The goal of this research is to study the feasibility of applying neural networks to the real-time discrimination between true defects and surface anomalies in float glass ribbon. At the present time the industry uses a laser scanner system capable of detecting defects in newly manufactured glass ribbon. The system does not distinguish between true defects and harmless surface anomalies which causes marketable glass to be wasted. Data collection will be performed in a laboratory setting using a laser scanner system identical to those currently working in the factory. The laboratory setup will be designed to simulate the conditions as closely as possible, so that any result from this study should be applicable to factory production. Because a large number of defect samples are available, the evaluation will focus on neural networks which learn in a supervised mode. At least two such techniques will be utilized: (1) backpropagation network with the generalized delta rule used as a learning algorithm; and (2) a technique which involves the use of the delta rule in a novel way.
