Feature Selection to Improve Performance of Yield Prediction in Hard Disk Drive Manufacturing

Abstract

Hard Disk Drive (HDD) manufacturing is one real-world application area that machine learning has been extensively adopted for problem solving. However, most problem solving activities in HDD industry tackle on failure root-cause analysis task. Machine learning is rarely applied in a task of yield prediction. This research presents the application of machine learning and statistical techniques to select appropriate features to be used in yield prediction for the HDD manufacturing process. The seven well-known algorithms are used in the feature selection step. These algorithms are decision tree (C5 and CART), Support Vector Machine (SVM), stepwise regression, Genetic Algorithm (GA), chi-square and information gain. The two prominent learning algorithms, Multiple Linear Regression (MLR) and Artificial Neural Networks (ANN), are used in the yield prediction modeling step. Yield prediction performance has been assessed based on the two evaluation metrics: Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). Yield prediction with MLR shows higher accuracy than yield estimation traditionally performed by human engineers. Resulting to conclusion that the proposed novel learning steps can help HDD process engineers to predict yield with the better performance, especially on applying GA as feature selection tool, the MAE is reduced from 0.014 (yield estimated by human engineer) to 0.0059 (yield predicted by MLR). That means error reduction is about 60%.