A Deep Learning-Based Model for Recognizing Wear Topography of Self-Lubricating Joint Bearings

The study of wear topography in self-lubricating joint bearings is of significant importance for evaluating their service life. In this work, an image dataset was acquired using a white-light interferometer, and the topographical height and color of the images were standardized. Images of worn bearing specimens subjected to 72,000 swing cycles at a frequency of 2 Hz under loads of 100 N, 150 N, 200 N, and 250 N were optimized to construct a processed image dataset. To overcome the limitations of traditional recognition methods in fully capturing both global and local image-based metrics, an improved residual neural network (ResNet) model was proposed. Comparative results with CNN, CapsNet, and conventional ResNet models indicate that, on the processed image dataset, the proposed method improved recognition accuracy by 34% relative to traditional approaches, and by 19% compared to the conventional ResNet. This study provides a novel approach and perspective for investigating the wear topography of self-lubricating joint bearings.