Surface defect detection, classification and root cause diagnosis in steel hot rolling process.

Steel hot rolling process is a complex manufacturing process. The recent development of sensing techniques makes it possible for inspecting the surface quality status online. In this thesis, an intelligent machine vision system was developed to detect, classify the surface defects and analyze root causes for the hot rolling process. There are three indispensable components of the system. Two different categories of efficient feature extraction methods based on image signals are developed for the defect detection. Local image features are calculated for non-repeating defects. Global features are extracted through autocorrelation for repeating defect patterns and the related parameter is optimized to make the features more robust. The online application of the algorithms shows they are fast enough to be incorporated in a real-time rolling process and robust to the variance of lighting, surface texture, and foreign inclusions. The extracted features are used as the input to the intelligent classification system for decision-making. A Support Vector Machine (SVM) technique based machine learning system was investigated. An efficient parameter selection method for training SVM classifiers is introduced based on one-class SVM. The incremental learning system is proposed based on one-class SVM and two-class SVM. The system can automatically detect the new defect categories by describing the current known defect categories as current knowledge domain. The system updates itself in an incremental fashion to accommodate new data without compromising classification performance on old data. Incremental class learning and incremental data learning are integrated into one learning system. The self-learning intelligent system that meets the requirement for the online application can expend its knowledge automatically. Two different root cause diagnosis approaches for rolling defects are explored according to the availability of the background knowledge about the mechanism of defect generation. Knowledge driven diagnosis approach is used when defect generation mechanism is known. Data driven diagnosis approach, Bayesian networks, may be applied to the cases that defect generation mechanism is unknown. With efficient defect root cause diagnosis, product quality can be improved by correcting the causal factors.D.Eng.Applied SciencesEngineering, Materials scienceIndustrial engineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studies College of Engineering Graduate Professional Programshttp://deepblue.lib.umich.edu/bitstream/2027.42/124997/2/3167146.pd