เล่มประชุมกรรมการครั้งที่ 2_2564 แก้ไขเพิ่มเติม

359 13th Eco-Energy and Materials Science and Engineering Symposium, Udonthani,Thailand, 1-4 December 2016 16 2-10-1 tansig traingd 0.750 0.0061 Table IV The initial parameters for neural network model. 17 2-10-1 tansig traingdm -0.554 0.0194 TABLE IV. PARAMETERS OF NEURAL NETWORK MODEL. 18 2-10-1 tansig trainlm 0.882 0.0034 Training parameter Value According to Table V, the experimental 2-5-1 topology in Maximum number of epochs to train 1000 model number 1 to 9, was trained with traingd, traingdm and 10-9 Minimum value of performance function 10-7 trainlm, involving transfer function that is purelin, logsig and tansig. R value of model 9 has 0.978 and MSE is 0.0007. R of Minimum value of gradient, g 0.001 model 9 are higher than other models. In addition, MSE is less Initial factor parameter, µ 0.1 than others. In model 10 to 18 based on the 2-10-1 topology, R Decreasing factor of µ, + µ value of model 15 is 0.965 which is higher than other models. Increasing factor of µ, - µ 10 Moreover, MSE, 0.0005, is less than others. Maximum value of µ, max 1010 Figure 4 and 5 show observed yield and output yield of D. Perfomance of neural network model neural network models. Mean square error (MSE) is used to measure the 90.0% Observed and output yield in model 9 performance of neural network model [10]. It can be define as EMSE 1 N Outputi Observedi 2 (10) 80.0% N i1 70.0% Outputi 60.0% where Observedi is the i-th network output 50.0% N is the i-th target data is the total number of output nodes. 40.0% 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Actual Simulation IV. EXPEREMENTAL RESULT Fig. 4. Observed and output yield in model 9 There are two topologies of neural network model were Observed and output yield in model 15 trained. In this proposed method, transfer functions, purelin, logsig and tansig, were applied to the models. Consequently, the 90.0% ANN models included the training algorithms, traingd, traingdm and trainlm to adjust weights. The performance of the neural 80.0% network model can be measure by MSE. The correlation between the model outputs and observed data can be represented 70.0% in regression, R values. If R value of 1 means a close relationship then it 0 means a random relationship otherwise it is close to -1 60.0% which is no relationship. The experimental results are as shown in Table V. 50.0% 40.0% 123 45 678 9 10 11 12 13 14 15 16 17 18 Actual Simulation Fig. 5. Observed and output yield in model 15 TABLE V. EXPERIMENT RESULT OF NEURAL NETWORK MODEL V. CONCLUSIONS Model Topology Transfer Training Regression MSE In this paper, neural network models are applied to optimize Function Algorithm (Training) the mineral talcum and water in mixture design. There are 1 2-5-1 0.0144 presented 18 models. Model 9 has 2-5-1 topology which is 2 2-5-1 purelin traingd -0.258 0.0185 trained by Levenberg Maquardt algorithm and has transfer 3 2-5-1 purelin traingdm 0.617 0.0056 function in tansig function. The correlation in experimental 4 2-5-1 purelin trainlm 0.455 0.0069 results, model 9 and model 15 have a small difference of R 5 2-5-1 logsig traingd 0.276 0.0057 value, only 0.013. It shows that model 9 has close relationship 6 2-5-1 logsig traingdm 0.368 0.0155 than model 15. Considering the performance of neural network 7 2-5-1 logsig trainlm 0.317 0.0101 models between model 9 and model 15 having a small difference 8 2-5-1 tansig traingd -0.615 0.0271 MSE, however the performance of model 15 is better than mode 9 2-5-1 tansig traingdm 0.079 0.0007 9. As a result, the neural network models to optimize mineral 10 2-10-1 tansig trainlm 0.978 0.0081 talcum components for mixture design that is model 15. 11 2-10-1 purelin traingd 0.413 0.0094 12 2-10-1 purelin traingdm 0.893 0.0055 Implementing the neural network model, the model 9 and 13 2-10-1 purelin trainlm 0.491 0.0080 model 15 can provide the optimal components for mineral 14 2-10-1 logsig traingd 0.954 0.0183 talcum process that produce the mineral talcum pellet with 15 2-10-1 logsig traingdm -0.432 0.0005 machine in Uttaradit provice. It will increase accuracy and logsig trainlm 0.965 productivity. Moreover, it helps production plan and defines standard of mixture. 92

360 13th Eco-Energy and Materials Science and Engineering Symposium, Udonthani,Thailand, 1-4 December 2016 REFERENT [1] C. MACLEOD, G. DROR and . G. MAXWELL, \"Training Artificial Neural Networks Using Taguchi Methods,\" Artificical Interigence Review, pp. 177-184, 1999. [2] M. . S. Adroera, X. . L. Parrab, I. B. Corralc and J. . V. Calvetca, \"Indirect model for roughness in rough honing processes based onartificial neural networks,\" Precision Engineering, pp. 1-9, 2015. [3] A. Yadollahi, E. Nazemi, A. Zolfaghari and A. M. Ajorloo, \"Application of artificial neural network for predicting the optimal,\" Progress in Nuclear Energy, vol. 89, pp. 69-77, 2016. [4] S. Ozsahin, \"Optimization of process parameters in oriented strand board,\" Springer-Verlag Berlin Heidelberg, vol. 71, p. 769 777, 2013. [5] D. J. Sarma and S. C. Sarma, \"Neural Networks and their Applications in Industry,\" Bulletin of Information Technology, vol. 20, pp. 29-36, 2000. [6] neuralpower, \"Integreat Electricity Demand and Price Forecasting,\" [Online]. Available: http://www.neuralpower.com/technology. [Accessed 7 October 2016]. [7] P. SIB, S. A. JONES and P. SIDDARTH, \"ANALYSIS OF DIFFERENT ACTIVATION FUNCTIONS USING BACK PROPAGATION NEURAL NETWORKS,\" Journal of Theoretical and Applied Information Technology, vol. 47, no. 3, pp. 1264-1268, 2013. [8] A. Bordes, L. Bottou and P. Gallinri, \"SGD-QN: Careful Quasi-Newton Stochastic Gradient Descent,\" Journal of Machine Learning Research, vol. 10, pp. 1737-1754, 2009. [9] M. . I. A. Lourakis, \"A Brief Description of the Levenberg-Marquardt Algorithm Implemened,\" Institute of Computer Science, pp. 1-5, 2005. [10] D. Rasch, J. Pilz, . R. Verdooren and A. Gebhardt, \"Optimal Experimental Design with R,\" Journal of Statistical Software, vol. 43, 2011. [11] S. Prechadet and C. Luksiri, \"Prediction of Silicon wafer Lapping Time by Artificial Neural etwork,\" Engineering Journal Kasetsart, vol. 77, pp. 1-11, 2011. 93

361

362

363

364

365

366

367

368

369

370

371

372

373

374

375 ระเบยี บวาระที่ 6 เร่อื งอ่ืนๆ เพ่อื พจิ ารณา -ถา้ ม-ี . . ความเหน็ คณะกรรมการ . . .. . . . . . . . . . \\มติที่ประชมุ . . . . . กรรมการบัณฑิตศึกษา . . ปิดประชมุ เวลา น.