Real-Time High-Performance Laser Welding Defect Detection by Combining ACGAN-Based Data Enhancement and Multi-Model...

Area of Science:

• Materials Science and Engineering
• Manufacturing Technology
• Artificial Intelligence in Manufacturing

Background:

• Current laser welding monitoring primarily detects post-process defects, lacking real-time capabilities crucial for mass production.
• Data-driven defect detection methods are hindered by the difficulty in building comprehensive laser welding datasets and insufficient experimental data.
• Real-time defect identification is essential for the mass production of electronic equipment utilizing laser welding, such as metal plates.

Purpose of the Study:

• To propose an intelligent welding defect diagnosis method using auxiliary classifier generative adversarial networks (ACGAN).
• To address the challenge of limited data in laser welding defect detection through data augmentation.
• To enhance the accuracy and reliability of real-time defect classification and recognition in laser welding processes.

Main Methods:

• Construction of a ten-class dataset comprising 6467 samples from optical and thermal sensory parameters during laser welding.
• Development of a structured ACGAN network model to generate synthetic data mimicking true defect feature distributions.
• Implementation of a data filtering and purification scheme using ensemble learning and Support Vector Machine (SVM) to refine generated data.

Main Results:

• The ACGAN model achieved 85.13% classification accuracy, while a Convolutional Neural Network (CNN) achieved 96.83%.
• Fusion models demonstrated superior performance: ACGAN-CNN reached 97.86% and ACGAN-SVM-CNN achieved 98.37% accuracy.
• The study validates ACGAN's utility for both classification and significant real-time defect recognition through data enhancement and model fusion.

Conclusions:

• The proposed ACGAN-based method effectively generates realistic data, overcoming limitations of small datasets in laser welding defect detection.
• Data purification techniques further enhance the quality of augmented data, leading to more distinct defect categories.
• ACGAN, especially when fused with CNN or SVM-CNN models, offers a powerful solution for accurate and real-time defect diagnosis in industrial manufacturing.