Industrial Inspection

Weld Inspection

Detecting weld defects: porosity, cracks, lack of fusion, slag inclusion.

0 datasets0 resultsView full task mapping →

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

2017

First CNN-based weld defect classification from radiographic images

2018

GDXray dataset provides standardized weld radiograph images with defect annotations

2019

YOLO adapted for real-time weld defect detection in radiographs

2020

U-Net applied to weld seam segmentation from camera images during welding

2021

Transfer learning from ImageNet dramatically reduces required weld defect training data

2022

Automated ultrasonic weld inspection using ML for TOFD and phased array data

2023

Multi-modal fusion (visual + thermal + ultrasonic) improves defect detection reliability

2024

In-process monitoring — detecting defects during welding from melt pool images

2025

AI-based weld inspection achieves ISO 5817/EN 12517 certification-equivalent classification

How Weld Inspection Works

1Image AcquisitionWeld joints are imaged usin…2PreprocessingRadiographs are enhanced fo…3Defect DetectionObject detection (YOLO4Defect ClassificationDetected regions are classi…5Acceptance GradingDefects are graded by sever…Weld Inspection Pipeline
1

Image Acquisition

Weld joints are imaged using radiography (X-ray/gamma-ray), ultrasonic testing, or visual cameras — each modality reveals different defect types.

2

Preprocessing

Radiographs are enhanced for contrast; visual images are cropped to the weld region; ultrasonic signals are converted to images (B-scan, C-scan).

3

Defect Detection

Object detection (YOLO, Faster R-CNN) localizes defect regions, or segmentation (U-Net) provides pixel-level defect boundaries.

4

Defect Classification

Detected regions are classified into standard defect types per ISO 6520: porosity, cracks, lack of fusion, undercut, slag inclusion.

5

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 GitHub

Related Tasks

Something wrong or missing?

Help keep Weld Inspection benchmarks accurate. Report outdated results, missing benchmarks, or errors.

0/2000
Weld Inspection Benchmarks - Industrial Inspection - CodeSOTA | CodeSOTA