Predictive modeling is a type of supervised learning to find the functional relationship between the input variables and the output variable. Predictive modeling is used in various aspects in manufacturing systems, such as automation of visual inspect...
Predictive modeling is a type of supervised learning to find the functional relationship between the input variables and the output variable. Predictive modeling is used in various aspects in manufacturing systems, such as automation of visual inspection, prediction of faulty products, and result estimation of expensive inspection. To build a high-performance predictive model, it is essential to secure high quality data. However, in manufacturing systems, it is practically impossible to acquire enough data of all kinds that are needed for the predictive modeling. There are three main difficulties in the data acquisition in manufacturing systems. First, labeled data always comes with a cost. In many problems, labeling must be done by experienced engineers, which is costly. Second, due to the inspection cost, not all inspections can be performed on all products. Because of time and monetary constraints in the manufacturing system, it is impossible to obtain all the desired inspection results. Third, changes in the manufacturing environment make data acquisition difficult. A change in the manufacturing environment causes a change in the distribution of generated data, making it impossible to obtain enough consistent data. Then, the model have to be trained with a small amount of data. In this dissertation, we overcome this difficulties in data acquisition through active learning, active feature-value acquisition, and domain adaptation. First, we propose an active learning framework to solve the high labeling cost of the wafer map pattern classification. This makes it possible to achieve higher performance with a lower labeling cost. Moreover, the cost efficiency is further improved by incorporating the cluster-level annotation into active learning. For the inspection cost for fault prediction problem, we propose a active inspection framework. By selecting products to undergo high-cost inspection with the novel uncertainty estimation method, high performance can be obtained with low inspection cost. To solve the recipe transition problem that frequently occurs in faulty wafer prediction in semiconductor manufacturing, a domain adaptation methods are used. Through sequential application of unsupervised domain adaptation and semi-supervised domain adaptation, performance degradation due to recipe transition is minimized. Through experiments on real-world data, it was demonstrated that the proposed methodologies can overcome the data acquisition problems in the manufacturing systems and improve the performance of the predictive models.