期刊名称：International Journal of Computer Science and Network Security
印刷版ISSN：1738-7906
出版年度：2009
卷号：9
期号：4
页码：123-137
出版社：International Journal of Computer Science and Network Security
摘要：In a hard disk media manufacturing, engineers rely on inspection machine to generate production yield temporal data that can be used for future analysis. To proactively perform process maintenance on the equipment in order to avoid unnecessary unplanned down time, they have to be able to predict the yield outcome before products arrive at the inspection machine. In this paper, we propose to predict the yield outcome by visualizing the historical data pattern generated from the inspection machine, transform the data pattern and trained it with machine learning algorithms. The trained visualized datasets can automatically generate a prediction model without the visual interpretation needs to be done by human. However, due to the nature of manufacturing process, majority class instances of the good yield are extremely outnumbers minority class instances of the bad yield. Comparison between the random under-sampling, over- sampling, and SMOTE + VDM sampling technique indicate that the sampling combination of SMOTE + VDM and random under-sampling dataset produced a robust classifier performance. Furthermore, the integration of K* entropy base similarity distance function with SMOTE, CNN+Tomek, and our novel SMOTE and SMaRT combination, extend the improvement of the classifiers F-Score robustness. Experimental results have indicated that the proposed approaches are viable to be applied to generate a predictive model, hence promoting the implementation of predictive maintenance in hard disk media industries.
关键词：Yield prediction; Predictive Maintenance; Pattern visualization; Data re-sampling; Robust classifier
Loading...
