Application of neural networks to a diagnostic problem in quality control

In order to properly diagnose quality problems that occur in manufacturing the diagnostician, be it human or computer, must be privy to various sources of information about the process and its behavior. This paper describes how neural networks can be used to provide knowledge of how a product characteristic changed, i.e. shift in mean or variability, when so noted by the control chart application. This is potentially useful because there might be some underlying knowledge about the physical phenomena in question that relates the behavior of the observed characteristic to its processing variables. When a change in the process is detected, a feature vector of process-related statistics is sent to the neural network. The neural network then classifies the change as a shift in mean or variance, and passes this information on to the diagnostician. This paper addresses various issues concerned with this problem, namely: process change detection, feature vector selection, neural network training patterns, and convergence error rates. Construction and corresponding simulation experiments show that the neural network approach training procedure yielded a net that correctly classified the problem 94 percent of the time, averaged across classes and magnitudes. Error rates for this configuration ranged from 1.2 to 19.3 percent for mean changes and 3.7 to 9.4 percent for variance changes. Clearly the change structure identification problem can be successfully addressed by a neural network approach, resulting in very low error rates and relatively robust performance across different magnitudes.