(The) hybrid modeling of machine learning approaches for intelligent quality system on large-scale of data

Abstract

In manufacturing industry with sequential processes, quality of the product is influenced by many factors that have complex causal relationships between processes and parameters. Although traditional approaches such as statistical process (SPC) have done a central role in manufacturing industry, they have limitations to control the processes considering a lot of variables simultaneously. Machine learning techniques have ability to learn from data and to handle uncertain, imprecise, and complex information in large scale of data. This thesis proposes three hybrid machine learning approaches of inductive learning and neural networks for solving quality problems. First, feature selection using neural networks is used to select a subset of features for higher correct prediction rate of inductive learning. Although a wrapper method which has known it yields good results in feature selection is difficult to apply to the population with large scale of data due to the complexity increased exponentially, this method has reduced its complexity to O(N) using sensitive information of network outputs about features from back-propagation neural networks after training. Second, clustering inductive learning method predicts quality characteristics of output with higher correct prediction rate by a hybrid method of SOM and inductive learning. Third, a pattern detection method is proposed to detect abnormal patterns of processes using self-organizing mapping (SOM) neural networks and test of goodness-of-fit. It gives a solution for multivariate process control on large parameters of data. This thesis presents these approaches as a unifying framework to solve quality problems with large scale of data. The hybrid machine learning approaches have applied to yield management of semiconductor industry.