The objective of this dissertation is to develop a new set of efficient process monitoring and diagnostic tools through their integration with data mining algorithms. Statistical process control (SPC) is one of the most widely used techniques for quality control. Although traditional SPC tools are effective in simple manufacturing processes that generate a small volume of independent data, these tools falter when confronted by the large streams of complex and correlated data found in modern manufacturing systems. As the limitations of SPC methodology become increasingly obvious in the face of ever more complex manufacturing processes, data mining, because of its proven capabilities to analyze and manage large amounts of data, has the potential to resolve the problems that are stretching SPC to its limits. This dissertation consists of three components.

First, we propose a new class of control charts that take advantage of available out-of-control information to improve the detection efficiency. The proposed charts integrate a traditional multivariate control chart technique with a supervised classification algorithm. We call the proposed chart the "Probability of Class (PoC) chart" because the values of the PoC, obtained from classification algorithms, are used as monitoring statistics. The control limits of PoC charts are established and adjusted by the misclassification cost. Second, we propose a collection of new control charts, based on one-class classification algorithms to improve both phase I and phase II analyses in SPC. The proposed one-class classification-based control charts plots a monitoring statistic that represents the degree of being an outlier obtained through the one-class classification algorithm. The control limits of the proposed charts are established based on the empirical level of significance on the quantile estimated by the bootstrap method. Third, we propose a nonparametric false isolation approach in multivariate SPC through monitoring statistics obtained from the one-class classification-based control charts.

The monitoring statistics obtained from one-class classification are decomposed into individual components that reflect the contribution of individual variables to the fault signal. The threshold derived from the bootstrap-quantile estimated method can help indicate the significance of these variables. The novelty of this dissertation is the integration of perspectives from data mining, quality engineering, and statistics that recognizes their shared goals while highlighting their key differences, so as to enable new methodologies for overcoming longstanding research problems and challenges appearing in modern manufacturing/service systems.