Back Propagation neural network modeling for warpage prediction and optimization of plastic products during injection molding
Fei Yin,Huajie Mao,Lin Hua,Wei Guo,Maosheng Shu
DOI: https://doi.org/10.1016/j.matdes.2010.12.022
IF: 9.417
2011-01-01
Materials & Design
Abstract:Warpage of plastic products is an important evaluation index for Plastic Injection Molding (PIM). A Back Propagation (BP) neural-network model for warpage prediction and optimization of injected plastic parts has been developed based on key process variables including mold temperature, melt temperature, packing pressure, packing time and cooling time during PIM. The approach uses a BP neural network trained by the input and output data obtained from the Finite Element (FE) simulations which are performed on Moldflow software platform. In addition, a kind of automobile glove compartment cap was utilized in this study. Trained by the results of FE simulations conducted by orthogonal experimental design method, the prediction system got a mathematical equation mapping the relationship between the process parameter values and warpage value of the plastic. It has been proved that the prediction system has the ability to predict the warpage of the plastic within an error range of 2%. Process parameters have been optimized with the help of the prediction system. Meanwhile energy consumption and production cycle were also taken into consideration. The optimized warpage value is 1.58 mm, which is shortened by 32.99% comparing to the initial warpage result 2.358 mm. And the cooling time has been decreased from 20 s to 10 s, which will greatly shorten the production cycle. The final product can satisfy with the matching requirements and fit the automobile glove compartment well. Keywords A. Polymers C. Molding F. Defects 1 Introduction Plastic Injection Molding (PIM) is an advanced technique to process plastic products. As one of the most significant evaluation indexes of the quality of plastic articles, warpage of plastic during the injection molding process has attracted more and more attentions. Many researches have been carried out to reveal the relationships between process variables and warpage of the plastics to reduce the warpage of injected plastic products. In 2010, Deng et al. proposed a hybrid optimization method that combines a Mode-Pursuing Sampling (MPS) method with the genetic algorithm for minimizing the warpage of injection molded plastic parts. And the warpage value of a food tray plastic part was minimized by using the proposed method [1] . Farshi et al. demonstrated an optimization method for injection molding process parameters by using sequential simplex algorithm. And they minimized the warpage as well as the shrinkage of a kind of automotive ventiduct grid by the proposed method [2] . Shi et al. proposed an adaptive optimization method to minimize the warpage of the injection molding parts based on neural network [3] . Kurtaran et al. considered mold temperature, melt temperature, packing pressure, packing time and cooling time as the key process parameters during PIM. And they got the optimum process parameters resulting in the minimum warpage value of a bus ceiling lamp base by using artificial neural network and genetic algorithm [4] . Ozcelik and Sonat performed the warpage analysis on the Computer-Aided Engineering (CAE) software Moldflow platform and they found that packing pressure is the most influential parameter on the warpage of Polycarbonate/Acrylonitrile Butadiene Styrene (PC/ABS) material by using Taguchi experimental method [5] . Huang and Tai also indicated that packing pressure has the greatest influence on the warpage of plastic parts, followed by mold temperature, melt temperature, and packing time, while the warpage is only slightly influenced by the gate dimension and the filling time in thin shell injection molding. And the optimum values for those parameters were revealed [6] . Kong and Kim studied the effects of processing conditions, package geometry and materials on the occurrence of the warpage, and they pointed out that a lower molding temperature and a smaller coefficient of thermal expansion can significantly reduce the warpage value of the plastic parts [7] . Erzurumlu and Ozcelik minimized the warpage of the plastic parts by using Taguchi optimization method, the signal-to-noise ( S / N ) ratio and the analysis of variance (ANOVA) [8] . Wang et al. minimized the warpage of the front panel of large Liquid Crystal Display (LCD) TV by the application of Taguchi experimental design [9] . Gao and Wang developed an effective optimization method by using the Kriging model to minimize the warpage value of plastic products during plastic injection molding [10] . All the researches cited beforehand accomplished their purpose of optimizing process parameters during PIM and minished the warpage value of the parts greatly via a series of optimization methods. However, these researches mainly focus on the procedure of optimization, a clear and reliable prediction model that could be used as the surrogate model during process optimization as well as a systematical procedure of the establishment of the Back Propagation (BP) neural network based warpage predictor model are still required. In addition, previous researches just optimized the warpage of the parts, while the energy consumption and the production cycle greatly affecting the economic benefits during PIM were rarely taken into consideration. In this study, mold temperature, melt temperature, packing pressure, packing time and cooling time were considered to be the critical process parameters that can greatly influence the warpage of the plastics during PIM referring to the cited literatures. Finite Element (FE) simulations were carried out under Moldflow software environment. And a mathematical function mapping the relationship between warpage value and process parameters was acquired by using BP artificial neural network and the experimental data obtained from the FE simulations. Warpage of plastic articles can be predicted accurately and quickly by using the warpage predictor. The prediction model was then employed as a calculation platform. Meanwhile, energy consumption and production cycle were also taken into consideration as constraint conditions to determine process parameters during PIM. A better and more suitable combination of the process parameters can be got by changing the input of the prediction system. Although the best solution resulting in the minimum warpage value of the part cannot be revealed via this simple method, a better and more suitable solution can be obtained to satisfy the demands of the manufacture. And a case study confirmed the validity and practicality of the proposed method. 2 BP neural-network model for warpage BP neural network has the powerful ability of non-linear interpolation. It can obtain the mathematical mapping reflecting the internal law of the experimental data [11–14] . Hence, it has been widely used in engineering application for prediction and optimization [15–18] . In 2000, Cook et al. developed a neural-network model to predict the value of a critical strength parameter in a particleboard manufacturing process [19] . Sadeghi predicted the quality of the injected parts by using the neural network [20] . Altan minimized the shrinkage of rectangular-shaped specimens by Taguchi, experimental design and the analysis of variance method. And neural network was employed to predict the shrinkage of the part. They suggested that neural network is an efficient predictive tool for shrinkage [21] . Lotti et al. developed a neural-network model to predict the shrinkage of an injection molded plaque, and they suggested that the artificial neural network has the best performance in the shrinkage prediction, even using limited experimental data, confirming its great capacity to model non-linear systems [22] . Referring to the cited literatures, a three-layer BP neural-network model is designed by using the Matlab neural network toolbox in this paper. And the structure of the BP neural network can be seen in Fig. 1 . Fig. 1 shows the structure of the BP neural network. The network consists of one input layer with five neurons standing for mold temperature, melt temperature, packing pressure, packing time and cooling time, respectively; two hidden layers with 20 neurons each and one output layer having one neuron representing warpage value of plastics. The input of each neuron comes from the output of the neurons contained in the preceding layer by the transition function shown as follow: (1) net i = ∑ j = 0 N ω ij x j where net i denotes the total input of the i th neuron in the computing layer; N denotes the number of the neuron in the forward layer; ω ij represents the connection weight of the jth neuron in the forward layer and ith neuron in the computing layer and x j denotes the output of the jth neuron in the forward layer. The ith neuron in the computing layer produces its output out i by processing the input through a transfer function f s , the tangent hyperbolic function chosen in this study as below: (2) out i = f s ( net i ) = 1 - e - net i 1 + e - net i The system should be trained by the experimental data before it can predict the warpage of plastics. During training process, the connection weight ω ij is calculated by minimizing the error between the predictive value and the actual value. In this paper, a kind of automobile glove compartment cap to be assembled on the automobile glove compartment is taken for an example, and the warpage of this plastic part is expected to be predicted accurately and quickly to instruct the production. 3 FE simulation of automobile glove compartment cap 3.1 3D FE model The automobile glove compartment cap utilized in this study is shown in Fig. 2 . Its width, length and thickness are 230 mm, 350 mm and 2.5 mm, respectively. The material of the glove compartment cap is polypropylene (PP). And the material mode of PP with the trade name of BP Amoco 1046 and manufactured by BP Chemicals which is from the library of the Moldflow software database was employed as the material of the part during simulations. The detailed material properties are given in Table 1 . The FE model of the automobile glove compartment cap is established under Moldflow software environment. Moldflow is good at solving pressure, flow and temperature fields during PIM [23] . The FE model includes 37,999 triangle elements, mesh condition and the filling system as well as the cooling system can be seen in Fig. 2 . 3.2 FE analysis FE analysis of the automobile glove compartment cap was performed and the parameters values during the analysis were set as shown in Table 2 . During the simulation, injection process was controlled by the injection time which has been set as 1s in this study. And the injection pressure was calculated by the computer automatically. The maximum injection pressure is 148.11 MPa, occurring at the transition point between injection process and packing process. Packing pressure was set as a percentage of the maximum injection pressure during packing process. The analysis took about 2500 s of CPU time on HP personal workstations. And the results of the warpage analysis of the glove compartment cap can be seen in Fig. 3 . Fig. 3 illustrates the warpage analysis result of the glove compartment cap. According to the result, the warpage of the part can be seen directly from the picture. The maximum warpage value of this part is 2.358 mm, occurring at the place far from the gates. 4 Warpage prediction system 4.1 Generation of training data To calculate the connection weights between each neuron, the designed Back Propagation 5-20-20-1 neural network should be trained by a group of experimental data. And the experimental data should reflect the inner relationship between the warpage and process parameters objectively and globally. A full factorial experimental design is no doubt an excellent idea, but that means there will be a large number of simulations, that is 243, should be taken, which will obviously consume much time. With the features of level uniformity and vertical uniformity of experimental data, the orthogonal experimental design method can reflect the inner law between the input data and output data globally and objectively, and at the same time the number of the CAE analysis is sharply declined. Consequently, the orthogonal experimental design method is used to conduct the FE simulations. Each experimental factor range is divided into levels between the lowest and the highest values. The setting of the processing parameters can be seen in Table 3 . 4.2 Training of the neural network The purpose of the network training is to calculate the connection weights to minimize the mean square error between the network prediction data and experimental data. At the beginning of the training, the weights are given at random, and then they are iteratively updated until convergence to the certain values by using the gradient descent method. The iterative procedure of the connection weights is set as below: (3) ω ij new = ω ij old + Δ ω ij (4) Δ ω ij = - η ∂ E ∂ ω ij out i where ω ij new denotes the contemporary connection weights of the net; ω ij old represents the previous connection weights of the net; E is the mean square error; outi is the output of the ith neuron; η is the learning rate parameter controlling the convergence rate of the net. Generally, the learning rate η is a constant between 0 and 1 and it was taken as 0.005 in this study. Training process takes about half an hour of CPU time on HP personal workstations for 900,000 epochs and the mean square error of the training data is set as 0.0001. Fig. 4 illustrates the training progress of the network. Fig. 4 shows the training process of the BP neural network. The training curve illustrates that with the updating of the connection weights, the mean square error between the network prediction data and training data declines gradually and converges to 0.0001 interminably within 9000,000 epochs. The mathematic mapping between the values of processing parameters and warpage value is stored in the trained net, and the mathematic function can be seen as follow: (5) W = f l ∑ w 3 f s ∑ w 2 f s ∑ w 1 X where W is the warpage value of the glove compartment cap; X = [ x 1 , x 2 , x 3 , x 4 , x 5 ]′ is the matrix consists of the values of mold temperature, melt temperature, packing pressure, packing time and cooling time; f l is the liner transfer function between hidden layer 2 and output layer; f s is the transfer function between input layer and hidden layer 1, as well as hidden layers 1 and 2; w 1 , w 2 , w 3 represent the connection weights between input layer and hidden layer 1, hidden layer 1 and hidden layer 2, hidden layer 2 and output layer, respectively. 4.3 Testing of the prediction system To test the accuracy of the neural network prediction system, 30 groups of process parameters not used in the training process are used in the testing. As shown in Fig. 5 , most of the predictive values can agree with the numerical experimental values very well. And a detailed comparison between the neural network prediction system and FE simulations can be seen in Table 4 . Table 4 shows the comparison between the neural network prediction system and FE simulations from the respect of running time and the warpage results. The average running time of the FE simulation is about 2500 s and that of neural network prediction system is only about 0.001 s; In addition, the forecast error of the predictive system is defined as follow: (6) F e = ( W P - W F ) / W F × 100 % where F e represents the forecast error of the prediction system, W P is the prediction value of the warpage, W F stands for the experimental value of the warpage. It can be seen from Table 4 that the forecast errors are ranging from −1.32% to 1.59%. The prediction system is in good agreement with the FE analysis, and the running time of the prediction system is sharply decreased comparing to that of the FE analysis. The efficiency of design and manufacture of the products can be dramatically improved by using the prediction system. 5 Parameter optimization using the prediction system For its high prediction accuracy and rapid response, the warpage prediction system was used as the surrogate of the simulative analysis to optimize the process parameters during PIM. Fig. 6 illustrates the technique line to optimize the process parameters during PIM. During the optimization, energy consumption, mainly influenced by packing pressure and packing time, and the production cycle, greatly determined by packing time and cooling time during PIM, were taken into consideration. The constraint functions employed in the optimization model can be formulated as follows: (7) P p ⩽ Upper p T p ⩽ Upper tp T c ⩽ Upper tc where P p represents the packing pressure; T p is the packing time; T c denotes the cooling time; Upper p , Upper tp and Upper tc represent the upper limit value of the packing pressure, packing time and cooling time, respectively. Here, Upper p , Upper tp and Upper tc were set as 120, 10 and 10, respectively. The process parameters optimized in this method are shown in Table 5 . And the predictive warpage value is 1.552 mm. Comparing with the process parameters recommended by the Moldflow software, the optimized process parameters have a lower mold temperature and a higher melt temperature. Meanwhile, a higher packing pressure and a longer packing time are needed to get a lower contractions and lower warpage. And the cooling time has been greatly decreased, which is from 20 s to 10 s, because of the consideration of the production cycle. A detailed comparison between the optimized process parameters and the recommended process parameters can be seen in Table 6 . It can be seen from Table 6 that the melt temperature, packing pressure and packing time of the optimized process parameters have been increased by 8.33%, 20% and 25% than that of the recommend process parameters, respectively. And on the other hand the mold temperature and the cooling time of the optimized process parameters have been decreased by 33.33% and 50% than that of the recommend process parameters, respectively. The optimized process parameters were then used in the FE simulation under software Moldflow environment. The warpage analysis result (shown in Fig. 7 a) is 1.580 mm, which is shortened by 32.99% comparing to the initial warpage result 2.358 mm (shown in Fig. 3 ). And Fig. 7 b illustrates the final product produced by using the optimized process parameters with the injection machine. With the instruction and control of the prediction system, the process parameters were optimized to decrease the warpage value of the part, meanwhile the energy consumption and production cycle were also restricted in a certain range from the economical point of view. Although the best solution resulting in the minimum warpage value of the part cannot be revealed through this simple optimization method, a better and more suitable solution can be obtained to satisfy the demands of the manufacture and engineering application. Fig. 8 shows the final product assembled on the automobile glove compartment on the assembly line. It can be seen from Fig. 8 that the final product can satisfy with the matching requirements and fit the automobile glove compartment well, which confirms the validity and practicality of the proposed method. 6 Conclusions In this study, a BP neural network predictor model is developed to predict and minimize the warpage of plastic during PIM. The following conclusions are drawn: (a) A Back Propagation 5-20-20-1 net is designed by using the Matlab neural network toolbox. Taking a kind of automobile glove compartment cap for example, the designed network is trained by the experimental data obtained from FE simulations. A mathematic function reflecting relationship between the process parameters values and warpage values of the plastic part is contained in the neural network predictive system. The neural network prediction system can predict the warpage of the plastic within 0.001s and the prediction error is less than 2% on average. (b) With the instruction and control of the prediction system, the process parameters during PIM are optimized to minimize the warpage value of the part. Meanwhile, energy consumption and production cycle were also taken into consideration. The optimized warpage value is 1.580 mm, which is shortened by 32.99% comparing to the initial warpage result 2.358 mm. At the same time the cooling time has been decreased by 50% comparing to that of the recommended process parameters. And the final product can satisfy with the matching requirements and fit the automobile glove compartment well. (c) The prediction system can also be used as a reliable mathematical model for global optimization of the process parameters automatically in future because of its high prediction accuracy and rapid response. Acknowledgement The work was supported by a Grant from National Science Fund for Distinguished Young Scholars (No. 50725217). The supports are gratefully acknowledged. References [1] Y.M. Deng Y. Zhang Y.C. Lam A hybrid of mode-pursuing sampling method and genetic algorithm for minimization of injection molding warpage Mater Des 31 2010 2118 2123 [2] B. Farshi S. Gheshmi E. Miandoabchi Optimization of injection molding process parameters using sequential simplex algorithm Mater Des 32 2010 414 423 [3] H.Z. Shi Y.H. Gao X.C. Wang Optimization of injection molding process parameters using integrated artificial neural network model and expected improvement function method Int J Adv Manuf Technol 48 2010 955 962 [4] H. Kurtaran B. Ozcelik T. Erzurumlu Warpage optimization of a bus ceiling lamp base using neural network model and genetic algorithm J Mater Process Technol 169 2005 314 319 [5] B. Ozcelik I. Sonat Warpage and structural analysis of thin shell plastic in the plastic injection molding Mater Des 30 2009 367 375 [6] M.C. Huang C.C. Tai The effective factors in the warpage problem of an injection molded part with a thin shell feature J Mater Process Technol 110 2001 1 9 [7] W.Y. Kong J.K. Kim Warpage in plastic packages: effects of process conditions, geometry and materials IEEE Trans Compon Packag Manufact 26 2003 2345 2352 [8] T. Erzurumlu B. Ozcelik Minimization of warpage and sink index in injection-molded thermoplastic parts using Taguchi optimization method Mater Des 27 2006 853 861 [9] G.L. Wang G.Q. Zhao H.P. Li Y.J. Guan Influence factor analysis of warpage and optimization of process parameters in plastic injection molding of thin-wall plastic parts based on Taguchi optimization method China Mech Eng 20 2009 488 492 [10] Y.H. Gao X.C. Wang An effective warpage optimization method in injection molding based on the Kriging model Int J Adv Manuf Technol 37 2008 953 960 [11] D. Fogel C. Robinson Computational intelligence: the experts speak 2003 Wiley-IEEE Press [12] S. Russell P. Norving Artificial intelligence: a modern approach 2nd ed. 2004 Posts & Telecom Press Beijing [13] . He JM, Huang HX. Predicting processing parameters for water-assisted injection molding based on inverse GA-LMBP neural network. SPE Technical conference ANTEC; 2007. p. 676–9. [14] B.H. Min B.C. Shin A study on volumetric shrinkage of injection molded parts based on neural networks J Injection Molding Technol 5 2001 201 207 [15] M. Younesi M.E. Bahrololoom M. Ahmadzadeh Prediction of wear behaviors of nickel free stainless steel–hydroxyapatite bio-composites using artificial neural network Comput Mater Sci 47 2010 645 654 [16] Z.M. Fu J.H. Mo L. Chen W. Chen Using genetic algorithm-back propagation neural network prediction and finite-element model simulation to optimize the process of multiple-step incremental air-bending forming of sheet metal Mater Des 31 2010 267 277 [17] K. Dehghani A. Nekahi Artificial neural network to predict the effect of thermomechanical treatments on bake hardenability of low carbon steels Mater Des 31 2010 2224 2229 [18] W.X. Yu M.Q. Li J. Luo S.S. Su C.Q. Li Prediction of the mechanical properties of the post-forged Ti–6Al–4V alloy using fuzzy neural network Mater Des 31 2010 3282 3288 [19] D.F. Cook C.T. Ragsdale R.L. Major Combining a neural network with a genetic algorithm for process parameter optimization Eng Appl Artif Intell 13 2000 391 396 [20] B.H.M. Sadeghi A BP-neural network predictor model for plastic injection molding process J Mater Process Technol 103 2000 411 416 [21] M. Altan Reducing shrinkage in injection moldings via the Taguchi, ANOVA and neural network methods Mater Des 31 2010 599 604 [22] C. Lotti M. Ueki R.E.S. Bretas Prediction of the shrinkage of injection molded iPP plaques using artificial neural networks J Injection Molding Technol l6 2002 157 175 [23] T.A. Osswald L. Turng Injection Molding Handbook 2002 Carl Hanser Verlag Munich