| A major problem with quality control charts analysis is the difficulty of judging whether a process has drifted from normal operation. The objectives of this doctoral work are, first, to cast this problem as a pattern recognition problem, and, second, to automate the decision-making process by implementing and testing various forms of statistical and neural pattern recognizers.;The process output presented on the control chart can be viewed as a discrete-time signal composed of two components: a random Gaussian noise signal plus an unnatural 'pattern' signal. The random noise component represents the natural and intrinsic variations of the process that are unavoidable, while the presence of an unnatural pattern indicates an external disturbance agent working in the process, that must be eliminated. Since several types of signals may appear on a control chart (i.e., various unnatural patterns), the problem becomes that of detecting and identifying unnatural patterns, that is, to recognize unnatural patterns.;For the statistical pattern recognition approach, a set of probability distribution functions (pdf's) are constructed for some unnatural patterns, thus providing the capability to derive optimal statistical decision rules. Other patterns, however, cannot be characterized by closed-form pdf's, and alternative approaches have to be pursued. Cross correlation analysis provides the basis for recognizing the latter set of unnatural patterns. Parameters of the pattern distributions are assumed to be unknown. These parameters, namely, the mean vector and the covariance matrix, are estimated from the training data using the Bayesian Estimation Technique. In addition, various information processing procedures, such as standardization, normalization, or coding (for the neural approach) are implemented to extract the maximum possible information from the given process data. A set of performance indices characterizing the operation of control charts are defined and evaluated for different operating conditions. The results obtained have shown the feasibility of the statistical approach; in fact the performance of the designed system outperformed existing systems in terms of Type I and Type II probabilities of error.;The self-organizing neural pattern recognition methodology is motivated by the need to implement a system that is capable of identifying unnatural patterns even when these are not known a priori. This pattern recognition methodology is primarily utilized to identify a change in the process structure. In addition, the potential of applying this technique to the classification of detected patterns is investigated. As a structure detector, the neural-based system is designed to operate in continuous learning mode. This strategy is based on the assumption that the process always starts in a state of statistical control. During this in-control period, called the training period, the output data of the process is streamlined to train the self-organizing network, that in turn, forms 'natural' clusters describing the in-control process. To operate the network in a testing mode, a set of labelled patterns taken during training are used to calibrate the network. This calibration or interpretation process is aimed at identifying which clusters (i.e., network nodes) correspond to the natural process output, which in turn, facilitates the detection of unnatural output. Performance indices, such as the false alarm rate, the rate of detection, and speed of convergence are evaluated for different network parameters. Performance analysis of the neural-based system reveals the viability of the approach to detect structural changes in the process data. |
