Wear Debris Segmentation of Reflection Ferrograms Using Lightweight Residual U-Net

Abstract

Extraction of lube oil wear debris morphological features is an important means for real-time monitoring of equipment wear, and online visual ferrograph (OLVF) is one of the representative technologies. In the current OLVF wear monitoring, the transmission ferrogram (TF) is basically relied on, but the more informative reflection ferrogram (RF) has not yet been applied, because its complex surface color distribution and bubble interference make it difficult to segment the RF. Accordingly, a convolutional neural network (CNN) model called lightweight residual U-net (Res-UNet) is constructed in this article. Simultaneously, with both RFs and TFs, an automatic labeling method is proposed to label the RFs and make a training dataset to implement network training. The experimental results demonstrate that the trained network can achieve accurate segmentation of RFs with excellent anti-interference performance. The proposed method lays the foundation for the feature extraction of reflection OLVF ferrograms and provides an alternative image segmentation method for other image wear debris sensors.