Weld Inspection
Detecting weld defects: porosity, cracks, lack of fusion, slag inclusion.
Weld inspection uses computer vision to detect defects in welded joints — porosity, cracks, lack of fusion, undercut, and spatter — from radiographic (X-ray), ultrasonic, or visual images. Deep learning has automated what was traditionally manual radiograph interpretation, with YOLO and U-Net variants achieving specialist-level detection.
History
First CNN-based weld defect classification from radiographic images
GDXray dataset provides standardized weld radiograph images with defect annotations
YOLO adapted for real-time weld defect detection in radiographs
U-Net applied to weld seam segmentation from camera images during welding
Transfer learning from ImageNet dramatically reduces required weld defect training data
Automated ultrasonic weld inspection using ML for TOFD and phased array data
Multi-modal fusion (visual + thermal + ultrasonic) improves defect detection reliability
In-process monitoring — detecting defects during welding from melt pool images
AI-based weld inspection achieves ISO 5817/EN 12517 certification-equivalent classification
How Weld Inspection Works
Image Acquisition
Weld joints are imaged using radiography (X-ray/gamma-ray), ultrasonic testing, or visual cameras — each modality reveals different defect types.
Preprocessing
Radiographs are enhanced for contrast; visual images are cropped to the weld region; ultrasonic signals are converted to images (B-scan, C-scan).
Defect Detection
Object detection (YOLO, Faster R-CNN) localizes defect regions, or segmentation (U-Net) provides pixel-level defect boundaries.
Defect Classification
Detected regions are classified into standard defect types per ISO 6520: porosity, cracks, lack of fusion, undercut, slag inclusion.
Acceptance Grading
Defects are graded by severity according to welding standards (ISO 5817, AWS D1.1) to determine accept/reject/repair decisions.
Current Landscape
Weld inspection AI in 2025 is transitioning from research to industrial deployment. Post-weld radiographic inspection using YOLO/Faster R-CNN is the most mature application, achieving inspector-equivalent accuracy on standard defect types. In-process monitoring (detecting defects during welding) is the growth area, enabling immediate correction rather than post-weld rework. The regulatory landscape is evolving — standards bodies (ISO, AWS) are beginning to address AI-assisted inspection, but full autonomous certification remains pending. Major players include Evident (Olympus), Baker Hughes, and specialized startups.
Key Challenges
Data scarcity — labeled weld defect datasets are small (hundreds, not millions) due to domain expertise required for annotation
Class imbalance — cracks (the most dangerous defect) are rare in production data
Modality diversity — X-ray, ultrasonic, and visual inspection each require different model architectures and preprocessing
Standard compliance — AI systems must demonstrate equivalence to certified human inspectors per industry standards
In-process noise — monitoring during welding introduces spatter, arc light, and smoke that obscure the weld pool
Quick Recommendations
Radiographic weld inspection
YOLOv8 / Faster R-CNN on GDXray
Proven detection performance on standard weld radiograph datasets
Visual weld surface inspection
U-Net + ResNet encoder
Pixel-level segmentation for surface defects visible to cameras
In-process monitoring
Custom CNN on melt pool images
Real-time monitoring during welding to catch defects as they form
Ultrasonic data interpretation
1D CNN or LSTM on A-scan data
Temporal signal processing for ultrasonic testing results
What's Next
The frontier is closed-loop weld quality control — AI that detects defects in real-time during welding and automatically adjusts welding parameters to prevent them. Expect digital weld certificates generated by AI inspection, multi-modal sensor fusion for comprehensive defect characterization, and federated learning across fabrication shops to build larger defect datasets.
Benchmarks & SOTA
No datasets indexed for this task yet.
Contribute on GitHubRelated Tasks
Something wrong or missing?
Help keep Weld Inspection benchmarks accurate. Report outdated results, missing benchmarks, or errors.