On line prediction of surface defects in hot bar rolling based on Bayesian hierarchical modeling

Steel bar manufacturing process is a multistage process in which melted steel passes through multiple steps to form into a bar or rod of varying diameters. The most common cause of rejection of such products is the surface defects such as seams, scales, cracks etc. Often automated inspection in the form of image analysis is done to predict the formation of such defects. These inspections suffer from the fact that they do not include the process information and hence are not very reliable. This paper presents a novel methodology of prediction of surface defects in bar manufacturing. In this method, the surface images obtained by inspection methodology are decomposed into individual defects based on process knowledge. These are then decomposed into the physical quantities affecting them which are decomposed into design parameters on rolling mill. To achieve this decomposition, Bayesian hierarchical modeling is used. Data collected from sensors installed on rolling mill are used for model building and design. After the model is built, it is converted into an automated system called Diagnostics on Rolling Mill (DORM). DORM is installed on rolling mill and its performance is evaluated for several days. It is found that the model performs very closely with manual inspection and predicts surface defects with good accuracy. The Bayesian model also gives an insight into process parameters and physical quantities which affect the individual defects.