MACHINE LEARNING APPROACH FOR FEATURE SELECTION AND CLASSIFICATION OF HIGH-DIMENSIONAL DATA

Abstract

Maintaining high product yield and quality consistency is always the ultimate target in modern manufacturing operations. In order to investigate the correlations between sophisticated process parameters and the system yield, this thesis aims at developing advanced feature selection and classification approaches for high-dimensional data in manufacturing processes. This thesis starts with proposing an optimal feature selection method for high-dimensional data sets. A new metric function that balances the tradeoff between the relevance contribution of adding a new selected feature and its cost penalty to the model was proposed. Apart from selecting the dominant features, an accurate hierarchical fuzzy rule-based classification model was developed. A novel regularized loss function was proposed to build up a multi-layer structure during training. Finally, this thesis proposed a time series classification model which aims at predicting the system yield of any molding process.