Poisson mixture model for defects prediction in steelmaking

Abstract

The modern steel industry adheres to providing high standard product quality. However, defects are always occurring due to variations during the manufacturing process. Thus, it is crucial to predict the occurrence of defects in real time. However, traditional probability models are inappropriate to model the observed defect data because they exhibit the unique characteristics of nonnegative integers, high-overdispersion, and heterogeneity. To deal with these problems, this work proposes an online defects prediction system based on the Poisson mixture model. Poisson mixture model consists of the component-specific models and mixing probability models. Each component-specific model captures the characteristics of that component while the mixing probability model takes the different sources of heterogeneity of the defect data into account. Compared to the standard Poisson model, Poisson mixture model is more flexible in dealing with extra-dispersion and heterogeneity problems in the defect data. The application results on the real steelmaking process have validated that the Poisson mixture model performs better than the PLS, Poisson, and negative binomial models in prediction accuracy.