Related Experiment Video
Updated: May 27, 2025
09:17
Surrogate Model Development for Digital Experiments in Welding
Published on: March 28, 2025
653
Lightweight DCGAN and MobileNet based model for detecting X-ray welding defects under unbalanced samples
Lei Zhang1,2, Haihong Pan1, Bingqi Jia1
1The School of Mechanical Engineering, Guangxi University, Nanning, 530004, China.
Scientific Reports
|February 20, 2025
Summary
This study introduces an improved DCGAN and DG-MobileNet model to generate diverse welding defect samples and enhance defect identification accuracy. The new model achieves 98.78% recognition accuracy, improving industrial inspection.
Area of Science:
- Manufacturing Engineering
- Artificial Intelligence
- Materials Science
Background:
- X-ray nondestructive testing is vital for weld inspection in manufacturing.
- Diverse welding defects and imbalanced data hinder accurate defect classification and cause model overfitting.
- Existing models struggle with low accuracy and poor convergence in identifying welding defects.
Purpose of the Study:
- To propose an improved DCGAN model for generating synthetic welding defect samples to address data imbalance.
- To introduce a lightweight DG-MobileNet model for accurate and efficient welding defect identification.
- To enhance the feature extraction capabilities and generalization ability of defect classification models.
Main Methods:
- An improved DCGAN model integrated with deep convolutional neural networks was developed to generate augmented welding defect datasets.
- A lightweight DG-MobileNet model incorporating dilated convolution modules and Squeeze-and-Excitation self-attention mechanisms was designed for defect identification.
- Model optimization included replacing the fully connected layer with global average pooling and integrating DropBlock with Batch Normalization.
Main Results:
- The proposed model achieved a high recognition accuracy of 98.78% for welding defect identification.
- The integrated approach effectively increased the number of training samples and mitigated classifier overfitting.
- The DG-MobileNet model demonstrated superior performance in terms of efficiency and lightweight design.
Conclusions:
- The developed models offer a robust solution for enhancing welding defect classification accuracy and addressing data imbalance issues.
- The proposed method shows significant potential for real-world industrial applications requiring efficient and accurate nondestructive testing.
- The study highlights the effectiveness of combining generative adversarial networks with advanced convolutional neural network architectures for industrial defect detection.
