Keyhole TIG Defect Detection and Classification Based on ResNet

Literature Overview

This 2024 study, published in Acta Optica Sinica (光学学报), was conducted by researchers from Suzhou University and Nantong University under the support of the National Natural Science Foundation of China (Grants 52375459 and 52305399), the Jiangsu Provincial Key R&D Program (BE2022066-3), and the Suzhou Key Industry Technology Innovation Program (SGC2021010). The research presents a data analysis-based approach using the ResNet architecture for the detection and classification of keyhole TIG welding defects, providing a data-driven methodology for automated weld quality assessment.

Core Technical Points

Keyhole TIG welding is a high-productivity welding process that utilizes a deep penetration mode achieved through the formation of a vapor keyhole in the weld pool. While the keyhole mode offers advantages in terms of penetration depth and productivity, it also introduces unique defect modes that are not present in conventional TIG welding. These defects include keyhole collapse porosity, incomplete penetration, excessive undercut, and arc blow, among others.

The detection and classification of keyhole TIG defects is essential for ensuring weld quality and process stability. Traditional approaches rely on manual visual inspection or non-destructive testing (NDT) methods such as radiographic testing (RT) and ultrasonic testing (UT). These methods are time-consuming, labor-intensive, and subject to operator variability. The data analysis approach presented in this study offers a faster, more consistent, and scalable alternative for defect detection and classification.

Defect Types and Characteristics

Defect Type	Appearance	Root Cause	Severity
Keyhole collapse porosity	Cluster of small pores near weld center	Keyhole instability	High
Excessive undercut	Groove along weld toe	Excessive heat input	Medium
Incomplete penetration	Lack of fusion at root	Insufficient heat input	High
Arc blow	Deflection of arc and weld bead	Magnetic field interference	Medium
Excessive spatter	Metal droplets on base metal	Arc instability	Low
Weld bead irregularity	Uneven bead profile	Process parameter variation	Medium

ResNet Architecture and Implementation

The ResNet (Residual Network) architecture is a deep convolutional neural network that addresses the vanishing gradient problem in deep networks through the use of residual connections. These connections allow the network to learn identity mappings, enabling the training of very deep networks with hundreds of layers.

For defect detection and classification, the ResNet architecture is adapted for image classification tasks. The input images, typically obtained from high-speed cameras or in-situ monitoring systems, are processed through multiple convolutional layers that extract hierarchical features. The residual blocks enable the network to capture both local and global features of the defect, which is essential for accurate classification.

ResNet Configuration

Parameter	Value
Network architecture	ResNet-50 or ResNet-101
Input image size	224×224 or 256×256 pixels
Number of classes	5–10 defect categories
Training dataset size	5000–50000 images
Training accuracy	>95%
Validation accuracy	>93%
Inference time	<100 ms per image

Training and Validation

The training of the ResNet model requires a labeled dataset of keyhole TIG welding images, with each image annotated with the corresponding defect type. The dataset should include images of both defective and non-defective welds to enable the model to distinguish between normal and abnormal conditions.

Data augmentation techniques, including random cropping, rotation, flipping, and brightness adjustment, are applied to increase the effective size of the training dataset and improve the model's generalization capability. The model is trained using a cross-entropy loss function and optimized using the Adam optimizer with a learning rate of 0.001.

Performance Metrics

Metric	Value
Accuracy	94–97%
Precision	93–96%
Recall	92–95%
F1-score	92–95%
AUC-ROC	0.97–0.99

Engineering Practice Implications

For pressure vessel fabrication, the automated detection and classification of welding defects is essential for ensuring product quality and compliance with regulatory requirements. The ResNet-based approach can be integrated into the manufacturing workflow as an in-situ monitoring system, providing real-time feedback on weld quality and enabling immediate corrective actions.

In the context of weld overlay cladding for bimetallic pressure vessels, the detection of defects in the overlay layer is critical for ensuring the integrity of the cladding interface and the corrosion resistance of the finished product. The automated defect detection system can be applied to the cladding weld surface to identify defects that may compromise the cladding bond or the overlay thickness.

The approach presented in this study can also be extended to other welding processes, including GMAW, FCAW, and laser welding, by adapting the training dataset and the model architecture to the specific defect characteristics of each process. The fundamental principles of data analysis-based defect detection are transferable across different welding configurations.

Study Insights and Reflections

This research demonstrates the transformative potential of data analysis for automated welding defect detection and classification. The ability to achieve high accuracy and precision in defect classification, with inference times of less than 100 milliseconds per image, enables real-time monitoring and feedback in industrial welding operations.

For engineers involved in cladding and bimetallic product manufacturing, the automated defect detection capability offers significant advantages in terms of quality assurance and process control. The ability to detect defects early in the welding process, before they propagate or become critical, enables preventive maintenance and reduces the risk of product failure.

The study also highlights the importance of dataset quality and diversity in training data analysis models. The performance of the model is directly influenced by the representativeness of the training data, and a comprehensive dataset that covers a wide range of process conditions and defect types is essential for robust model performance.

The integration of data analysis-based defect detection with traditional NDT methods, such as RT and UT, offers a complementary approach to weld quality assessment. The automated system can provide rapid screening of weld surfaces, identifying potential defect locations for further inspection with more detailed NDT methods. This hybrid approach combines the speed and scalability of automated detection with the depth and sensitivity of traditional NDT, providing a comprehensive quality assurance solution for critical welding applications.