Application of Connectionist Learning Methods to Manl~farturing Process Monitoring

One of the projects in the Machine Learning Department a t GTE Laboratories involves studying quality control of a fluorescent bulb manufacturing line. All manufacturing processes are subject to incompletely understood changes due t o variations in raw materials, environmental factors such as weather, wearing and aging of the machinery, and changes in operators. This can result in iiiarked variations in yield, quality, and rejection rates. These problems can occur even in the most mature and established manufacturing processes. Our goal is Computer Integrated Manufacturing (CIA4), the autonomous computer-based monitoring of plant behavior, determuatioii of causal influences, and, ultimately, adaptive control of the plant process. The first stage is process monitoring with respect to yield and other performance measurements. By using past experience to find correlations between approximately one hundred sensory measurements, we will determine which process variables most affect quality. Two approaclies are being compared. One employs standard statistical procedures to find correlations between sensor me,isureinents and quality. The sensor data from the production line are collected over a period of time and correlatious are made olf-line at infrequent intervals using analyses such as linear regression. The second approach is t o estimate the correlations incrementally, as the data are collected, on-line aiid i n real-time. The estimates are updated incrementally using connectionist learning procedun s.