB 7.0.1 software. The training was performed in batch mode with the following learning factors: Learning rate = 0.1; Momentum constant = 0.5; Target for MSE =0.0000001; Maximum number of epochs = 1000. During training, the network compares its predicted value to the actual output and adjusts all the weights to improve the model. Once the MSE of the raining data reached the target value, the training is terminated and the weights and biases are automatically saved by the program. The MSE target is achieved in 292 epochs and the variation of MSE during the training is shown in Figure 3.

V. COMPARISON OF RSM AND ANN MODELS The developed RSM and ANN predictive models were compared on the basis of their prediction accuracy. The measured values of springback and bend force from the 15 test experiments and the corresponding values of the ANN and RSM models were compared to test the interpolation of the prediction models. The accuracy of the developed models is assessed by means of absolute error function and it has been defined as

Vm Vexp % Absolute error =

Vexp

100

(10)

where Vm is the predicted value of the model and Vexp is the measured experimental value. Figures 4 and 5 illustrate the comparison of the error profiles for springback and bend force respectively, for 15 test data points. The average absolute error for springback and bend force in ANN model is 1.24% and 1.41% while for RSM model it is 3.50 % and 3.92% respectively. Despite the error being within the acceptable level for both the modeling techniques, ANN outperforms the RSM technique. VI. DISCUSSIONS Though the researchers [4-8] developed ANN models in air bending earlier, they concentrated mainly on the prediction of springback and not on bend force, which is another important parameter to be considered in air bending. The important parameters like die opening and die radius which are having major influence on springback and bend force have not been considered by Narayanasamy and Padmanabhan [7]. The main advantage of the present ANN is that it can be used as a computer aided tool for providing easy, quick and precise predictions in process design in air bending of electrogalvanised steel steel. This work is a new attempt on electrogalvanised steel sheet and it can be further extended for other types of coated steel sheets also.

Figure 3. Variation of MSE during training the network

CONCLUSIONS

Figure 4. Comparison of error profiles of ANN and RSM for Springback

This study described a comparative analysis of the two modeling approaches namely artificial neural network (ANN) and response surface methodology (RSM) for the prediction of springback and bend force in air bending of electro galvanised steel. A five level central composite design of experiments was employed to create the input-output data pairs for developing both the models. The ANN model was based on a multilayer feed forward topology and trained with LM back propagation algorithm. The performance of the models was compared based on their prediction accuracy using a new set of 15 input-output pairs. The performance of the ANN model for predicting springback and bend force was found to be better than RSM models.

Figure 5. Comparison of error profiles of ANN and RSM for Bend force

© 2012 AMAE DOI: 03.MES.2012.2. 2

77

Short Paper Proc. of Int. Conf. on Recent Trends in Mechanical, Instrumentation and Thermal Engineering 2012 REFERENCES

[12] C.Bruni, A.Forcellese, F.Gabrielli, and M.Simoncini, “Air bending of AZ31 magnesium alloy in warm and hot forming conditions”, J. Mater. Process Technol., vol.177, No.1-3, pp.373-376, 2006. [13] D.Fei,. and P.Hodgson, “Experimental and numerical studies of springback in air v-bending process for cold rolled TRIP steels”, Nucl. Eng. Des., vol.236, No.18, pp.1847-1851, 2006. [14] M.L.Garcia-Romeu, J.Ciurana, and I.Ferrer, “Springback determination of sheet metals in air bending process based on an experimental work”, J. Mater. Process Technol., vol.191, pp.174-177, 2007. [15] S.Y.Kim, W.J. Choi, and S.Y.Park, “Spring-back characteristics of fiber metal laminate (GLARE) in brake forming process”, Int. J. Adv. Manuf. Tech., vol. 32, No.5-6, pp. 445-451, 2007. [16] R.Narayanasamy, and P. Padmanabhan, “Effects of material orientation on air bending of interstitial free steel sheet”, J. Manu. Eng., vol.3, No.3, pp.163-169, 2008. [17] K.Yilamu, R.Hino, H.Hamasaki, and F.Yoshida, “Air bending and springback of stainless steel clad aluminium sheet”, J. Mater. Process. Technol., vol. 210, pp.272-278, 2010. [18] D.Vasudevan, R.Srinivasan, and P.Padmanabhan, “Effect of process parameters on springback behaviour during air bending of electro galvanised steel sheet”, J. Zhejiang UniversityScience A, vol 12, No.3, pp.183-189, 2011. [19] R.Srinivasan, D.Vasudevan, and P.Padmanabhan, “Experimental investigations of bend Force in air bending of electrogalvanized steel sheets”, J. Mater. Engg. Perform., DOI: 10.1007/s11665-011-0029-7. [20] K.C.Kapoor, and Q.Feng, , “Statistical method for product and process improvement”, in Springer hand book of engineering statistics, Pham, H. Ed. Springer-Verlag, London, 2006. [21] K.Hornik, M.Stinchcombe and H.White, “Multilayer feedforward networks are universal approximators”, Neural Networks, vol. 2, No. 5, pp.359-366, 1989. [22] R.Z.Saen, “The use of artificial neural networks for technology selection in the presence of both continuous and categorical data”, World Appl. Sci. J., vol. 6, No. 9. pp.1177-1189, 2009. [23] M.T.Hagan, and M.B.Menhaj, “Training feedforward networks with the Marquardt algorithm”, IEEE Trans Neural networks, vol 5, No.6, pp.989-993, 1994. [24] M.Sanjari, A.K.Taheri, and M.R.Movahedi, “An optimization method for radial forging process using ANN and Taguchi method”, Int. J. Adv. Manuf. Technol., vol.40, No. 7-8, pp. 76–784, 2009. [25] M.T.Hagan, H.B. Demuth, and M.H.Beale, Neural network design, PWS Publishing company, Boston, 1996.

[1] L.J. de Vin , “Curvature prediction in air bending of metal sheet”, J. Mater. Process Technol., vol. 100, pp. 257-261, 2000. [2] H.M. Jiang, X.P.Chen, C.Wu, and Li, H.H. “Forming characteristics and mechanical parameter sensitivity study on pre-phosphated electrogalvanised sheet steel”, J. Mater. Process Technol., vol. 151, pp. 248-254, 2004. [3] R.Narayanasamy, and P.Padmanabhan, “Application of response surface methodology for predicting bend force during air bending process in interstitial free steel sheet”, Int. J. Adv. Manuf. Technol., vol.144, No.1–2, pp. 38–48, 2009. [4] A.Forcellese, F.Gabrielli, and R.Ruffini, “Effect of the training set size on springback control by neural network in air bending process”, J. Mater. Process. Tech., vol. 80-81, pp. 493-500, 1998. [5] M.Inamdar, P.P.Date, K.Narasimhan, S.K.Maiti, and U.P.Singh, “Development of an artificial neural network to predict springback in air vee bending”, Int. J. Adv. Manuf. Technol., vol. 16, pp. 376-381, 2000. [6] M.L.Garcia-Romeu, and J.Ciurana, “ Springback and geometry prediction – Neural networks applied to the air bending process”, Huang, D.S., Li, K., and Irwin, G.W. Eds. ICIC 2006, LNCS 4113, pp. 470–475, Springer-Verlag, Berlin Heidelberg, 2006. [7] R.Narayanasamy, and P.Padmanabhan, “Comparison of regression and artificial neural network model for the predicting springback during air bending process of interstitial free steel sheet”, J. Intell. Manuf., DOI: 10.1007/s10845-009-0375-6, 2010. [8] Z.M.Fu and J.H.Mo, “Springback prediction of high-strength sheet metal under air bending air bending forming and tool design based on GA-BPNN”, Int. J. Adv. Manuf. Technol., vol. 53, Nos. 5-8, pp. 473-483, 2011. [9] C.Wang, G. Kinzel, and T.Atlan, “Mathematical modeling of plane-strain bending of sheet and plate”, J. Mater. Process. Technol., vol. 39, pp. 279-304, 1993. [10] Y.M.Huang, and D.K.Leu, “Effects of process variables on V-Die bending process of Steel Sheet”, Int. J. Mech. Sci., vol.40, No.7, pp.631–650, 1998. [11] M.V.Inamdar, P.P. Date,. and S.V.Sabnis, “On the effects of geometric parameters on springback in sheets of five materials subjected to air vee bending”, J. Mater. Process Technol., vol.123, No.3, pp.459-463, 2002.

© 2012 AMAE DOI: 03.MES.2012.2. 2

78