Wer Debris Recognition and Quantification in Ferrography Images by Instance Segmentation

Abstract

Due to the diversity of debris morphologies and their states in the images, automatic wear debris analysis is still a challenge. Instance segmentation is an advanced technique of deep learning used for computer vision tasks. It can detect and delineate each distinct object of interest appearing in an image. An instance segmentation model using a modified mask branch with larger feature map and dilated convolution is proposed and applied to the ferrography images, so that five typical types of wear debris corresponding to different wear conditions can be located, segmented, and recognized, even if they are small, blurred, or overlapped. The mean average accuracy on the test set is about 88% when the intersection over union threshold is 0.5. This method realizes an end-to-end process of wear debris identification that can eliminate the error caused by the intermediate steps. In addition, it obtains accurate quantitative debris information—that is, the type, quantity, and size of debris—which provides a solution for intelligent wear debris analysis and automatic wear detection.

Keywords:

• Wear debris analysis
• ferrography
• instance segmentation
• convolutional neural network

Additional information

Funding

This research was financially supported by the National Key Laboratory of Science and Technology on Helicopter Transmission (Grant No. HTL-A-21G03; Nanjing University of Aeronautics and Astronautics).