A DYNAMIC LEARNING MODEL FOR ON-LINE QUALITY CONTROL USING THE TAGUCHI APPROACH

Abstract

An important philosophy of quality control due to Taguchi rests on the belief that, once quality is designed into the product and the process, very little inspection is necessary. Taguchi evaluates quality in terms of costs incurred, including inspection and adjustment costs apart from quality loss. Using these costs, he determines an optimal interval for periodic inspection of products. At each inspection, the first item on the process line is checked for defects. If no defects are found, the process is assumed to be under control. On the other hand, if the first item turns out to be defective, the process is immediately stopped and adjusted. The underlying assumption here is that, once a defect is observed, it is construed that the process has drifted and all subsequent items are defective.

This article challenges the underlying assumption by recognizing that the observation of one defective item does not necessarily imply that the process is out of control. Noisy signals may occur as a result of nonprocess disturbances such as measurement error or operator error. The result of the relaxation of Taguchi's assumption is that the decision of whether to stop the process would have to be made dynamically. The decision model developed in this paper is based on the concept of state-space search and incorporates Bayesian and regression-learning mechanisms for parameter estimation. Subsequently, a framework for an expert controller for automating the system is developed. In this framework, knowledge is identified in the form of three components: assessment knowledge, simulation knowledge, and process knowledge. The problem studied has wide applications in the implementation of the Taguchi approach to quality control in manufacturing environments. Possible extensions and directions for-future research are also presented.