Automated Machine Learning System for Defect Detection on Cylindrical Metal Surfaces.

Area of Science:

• Materials Science and Engineering
• Computer Science and Artificial Intelligence

Background:

• Metal workpieces are crucial in manufacturing, but surface defects like scratches and dirt compromise product quality and safety.
• Manual inspection is labor-intensive, prone to errors, and inefficient for large-scale production.
• Automated optical inspection (AOI) is preferred, but traditional algorithms struggle with metal surface defect detection.

Purpose of the Study:

• To compare the effectiveness of three Convolutional Neural Network (CNN) models (VGG-16, ResNet-50, MobileNet v1) for metal surface defect detection using transfer learning (TL).
• To develop and evaluate a novel automated machine learning (AutoML) model for enhanced metal surface defect detection.
• To assess the performance of TL, AutoKeras, and the developed AutoML model on a metal defect dataset.

Main Methods:

• Transfer learning (TL) was applied to VGG-16, ResNet-50, and MobileNet v1 for defect detection.
• A novel AutoML model was developed using random search on core layers of TL models, incorporating a retraining criterion for poor performance.
• AutoKeras was employed to identify a suitable model for the metal surface defect dataset.

Main Results:

• Transfer learning yielded accuracies of 91% (VGG-16), 59.00% (ResNet-50), and 50% (MobileNet v1).
• The AutoKeras model achieved the highest accuracy at 99.83%.
• The developed AutoML model reached 95.50% accuracy, effectively recognizing defects with low training costs.

Conclusions:

• The developed AutoML model demonstrates high accuracy and efficiency in detecting metal surface defects, comparable to AutoKeras.
• This research contributes to advancing diagnostic technologies for smart manufacturing through improved automated quality control.
• The proposed AutoML approach offers a viable alternative for defect detection, especially when dealing with limited defect samples.