Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
Characterization of wear in images of samples from the HFRR test
Submission Author:
Maxwell Jácome , RN
Co-Authors:
Maxwell Jácome, José Josemar de Oliveira Júnior
Presenter: José Josemar de Oliveira Júnior
doi://10.26678/ABCM.COBEM2021.COB2021-2053
Abstract
Tribological systems evaluate the wear mechanisms that occur in contact between metals and the lubricity of fluids. The HFRR (High Frequency Reciprocating Rig) test measures the level of wear and fuel lubricity through the wear of a ball rubbed on a flat surface. This test is evaluated through the parameter Wear Scar Diameter (WSD), being a measure obtained from a 2D image. However, this measure is limited to analyzing the geometry of the image. The HFRR images show structural similarities with images of melanomas, an area of medicine in which several studies on classification between benign and malignant are found. The literature presents nine descriptors extracted from melanoma images for classification through an ANN (Artificial Neural Network): diameter, symmetry (x-axis and y-axis), mean of colors (channels R, G and B), color variance (channels R, G and B). From a group of 56 images, image processing techniques were applied using the Matlab image toolbox and the OpenCV library. Applying the Hough Transform to detect circles, the circumference that best approximated the wear scar was obtained. From it, the diameter and geometric center were obtained. For the symmetry criterion, the images were segmented and divided into four quadrants from the geometric center. The difference between the total number of filled pixels generated the descriptors of symmetry in relation to the vertical and horizontal axes. As for the means and color variance, histograms were obtained in channels R (red), G (green) and B (blue). Then, mean and variance of each data set were obtained. An ANN was built to verify the efficiency of the descriptors in characterizing the group of samples. A Feed Forward Back Propagation network was built with an input layer (descriptor data), two intermediate layers (one with 60 neurons and the other with 1 neuron) and an output layer (with 2 qualitative values, -1 indicating a group and 1 indicating another). The images were divided into a training group (70%) and a validation group (30%). After training, validation data were applied to the already built network, observing the output result. The network showed 75% efficiency when classifying images between two groups of samples with different lubricants. The results demonstrate that the images of the HFRR test provide data capable of recognizing and classifying patterns, be it the type of lubricant used or the most active wear mechanism.
Keywords
Tribology, lubricity, Artificial neural networks
