| In today's computer integrated manufacturing (CIM) environment, one of the difficulties encountered in implementing statistical process control (SPC) is how to automatically identify special disturbances to the process. It has been proven that statistical quality control charts are useful tools for monitoring manufacturing processes. However, two major deficiencies of conventional Shewhart control charts are: (1) the decision is completely based on the last observation plotted against the control limits and; (2) no information about any special disturbances to the process is indicated explicitly. Although Western Electric rules and other control chart schemes, such as CUSUM, are frequently proposed to address some of the problems, information about special disturbances is usually overlooked by these schemes. In this research, a general pattern-recognition methodology for control chart applications was developed to resolve the deficiencies mentioned above. Two types of pattern recognizers, namely, the Back-Propagation pattern recognizer (BPPR) and the Boltzmann machine pattern recognizer (BMPR), are proposed and tested. These pattern recognizers were trained and tested with the consideration of Type I errors as well as Type II errors. Training patterns were selected by using a training pattern refinement procedure. A pattern classification algorithm, which is suitable for real-time statistical process control, was also developed. Extensive experiments and simulations were conducted to validate the usefulness of the proposed methodology. The first step of the study was a factorial design in which a favorable network was obtained. Then, in the secondary study, a more extensive performance evaluation was conducted. Four newly proposed performance measures, namely, rate of target (ROT), rate of non-target (RONT), average target pattern run length (ATPRL), and average run length index (ARLIDX) were used to evaluate these trained recognizers. The characteristics of learning behavior and the performance for both types of recognizers are analyzed and compared. Numerical results indicate that the performance of these pattern recognizers on various unnatural process patterns is quite promising. Finally, the relationships of the neural network approach to multiple regression analysis and to fuzzy set theory are discussed. |
