Transactions on Machine Learning and Data MiningP-ISSN: 1865-6781, E-ISSN 2509-9337ISBN 978-3-942952-89-7
Volume 14 - Number 2 - October 2021 - Pages 55-69
Bearing Lubricant Defect Segmentation Using Synthetic Data
Richard Bellizzi1, Jason Galary2, Alfa Heryudono31,2 Applications Development & Validation Testing Lab, Nye Lubricants, USA
3 University of Massachusetts Dartmouth, USA
Abstract
Lubricant testing often requires a post-test examination of specimens
to obtain the desired critical measurement. Advancing these analysis methods aids in develop-ing higher-performing
products by allowing for improved insight into the lubri-cants’ performance. Building off previous work in Computer
Vision and Ma-chine Learning, this work aims to extend the use of these methods into the lubri-cant testing realm.
Minimizing defects is a desirable outcome since part of the lubricants’ role is to protect the bearing’s surface.
While large-scale defects are easy to interpret, it becomes difficult to differentiate between test results
when comparing bearing examples with less apparent defects. Providing a more con-sistent, granular analysis of
these tests can help lubricant development withstand stringent requirements.
R-Mask CNN methods provide an option to apply instance segmentation tech-niques to classify areas of interest,
allowing for an image with multiple instances of these defects. Since big data is the fuel for a system like this,
there are certain limitations regarding the number of examples for lubricant bearing surface defect data. Leveraging
data amplification techniques allows for a synthetic ‘big’ data set to accommodate the model’s needs. This paper lays
out how these tools work synergistically to provide a model that can operationalize for a company sooner than waiting
to generate a complete set of ideal data.
Keywords:Machine Learning, Computer Vision, Defect Segmentation, Synthetic Data
Download Paper (379 KB)