Wear mechanisms and severity level classification in iron ore transfer chute linings by propagating regional labels coded as embedding deep learning vectors

Abstract

Wear mechanism severity analysis constitutes one of the principal strategies for selecting more wear-resistant materials. A machine learning model was developed to characterize and recognize wear mechanisms and severity in worn surfaces of industrial equipment. Particularly, deep features were computed from a convolutional net (VGG16 and Resnet50), while wear mechanisms and severity classification were computed from a Random Forest (RF) classifier. The histogram of oriented gradient (HoG) was also implemented to compute an image descriptor and thereafter a classification was carried out with Random Forest classifier. From a set of images of worn surfaces of transfer chute linings, two different datasets were employed to search independently among five wear mechanisms and four levels of wear severity. The proposed approach was successful in recognizing and mapping wear mechanisms and determining their severity levels, as confirmed by the high values of metrics obtained using the deep features. A comparison was carried out between the mass loss performance of the materials tested in the field and their wear severity levels recognized by the machine learning model. A satisfactory correlation was identified between the materials with the lowest and highest wear, as determined by their measured mass losses from a field test, and the corresponding levels of wear severity on their worn surfaces obtained by the machine learning model. This result demonstrates the potential of using machine learning for the worn surface analysis to support industrial equipment wear analysis.