Surface defect detection with predictive models in the galvanizing process

Abstract

Hot-dip galvanizing is a widely used process worldwide to provide metal products with a protective layer that enhances its corrosion resistance. The effectiveness of such layer relies on the uniformity of the coverage, thus, any alteration in the galvanizing layer may be considered as a defect. These defects are catalogued as surface defects where two groups are identified: Bare Spots and Dross-Derived defects. Currently, these defects are detected at the end of the line where no preventive actions can be performed. Consequently, the surface defects’ occurrence is not avoided, increasing in turn the expenses of the company. For that reason, a project oriented to these defects’ prediction is proposed.

This project consists on a set of predictive models, which are tested to be able to predict these defects’ occurrence at an early stage that let the people of the galvanizing line to design and unleash preventive actions that could alleviate the surface defects’ incidents. Four models are studied: Stepwise Logistic Regression, Random Forest Classifier, Gradient Boosting Classifier, and Low FNR Low FPR Random Forest Classifier (LFNR-LFPR RFC) ensemble. LFNR-LFPR RFC is a custom-made multi-objective ensemble designed in this project, which basic learners are two Random Forest Classifiers. To test the models’ performance, the False Negative Rate (FNR) and False Positive Rate (FPR) scores are employed, where the acceptance criteria is to at most have a 15% of FNR and a 25% FPR. From the models tested, LFNR-LFPR RFC was able to outperform the others while achieving FNR and FPR scores under the acceptance criteria for most of the studied cases (two out of three for Bare Spots and one out of two for Dross-Derived defects). Furthermore, the importance of the variables selected for the LFNR-LFPR RFC model was evaluated. As a result, variables from different sources, such as the galvanizing line per se, the chemistry of the coil and from upstream processes, were obtained. In turn, these lists of variables can provide insights on how to design preventive actions that could decrease the surface defects’ occurrence. Finally, the economic impact of the defects and the predictive models is assessed, where, according to the LFNR-LFPR RFC ensemble’s results, savings are possible.