Cascaded Segmentation U-Net for Quality Evaluation of Scraping Workpiece

Area of Science:

• Manufacturing Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Hand-scraping is crucial for high-precision machine tools, but its quality assessment relies on subjective, time-consuming human judgment.
• Inconsistent quality evaluation impacts machine tool accuracy and service life.
• Objective and automated methods are needed for reliable scraping quality assessment.

Purpose of the Study:

• To develop an edge-cloud computing system for objective evaluation of hand-scraping quality in machine tools.
• To introduce a novel cascaded segmentation U-Net for accurate segmentation of height of points (HOP).
• To automate the calculation of key quality parameters: percentage of points (POP) and peak points per square inch (PPI).

Main Methods:

• An edge-cloud computing system was designed to capture and process scraping workpiece data.
• A novel cascaded segmentation U-Net architecture was developed for high-quality segmentation of height of points (HOP), trained on small datasets.
• A post-processing algorithm was implemented to automatically calculate POP and PPI from segmented HOP data.

Main Results:

• The developed system achieved a low error rate of 3.7% for POP and 0.9 points for PPI.
• The cascaded segmentation U-Net achieved a high Intersection over Union (IoU) score of 90.2% for HOP segmentation.
• The network architecture, based on identity functions, effectively handles oil ditches and residual pigment, enabling end-to-end training.

Conclusions:

• The proposed edge-cloud system and cascaded segmentation U-Net provide an effective and objective method for evaluating hand-scraping quality.
• Automated assessment using POP and PPI significantly enhances accuracy and efficiency compared to traditional subjective methods.
• The novel network architecture demonstrates robust performance, even with limited training data, paving the way for improved machine tool manufacturing.