Defect detection with predictive models in the galvanizing process

Abstract

In the steel industry, having better control over the final mechanical properties of the steel coils is something highly desired by companies, as this would allow them to reduce the number of defective products they manufacture and reduce the costs associated with them. In a galvanizing line, modeling the yield strength and elongation properties of steel coils can be done before subjecting the coils to the galvanizing process, therefore preventing the waste of zinc, and improving the overall quality control of the line. In this thesis, an ensemble of two quantile random forest regressors was employed to predict the mechanical properties of galvanized steel coils using real-life data from a steel manufacturing company in order to identify defective and non-defective products. The ensemble was designed with goal-specific components in order to optimize the false negative rate and false positive rate of the model. Out of the six clusters of data built from the dataset, four were properly modeled with this approach, while one was best modeled with an individual quantile random forest regressor. Results revealed that a combination of chemistry, segmentation, previous processes, and galvanizing process parameters are required to effectively predict the yield strength and elongation properties. Additional testing of this ensemble model in different industrial contexts and with different performance metrics is recommended to further validate its efficacy.