Grad-MobileNet: A Gradient-Based Unsupervised Learning Method for Laser Welding Surface Defect Classification.

Area of Science:

• Materials Science and Engineering
• Artificial Intelligence and Machine Learning
• Manufacturing Technology

Background:

• Laser welding is critical in new energy vehicle battery manufacturing.
• Welding defects, such as lack of fusion and porosity, can compromise battery safety and performance.
• Acquiring balanced datasets of welding defects from industrial sites is challenging for deep learning models.

Purpose of the Study:

• To develop an effective deep learning model for detecting laser welding defects in power battery modules.
• To address the challenge of limited and imbalanced defect data in industrial manufacturing.
• To propose an unsupervised learning approach for defect detection.

Main Methods:

• Construction of the RIAM dataset, comprising industrial images of laser welding with four categories: Normality, Lack of fusion, Surface porosity, and Scaled surface.
• Proposal of Grad-MobileNet, a gradient-based unsupervised model utilizing MobileNetV3 architecture.
• Feature extraction based on image gradient distributions for defect classification.

Main Results:

• The proposed Grad-MobileNet model achieved 99% accuracy in detecting welding defects.
• The unsupervised model demonstrated superior performance compared to expectations for supervised learning methods.
• The model effectively utilizes limited normal data for defect identification.

Conclusions:

• The Grad-MobileNet model offers a robust solution for unsupervised detection of laser welding defects in power battery manufacturing.
• This approach mitigates the need for extensive, balanced defect datasets, making it practical for industrial applications.
• The gradient-based feature extraction proves highly effective for identifying subtle welding anomalies.