Weld Inspection: Quality Control AI on GPU

The Challenge: 2,500 Welds and Three Certified Inspectors

A structural steel fabrication company in Sheffield produces beams, columns, and connections for commercial construction projects across the UK. Every production day generates approximately 2,500 welds requiring inspection against BS EN ISO 5817 quality levels. The three certified welding inspectors (CWIs) currently review radiographic (RT) and visual test images, classifying each weld as acceptable or identifying specific defect types: porosity, lack of fusion, undercut, slag inclusions, or cracking. Each inspector can evaluate roughly 200 radiographs per shift, creating a persistent backlog that delays project sign-off by two to three days. With one inspector approaching retirement and recruitment for qualified replacements proving difficult, the bottleneck is worsening.

The company needs AI-assisted inspection that pre-screens radiographs and flags likely defects for human review, allowing inspectors to focus their expertise on borderline cases rather than reviewing the 78% of welds that pass without issues.

AI Solution: CNN-Based Weld Radiograph Classification

A convolutional neural network trained on labelled radiographic images classifies weld quality and detects specific defect types. Models built on architectures like EfficientNet, ResNet-50, or custom U-Net variants for segmentation can identify porosity clusters as small as 0.5mm on standard radiographic images. The vision model processes each radiograph on a dedicated GPU server, outputting a classification (pass/review/reject) with defect type labels and bounding boxes overlaid on the image.

The AI does not replace the CWI — regulatory compliance still requires human sign-off — but it triages the queue so inspectors only examine the 22% of welds that need expert assessment. The remaining 78% receive AI-confirmed pass status, with the CWI batch-approving them in minutes rather than hours.

GPU Requirements

Weld radiograph analysis typically uses high-resolution greyscale images (2048×2048 or larger). The classification and segmentation models must run at sufficient resolution to detect sub-millimetre defects. Running both a classifier and a segmentation model simultaneously requires moderate VRAM.

At 2,500 welds daily, any GPU in the range processes the entire day’s output in under five minutes. The private AI hosting option keeps proprietary client project data and quality records within GDPR-compliant UK infrastructure.

Recommended Stack

• EfficientNet-V2 or ResNet-50 for radiograph classification (pass/review/reject).
• U-Net or Mask R-CNN for defect segmentation, producing pixel-level defect maps.
• NVIDIA TensorRT for optimised inference at maximum throughput.
• Label Studio for CWI-assisted annotation of training data, with defect type and severity labels.
• Optional: document AI or PaddleOCR for reading weld identifiers and job numbers from radiograph film labels.

For generating inspection reports, add an open-source LLM via vLLM to convert defect detection results into structured text compliant with inspection documentation standards.

Cost Analysis

The three CWIs cost approximately £180,000 annually in salaries. AI pre-screening reduces their effective workload by 70%, meaning the same team can handle 70% more volume — critical as the company wins larger contracts. Alternatively, when the retiring inspector leaves, the remaining two can maintain current throughput without a replacement, saving £60,000 annually in recruitment costs.

Faster inspection turnaround — from three days to same-day — accelerates project delivery. On a typical £2 million fabrication contract, two days saved in inspection translates to earlier billing and reduced holding costs, worth approximately £8,000 per project. With 15 active projects per quarter, the throughput improvement alone justifies the investment.

Getting Started

Compile a dataset of 5,000 labelled radiographic images with defect classifications from your CWIs. Split into training and validation sets, training the classifier first and adding segmentation once classification accuracy exceeds 95%. Deploy alongside the existing inspection workflow, with AI results displayed as an overlay for CWI review. Track agreement rates between AI and human inspectors over 60 days before expanding to full pre-screening mode.

GigaGPU provides UK-based dedicated GPU servers for industrial vision applications. Add an AI chatbot for quality documentation queries, or integrate with your existing MES for automated inspection data logging.