Real-Time Ferrogram Segmentation of Wear Debris Using Multi-Level Feature Reused Unet

Area of Science:

• Mechanical Engineering
• Materials Science
• Artificial Intelligence

Background:

• Real-time monitoring of machinery is essential for effective maintenance.
• Analyzing wear debris in lubricating oil is a key method for assessing equipment health.
• Existing semantic segmentation models face challenges in accurately identifying minute wear particles.

Purpose of the Study:

• To develop an improved method for segmenting tiny wear debris in oil images.
• To enhance the accuracy of wear particle analysis using the online visual ferrograph (OLVF) technique.
• To mitigate segmentation inaccuracies caused by reflections and bubbles in oil samples.

Main Methods:

• Comparative experiments using various semantic segmentation models (DeepLabV3+, PSPNet, Segformer, Unet).
• Proposal and implementation of a novel Multi-Level Feature Reused Unet (MFR Unet).
• Enhancement of the residual link strategy within the Unet architecture.

Main Results:

• The proposed MFR Unet demonstrated superior performance in segmenting minute wear debris.
• The enhanced model effectively reduced the influence of reflections and bubbles on image segmentation.
• MFR Unet achieved more accurate identification of wear particles compared to other tested models.

Conclusions:

• The MFR Unet offers a significant advancement for the real-time monitoring and fault diagnosis of machinery.
• Accurate segmentation of wear debris is critical for predictive maintenance strategies.
• This approach holds promise for improving the reliability and efficiency of industrial equipment maintenance.