Multivariate state space mapping models for process and quality improvement

The rapid growth of Distributed Control Systems and the ability to archive massive quantities of data has led to a need to better utilize the information contained in these process and quality databases. The stored data contains important information on how the process is and can be operated to improve efficiencies and product quality. However, these massive multivariate data sets can be difficult to analyze using traditional statistical methods. Recently, a variety of data mining techniques have begun to be applied to these databases in an attempt to determine fundamental multivariate patterns and relationships.; In manufacturing applications, it is important to allow for rather nasty characteristics: multiple change points, arbitrary types of change, arbitrary duration times in states, and closed-loop control. For example, in continuous manufacturing processes, these conditions are the norm, not the exception. Traditional views of processes saw these as streams of individual variables. However, it is possible to look at this same process as a series of process states that evolve over time. Rather than analyze multivariate relationships between variables (such as correlations), this research proposes that it can be more effective to develop a basis of process states and analyze the behavior of these states over time. The multivariate measurements of the process variables represent each process state. The goal is to link the process state characterization with manufacturing performance and quality. In this manner, a partition can be used to relate input and output performance.; There might be a variety of different conditions at which the process runs well, but many other conditions might produce poor performance. A fundamental role of the massive, historical data is to identify good conditions. Then, studies can be made of the properties of good conditions to gain process insight and knowledge, decide how to transition from poor to good conditions for a new type of process control, and combine data with engineering and scientific knowledge for better process operation. These benefits start with the partitions.; Consequently, one needs a flexible tool to generate process states. Detecting and quantifying these hidden patterns or structure in large multivariate data sets can be of great value in understanding process behaviors. The Self-Organizing Map (SOM) developed by Kohonen is one method that can be used to develop this understanding.; This approach can be applied to process data to determine a spanning set of process states, which represent the fundamental multivariate states the process exhibits. The method greatly reduces the dimensionality of the space under study. This research applies a state space approach to modeling both the process and quality variables. The dimensionality is greatly reduced and a mapping of the multivariate process state space to multivariate quality state space is developed. This provides information on which process states lead to the most desirable quality states. The information from this mapping can be used to determine standard operating conditions that are stable and improve quality.