Short answer
Weld inspection data comes in two families: surface images of weld beads, and volumetric or radiographic images such as X-ray films and digital radiographs, usually with written reports. Public data is scarce; GDXray provides 19,407 X-ray images in five groups, including welds, for research and education. Private archives sit with fabricators and inspection firms, and need standards context, relabeling and rights review.
Why weld data is a separate sourcing problem
A weld is a joint, not a surface. Its quality depends on what is inside it as well as what shows on the outside, so inspection draws on several methods: visual examination of the bead, radiography, ultrasonic testing and others. Each method yields a different data type, and each defect class (porosity, lack of fusion, cracks, undercut, slag inclusions, incomplete penetration) is visible in some methods and invisible in others.
Acceptance depends on whichever weld code and customer specification apply, and these differ by industry; check the applicable standard for each archive rather than assuming one (see acceptance standards as label definitions). A weld that passes under one set of criteria may be rejected under another, so labels are only meaningful with the criteria attached. That makes weld data distinct from generic defect image datasets and from casting defects, even though X-ray reading skills overlap; the general modality is covered in X-ray and CT inspection images. See the inspection cluster and the humanoid data section for scope.
What public weld data exists
The main public resource is GDXray, a database of X-ray images for nondestructive testing. It holds 19,407 X-ray images in five groups: castings, welds, baggage, natural objects and settings, with annotations as bounding boxes or labels; it is free for research and education, and the GDXray+ site reports more than 21,100 images (Mery et al., 2015).
Two cautions follow. The weld group is one part of that total, so do not read the headline count as weld images. And "free for research and education" does not equal permission for commercial model training; read the terms and consult counsel. This is not legal advice.
For other weld modalities, such as surface bead images or phased-array ultrasonic scans, published open data is thin, which pushes buyers toward private archives.
What weld data can and cannot support
It can support defect detection and classification on radiographs, segmentation of indications, bead-geometry assessment from surface images, and report generation or triage models trained on images plus findings. It can support evaluation sets that check a system against qualified inspectors' reads.
It cannot show a robot how to weld or how to position a probe. Trajectories, torch angles, travel speed, standoff and interpass handling are process and control data, which archives of finished-weld images do not contain. Weld images also reveal nothing about the welder's skill unless procedure records are linked. For robots that perform inspection in the field, probe handling would need separate active inspection data.
Types of weld data a buyer might request
| Data type | What it contains | Typical holder | Main caveat |
|---|---|---|---|
| Surface (visual) images | Bead appearance, spatter, undercut, misalignment | Fabricators, shipyards, pipeline contractors | Lighting and angle vary widely |
| Radiographic images | Film scans or digital radiographs of joints | Inspection firms, fabricators | Sensitive to technique and film quality |
| Ultrasonic scans | A-, B- or C-scan data from probes | Inspection firms | Needs instrument and setting context |
| Inspection reports | Findings, locations, acceptance decision | Quality teams, inspection firms | Free text and codes need parsing |
| Weld procedure and joint records | Process, material, joint design | Engineering departments | Proprietary parameters |
| Repair records | Excavation and re-weld history | Quality teams | Links original finding to outcome |
Illustrative package for a weld inspection request
Illustrative; values are synthetic.
| Field | Example |
|---|---|
| Joint context | Butt weld, carbon steel, 12 mm thickness, joint_id J-0412 |
| Method | Digital radiography, 2 views per joint |
| Image | 16-bit, scale marker and penetrameter visible |
| Finding | Indication type, location along the weld, size |
| Acceptance basis | Named code and class, version |
| Decision | Accept, repair, reject |
| Repair outcome | Re-radiograph result |
| Reader | Qualification level of the person who read it |
| Split key | Project, weld procedure, welder or crew |
Reading the data: three practical points
Indications are not defects. Radiographs show indications that a qualified reader interprets against criteria. A dataset of boxes drawn around dark spots mixes real discontinuities, film artifacts and geometry effects such as root reinforcement. Ask how artifacts were separated from findings.
Small crops lose context. A cropped indication without the weld axis, thickness and marker scale cannot be sized or classified reliably. Keep full-width images with annotations rather than only patches.
Inspection is sampled. Many fabricators radiograph a fraction of joints, so the archive over-represents joints selected for inspection rather than the whole population. State the sampling rule before using prevalence statistics.
Quality and acceptance criteria
Acceptance basis recorded. Without the code and class, a "reject" label has no fixed meaning. Require it per image or per project.
Reader qualification and agreement. Radiograph interpretation is skilled work. Ask who read the images and whether a second reader reviewed a sample; see inter-inspector agreement.
Technique metadata. Source type, energy, film or detector, geometry and exposure all change what indications look like. Without them, cross-supplier pooling is risky; see imaging conditions and metadata.
Outcome linkage. Link each finding to repair and re-inspection. A flagged indication that was repaired and passed is better evidence than a flag alone.
Balance and rarity. Most welds pass. Cracks and lack-of-fusion cases are rare and matter most; see rare defects and class imbalance.
Split design. Split by project, procedure or crew. Images of adjacent welds from one pipeline are strongly correlated.
Supplier types
Fabricators and shipyards keep weld images and records for their own projects, often with customer restrictions. Pipeline and pressure-vessel contractors generate large radiograph volumes with formal reports. Third-party inspection and NDT firms hold reports and images on behalf of clients; rights often stay with the client, as discussed in third-party inspection firms as suppliers. Welding robot integrators sometimes hold in-process camera data, which is a different asset from post-weld inspection images.
Preparation the data holder needs to do
Convert films and proprietary detector formats to documented image formats without lossy compression. Attach reports and parse them with care; see inspection report text. Remove client, project and personnel identifiers. Radiographs can carry burned-in or marker text with job numbers, which must be handled as part of the image, not only the metadata. Confirm who owns the images; a fabricator paid by a client may not be free to license them. Ask counsel about the associated engineering records too; this is not legal advice.
Limitations and extra preparation
Weld data is hard to reuse outside the context that produced it. Labels depend on the inspection method, the reader's qualification and the code and class applied at the time; check the applicable standard for each archive rather than assuming one interpretation across sources. Radiographs digitized from film differ from native digital radiographs in contrast and noise, and surface images of beads say nothing about internal discontinuities. Many archives hold only rejected or repaired welds, which skews the class balance and hides what acceptable variation looks like.
None of this data describes how to weld, grind or handle a part; it supports inspection, triage and evaluation, not robot control. Plan for extra preparation: digitizing and calibrating film, separating indications from confirmed defects, attaching the acceptance basis to every label, removing client, project and location identifiers, and arranging review by qualified inspectors where labels are missing or inconsistent. Splits should be made by component or project, because multiple exposures of the same weld are highly correlated.
Questions to ask suppliers
- Which inspection methods are represented, and how many joints per method?
- What code and acceptance class governed each decision?
- Who read the images, at what qualification level, and was there a second read?
- What technique and detector were used, and are settings recorded?
- How many findings were repaired, and is the re-inspection result linked?
- What share of joints are rejects, and what defect types dominate?
- Do images contain burned-in job numbers or customer names?
- Who holds rights to the images and reports, and does the contract with the client allow licensing?
Discuss your requirements
If you need weld inspection images or records from US businesses, describe the method, material, defect types and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match.
Discuss your humanoid training-data requirements with SourceX
Sources
- GDXray: The Database of X-ray Images for Nondestructive Testing — GDXray: 19,407 X-ray images in five groups including welds and castings, annotated with boxes or labels, free for research and education.