Optimization of surface appearance defect reduction for alumina substrate using design of experiment and data mining technique

Abstract

This paper presents an integrated application of design of experiments (DOE) with support vector machine (SVM) for manufacturing process modeling in order to achieve a high accuracy model. The proposed method is as follows. First, DOE is applied to indicate the critical parameters of the process. Due to the nature of defective distribution is binomial; fundamental DOE assumptions may be violated. Consequently, Freeman and Turkey (F&T) transformation was applied to the percentage of the surface appearance defect. Then the residual analysis was opted for model adequacy checking. Last but not least, the response surface methodology (RSM) model is sufficient following DOE assumptions. Then, support vector regression (SVR) was used to establish the nonlinear multivariate relationships between process parameters and responses. Data obtained from designed experiments were used in the training process. Last a grid search was adopted to the SVR model to find the optimum parameter setting. Data from real experiments of the powdering process parameters for alumina substrate sheet for product A were used to demonstrate the proposed method. Other prominent approaches, namely RSM data and artificial neural networks (ANN) learning with quick propagation algorithm (Quickprop), were conducted for comparison purposes. The experimental results suggested that the SVR was capable of high accuracy modeling and resulted in much smaller error in comparison with the results from ANN learning with quick propagation algorithm and full factorial. After searching the optimum condition from the SVR model, it was 6.5 cm of distance and the minimum level of other factors. After performing verification runs, it can be efficiently employed to reduce the percentage of surface appearance defect from 5.8% to 4.0 %. As a result, the direct cost of company has been cut down by approximately $5,200 per month from the enchanting operation of powdering machines for alumina substrate sheet. © Springer-Verlag Berlin Heidelberg 2010.