Abstract
Finding a fast, automated, and more accurate method of inspection and qualification of corrosion in bearing races is one challenge in the lubricant industry. The goal is to inspect a section or multiple sections of a bearing and determine whether corrosion is present or not. The lubricant industry uses the EMCOR method (ASTM D-6138) and the Standard Corrosion methods (ASTM D-1743 and D-5969) to test how lubricants interact within bearings under corrosive conditions. Several factors that create difficulties when integrating such methods include limited available data sets, costly sample generation, uncertainties in visual observations, and bias in finite ratings. Moreover, for smaller companies, limited resources are considered additional development constraints.
Recent advancements in machine learning algorithms allow smaller companies to incorporate computational methodologies as supporting roles, such as developing solutions for inspection and quantifying corrosion on bearings. This paper designs a solution model using the limited amount of 'real' data from in-house testing coupled with Convolutional Neural Networks (CNN) and Transfer Learning (TL). We show that conflicting results caused by human error factors become mitigated through a more reliable and computational-based method. We also demonstrate a repeatable and accurate method for visualizing and classifying corrosion on bearings utilizing CNN and TL to bridge the gap in technology in the lubricant industry. The method can aid smaller companies in a digital transformation transition and provide more insight into products and improve development capabilities.