Automated NDT evaluation (industrial X-ray, CT, ultrasonic) for castings & welds

Non-destructive testing (NDT) reveals internal flaws that no camera can see: gas porosity and shrinkage in aluminium castings, lack of fusion in welds, delaminations in composites. Evaluating the resulting radiographs, CT volumes and ultrasonic scans is slow and subjective. Deep-learning models pre-screen the data, segment defects in 3D and apply acceptance rules, so certified inspectors only review borderline and flagged cases.

Problem & value

• Today: film or digital radiographs and CT data are read by certified inspectors (ISO 9712 levels 2/3) or by classical automatic defect recognition (ADR) with hand-tuned thresholds. Thresholds break on noise, beam-hardening artefacts, varying wall thickness and complex geometry, so false calls are high and skilled reviewers are scarce.
• AI value: consistent first-pass reading, 3D defect segmentation in CT volumes, fewer false rejects (scrap and re-work), faster throughput for 100 % inline inspection, and a documented, reproducible decision trail. Vendors report large reductions in manual review effort; independent, comparable figures are rare, so measure on your own parts [1][2][3].
• Why distinct from VIS-001: the data are volumetric or signal data, not surface images; decisions follow acceptance standards and certified-personnel workflows.

Variants

• Automotive and e-mobility castings: inline CT or DR of aluminium die castings (engine blocks, battery housings, structural parts), 100 % or sampled.
• Weld inspection: radiography and phased-array ultrasonic testing (PAUT/TFM) of pipelines, pressure vessels, rail and shipbuilding welds.
• Aerospace and energy: CT of turbine blades, additive parts and composites (see VIS-057 for in-process monitoring of additive manufacturing).
• Batteries and electronics: CT/X-ray of cells, solder joints and BGA packages (see VIS-056).
• Service labs: NDT service providers that offer AI-assisted reading as a quality-assured pre-screen.

How it works

1. acquisition (X-ray / CT / PAUT)
2. reconstruction & calibration (FDK, artefact correction)
3. pre-processing (normalisation, ROI from CAD/registration)
4. detection / segmentation model (2D or 3D, anomaly branch for unseen defects)
5. measurement (size, volume, position, wall distance, porosity %)
6. rule engine (acceptance class, e.g. ISO 10675 weld / casting specs, customer drawing)
7. verdict: auto-accept | reject | human review (certified inspector)
8. report, archive of raw data + model version, feedback loop

• Models: U-Net / nnU-Net-style 3D segmentation for porosity and cracks in CT, YOLO- or DETR-type detectors for 2D radiographs, and autoencoder / feature-embedding anomaly models when defect examples are scarce. Synthetic data from CT simulation (for example Fraunhofer EZRT or aRTist-style simulators) and defect insertion help with rare defect classes.
• Rule layer stays deterministic: acceptance limits come from the standard and the drawing, not from the model; the model only supplies measured defects.
• Human-in-the-loop: low-confidence or safety-critical results go to a certified inspector; release of safety-relevant parts remains a human, qualified decision. Reviewer corrections feed retraining.

Data preparation

Data are radiographs, CT volumes and ultrasonic scans together with CAD models, drawings, acceptance standards and inspector reports. They live on the inspection devices' workstations, local file shares, the MES/QMS and sometimes on paper reports. Typical gaps: proprietary raw formats, missing or inconsistent calibration metadata, no link between a scan and the part, serial number and drawing revision, very few defect examples, defects labelled only in free-text reports, and scattered inspector knowledge.

Preparation steps: export scans to open formats (DICOM-like, TIFF, HDF5) with metadata; link each scan to part number, batch and production parameters via the ERP/QMS (PREP-005); restructure Excel inspection lists into tidy tables (PREP-011); digitise paper reports (PREP-006); harvest labels from past inspection verdicts (PREP-020); define the defect class scheme per standard (PREP-019); and build a gold test set labelled by certified inspectors (PREP-018). CT simulation and synthetic defect insertion help for rare classes.

Rough share of effort: data preparation and labelling typically take well over half of the project; model selection is the smaller part. Plan inspector time for labelling.

Implementation paths

Which path for whom

• SMB (foundries, weld shops, small manufacturers): Path A, because mature NDT software from X-ray/CT vendors and inspection service providers fits certified processes and requires no ML team.
• Enterprise (series production, automotive, aerospace suppliers): Path C, because inspection is core to quality, volume makes seat or API cost high, and part designs and CT data are trade secrets that should stay on-prem.
• Developer: Path C, because open 3D segmentation and anomaly models (nnU-Net, YOLO-type) plus CT simulation allow building custom inspection pipelines.

Path A — Buy (recommended for most manufacturers)

• Inspection hardware vendors: Waygate Technologies (Baker Hughes; Inspection Works / ADR software, phoenix CT), Zeiss Industrial Metrology (CT with analysis software), Yxlon (Comet group, ADR for castings), Nikon Metrology, North Star Imaging, Comet-Yxlon; EU-based options include Zeiss, Yxlon, Fraunhofer EZRT (research and transfer) and Volume Graphics (Heidelberg, part of Hexagon).
• Analysis software: VGSTUDIO MAX (Volume Graphics, with machine-learning based defect detection), Dragonfly (Object Research Systems, ML segmentation), Thermo Fisher Avizo, Hexagon software; Eddyfi and Olympus/Evident for ultrasonic (PAUT/TFM) with automated flaw recognition features.
• Choose A when: parts are standard (aluminium castings, steel welds), you buy X-ray/CT anyway, and the vendor can validate against your acceptance classes.

Path B — Build on cloud APIs

• Limited fit: raw CT volumes are huge (GB per part) and often confidential. Cloud makes sense for training and annotation (EU regions of Azure ML, AWS, Google Vertex, or EU providers such as OVHcloud and STACKIT), while inference runs at the edge or on-prem. A VLM (Claude, Gemini) can draft inspection reports from the structured findings but must not make the accept/reject decision.

Path C — Self-host (open source)

• Frameworks: PyTorch, MONAI, nnU-Net for 3D segmentation, Ultralytics YOLO or MMDetection for 2D, anomalib for anomaly detection, ONNX Runtime / TensorRT / OpenVINO for deployment; ASTRA Toolbox or TIGRE for CT reconstruction; 3D Slicer or napari for review.
• Public data: GDXray (X-ray images of castings and welds) for benchmarking, not for production [4].
• Hardware: HW3 (one 48 GB GPU) for 3D segmentation of full CT volumes; HW2 suffices for 2D radiographs at line speed; inline use needs a latency budget matched to the cycle time.

Path D — Custom training

Usually needed for a new part family or alloy: collect scans with ground truth from metallography, destructive sectioning or expert consensus (hundreds to thousands of annotated defects per class; defect-free parts are plentiful, defects are rare). Use simulation and defect insertion to augment, domain adaptation across scanners, and a locked validation set. Re-validate after any change of detector, voltage, filter, alloy or part design.

State of the art (2026)

• Vendors ship ML-based defect detection inside CT analysis and ADR products; trend is inline CT for castings with automated verdicts, and 3D segmentation replacing threshold-based porosity analysis [1][2][3].
• Weak spots: rare defect types, tiny cracks near the resolution limit, artefacts (metal streaks, ring artefacts), domain shift between scanners and settings, and over-confident scores. Calibrate uncertainty and keep a review band.
• Research: self-supervised pre-training, foundation-style 3D segmentation models and simulation-to-real transfer are active; production systems remain mostly supervised, validated per part.
• Qualification: standards for AI in NDT are still developing; organisations such as the German Society for NDT (DGZfP) and ASNT discuss qualification and validation of ADR, and ISO 9712 continues to govern personnel [5].

Shared resources & combinations

• Same stack (image-classification, segmentation, anomaly-detection): VIS-001 surface inspection, VIS-056 wafer-map defects, VIS-057 additive-manufacturing monitoring, VIS-043 pipe CCTV inspection.
• 3D data pipeline is shared with VIS-040 point-cloud workflows.
• Downstream: PRED-025 field-failure early warning, traceability linking scan results to serial numbers and production data (MES/QMS), supplier quality feedback.
• Annotation tooling and edge GPU servers are reusable across inspection use cases.

Effort, cost & KPIs

• Path A: weeks to months for commissioning and validation on your parts; cost is dominated by the CT/X-ray system and licences.
• Path C/D: 6-12 months incl. data collection, annotation, validation and documentation; ongoing retraining and revalidation.
• KPIs: probability of detection (POD) by defect size, false-reject rate, escape rate (missed defects), review rate, scan-to-verdict time, scrap and re-work cost, agreement with certified inspectors.

People & change

Involve management for budget and quality responsibility, then the quality or NDT lead as process owner. Certified inspectors are the key group: they label data, validate the model and keep release authority for safety-relevant parts. IT connects scanners, storage and the MES/QMS. Customers may need to accept AI-assisted inspection, so involve key customers and auditors early, and check qualification rules for NDT personnel and the applicable standards with legal/compliance.

Works council: depends. The system evaluates parts, not people, but if verdicts, override rates or throughput are logged per inspector, §87(1) no. 6 BetrVG applies (BAG: objective suitability is enough). §90(1) no. 3 requires informing the council in good time about AI-supported work processes. A works agreement on result logging helps. Personal data are minimal (inspector IDs), so a DPIA is usually not needed; the DPO confirms.

Support AI literacy (AI Act Art. 4): train inspectors on the model's limits, missed-defect risk and automation bias. Run a pilot on one product line in parallel with manual evaluation and compare miss rate and review time. Typical concern: deskilling or replacement; answer that the model pre-screens and humans decide on low-confidence and safety-critical cases. See GUIDE-009, GUIDE-008, GUIDE-004.

Risks, regulation & compliance

• EU AI Act: classification depends on use. If the AI is a safety component of a product under Annex I legislation (e.g. machinery, pressure equipment, medical devices) it may be high-risk (Art. 6); pure internal quality screening with human release is typically minimal risk. Not an Art. 5 practice. AI literacy (Art. 4) applies in all cases [6].
• Sector rules: ISO 9712 personnel certification, ISO 10675 (weld radiography acceptance levels), ASTM E155 reference radiographs, customer specifications, aerospace (Nadcap) and pressure-equipment rules often require validated procedures; changes to the software are changes to the NDT procedure [7].
• Radiation protection: X-ray and CT systems fall under national radiation protection law (e.g. StrlSchG in Germany); not changed by AI.
• GDPR: no personal data in the scans; operator and reviewer logs may be personal data.
• Risks and mitigations: missed defects (use POD studies, conservative thresholds, review band), model drift after scanner or process changes (monitoring, revalidation), automation bias (blind re-reads, sampled audits), trade-secret CAD data (on-prem or EU hosting).

Getting started

1. Define acceptance classes, defect types and the cost of misses versus false rejects per part family.
2. Collect 200-500 representative scans incl. known defects with ground truth (sectioning, expert consensus).
3. Benchmark the ADR/CT-analysis modules of two vendors on your data; compare against inspectors.
4. Run in shadow mode next to certified inspectors; measure POD and false-reject rate.
5. Document the validated procedure, define the review band and revalidation triggers, then go live.

Sources

1. Waygate Technologies — Inspection Works / automated defect recognition — https://www.bakerhughes.com/waygate-technologies
2. Volume Graphics — VGSTUDIO MAX CT analysis software — https://www.volumegraphics.com
3. Fraunhofer IIS — Development Center X-ray Technology (EZRT) — https://www.iis.fraunhofer.de/en/ff/zfp.html
4. Mery et al. — GDXray: database of X-ray images for NDT — https://domingomery.ing.puc.cl/material/gdxray
5. ISO 9712 — NDT qualification and certification of personnel — https://www.iso.org/standard/78552.html
6. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
7. ISO 10675-1 — NDT of welds, acceptance levels for radiographic testing — https://www.iso.org/standard/72118.html
8. nnU-Net — https://github.com/MIC-DKFZ/nnUNet