AI-Driven Automated Defect Detection in Weld Radiographs

Use Case:

Deployment of an AI-powered tool to detect localize and classify weld defects in radiographic images enabling batch-level automated inspection and reducing manual inspection errors and costs.

Outcome:

• Technology reached TRL7 with full MVP demonstrator
• AI model precision of up to 95% on critical defect classes (e.g., voids)
• Average classification accuracy >85%, outperforming traditional methods by 15%
• Automated pipeline for pre-processing, inference, and defect visualization

Ecosystem Support:

StairwAI project support for AI voucher guidance and expert collaboration; partnership with Preste (AI experts) for dataset annotation, training, and model optimization.

AI Relevance:

The ZeroDefectWeld project exemplifies industrial AI adoption through: I) explainable and modular AI pipelines; II) significant defect detection accuracy using open datasets and expert annotation; III) robust deployment flexibility (local/cloud); IV) reduced inspection overhead and scalable integration for SMEs.

Summary:

StreamOwl, a Greek SME focused on data analytics and applied AI, developed ZeroDefectWeld to address the need for scalable, precise, and automated inspection of welded joints in industrial environments. The system uses Deep Learning to detect and classify defects from X-Ray radiographic images, a domain where traditional methods struggle with low precision and high labor intensity.

In collaboration with Preste, a team of AI experts in computer vision, StreamOwl built an end-to-end pipeline encompassing image pre-processing, defect segmentation, and classification. The team trained UNet and EfficientNet-b0 architectures using the publicly available GDX-Ray dataset, after performing advanced annotation and data augmentation to enable accurate multi-class predictions. The system annotates each image with bounding boxes and defect types, supporting both human validation and fully unattended batch processing.

A key highlight of the project is its dual interface: a command-line tool for automated deployment in production scripts, and a user-friendly web interface for manual review. The solution demonstrated exceptional performance in detecting critical defects (e.g., voids, cracks) with over 95% precision and achieved generalization across diverse image shapes and sizes.

With StairwAI’s support, StreamOwl not only developed the technology but also finalized a business model targeting SMEs in welding, construction, and manufacturing industries. The model emphasizes ease of deployment, reduced inspection cost, and AI-enhanced safety assurance. The solution is now poised for commercialization, laying the groundwork for wider adoption of intelligent quality assurance in heavy industry.

Deployment of an AI-powered tool to detect localize and classify weld defects in radiographic images enabling batch-level automated inspection and reducing manual inspection errors and costs.