Wear Debris Classification and Quantity and Size Calculation Using Convolutional Neural Network

Abstract

The steel production equipment faults are mostly caused by wear faults, and the classification of wear debris in its lubrication system can monitor the wear status of the machine. The traditional methods of wear debris image classification mostly use digital image processing technology by extracting color, shape, texture and other multi-dimensional features of wear debris. It is so difficult to extract suitable multi-dimensional features that the classification accuracy is always kept at a low level. Convolutional Neural Network can directly take the image pixels as input, and extract features automatically, avoiding the poor applicability of manual extraction methods and complicated image pre-processing. An improved lightweight convolutional neural network for wear debris image classification named UstbNet is proposed in this paper. Data augmentation, number and size adjustment of convolution kernels, batch normalization and loss function optimization are used to speed up the model convergence and improve the classification accuracy. The classification accuracy of UstbNet model reaches 96%. After the step of determining the existence of wear debris, we use Faster RCNN to detect the quantity and size of wear debris and further improve it. Grabcut is applied to segment wear debris image based on detected region proposals.