Product Quality Management Using a Real-time Expert System

Abstract

Improving plantwide product quality is a universal concern for every manufacturer. A prototype of a real-time expert system for this purpose has been developed. One of the main functions of the system is quality prediction. The quality of a steel product is predicted by the plantwide quality Expert System (PWQES) by utilizing process variables and operating conditions that change over time. An important issue is how to predict the quality of the product using both real-time data and expert knowledge. Because the knowledge of a steel mill expert includes many uncertainties with regard to the exact causes of specific defects, a method to handle this uncertainty is required in the PWQES. Several methods were examined that deal with uncertainty in prediction of defects in the steel products: production rules with certainty factors, artificial neural networks, and nearest neighbor classifiers. First, a rule set module was built for PWQES that includes steel mill expert knowledge about defects and their causes. Each prediction rule has a certainty factor associated with it. Then, an artificial neural network was constructed and trained using actual process data from the same mill. Finally, a nearest neighbor technique was used to classify products as defective or not. Utilizing five different measurements of prediction error, these different methods then were compared for the mill. As a result, a hybrid system was implemented that utilizes the production rule set and nearest neighbor methods for the quality prediction function.