Non-Destructive Testing (NDT)¶
Last updated: 2026-04-11
Recent Finds¶
Phased-Array Thermography (PAT): Beam-Steering Applied to Infrared NDT (Scientific Reports, Nov 2025)¶
Southampton aeronautics researchers introduce PAT — applying phased-array beam-steering logic to infrared thermography rather than ultrasound. An array of independently controlled embedded heating elements steers and focuses thermal wavefronts into the structure. This overcomes a core limitation of conventional pulsed thermography: the inability to detect defects perpendicular to the heated surface (inclined cracks, certain delaminations). Validated on aerospace specimens — aluminium plates with flat-bottom holes and CFRP composite panels with impact damage. PAT found flaws that standard pulsed thermography missed entirely. A 3D post-processing step further resolved small, deep defects invisible in 2D analysis. Particularly promising for in-situ structural health monitoring (SHM) of composite airframes because heating elements can be embedded into the structure.
Eddy Current In-Line Inspection for Oil & Gas Pipelines: Systematic Review (MDPI Processes, Jan 2026)¶
Open-access systematic review comparing five ECT methods deployed on pipeline inspection gauges (PIGs) running inside live O&G lines: conventional ECT, multi-frequency ECT, remote-field ECT (RFECT), pulsed ECT (PECT), and array ECT (AECT). Key differentiators: RFECT excels at through-wall corrosion in ferromagnetic pipes; PECT (broadband excitation) gives superior depth-profiling of wall-thickness loss; AECT provides high-speed, high-resolution mapping suitable for corrosion-under-insulation (CUI). The review maps each technique to defect type (corrosion pits, stress corrosion cracking, weld anomalies) and identifies AI-augmented signal interpretation + miniaturized sensor arrays as the next frontier for high-pressure gas transmission infrastructure.
CATT-S: CNN + Transformer for PAUT Defect Detection at 99.4% Accuracy, Real-Time CPU (Sensors MDPI, Oct 2025)¶
Changwon National University benchmarks deep learning architectures on PAUT data from welded and composite specimens. Their novel CATT-S model (Convolutional Attention Temporal Transformer for Sequences) combines CNN layers for local signal morphology with Transformer self-attention to capture cross-beam spatial dependencies — the relational information between adjacent scan lines that pure CNNs miss. Result: 99.4% accuracy, F1 = 0.905 on real experimental data vs. 94.9% for standard CNNs. Critically: the exported ONNX model runs real-time on CPU, making deployment on portable field hardware practical without GPU. Directly applicable to aerospace weld inspection and pipeline girth-weld screening where throughput and operator-independence are both requirements.
Advances of Machine Learning in Phased Array Ultrasonic NDT (MDPI, 2025)¶
Comprehensive review of ML applied to phased array ultrasonic testing (PAUT). Key finding: ML integration improved average detection precision from 78% to 93.3% and recall from 66% to 91.5% across benchmark datasets. The paper surveys CNN-based defect classifiers, U-Net segmentation for B-scan images, and transformer-based anomaly detection. Takeaway: the bottleneck is now labeled data quality, not model architecture — industrial NDT datasets are small and proprietary.
Advanced NDT Integration in Steel Processing (MDM Metals, March 2026)¶
Practical engineering overview of multimodal NDT in steel manufacturing. Combines radiographic testing (RT), eddy current (EC) for surface cracks, and infrared thermography for subsurface delamination. Real-time AI analytics fused across modalities enables comprehensive coverage without the blind spots of any single technique. Highlights the trend toward inline NDT (in the production line) vs. offline batch inspection.
Automated Defect Detection in Multimodal Phased Array Plane Wave Imaging (NDT.net)¶
Research on automating defect detection across multiple PAUT imaging modalities (plane wave, full matrix capture). ML classifiers trained on multimodal representations outperform single-modality models. Particularly relevant for aerospace structural health monitoring where false negatives have catastrophic consequences and false positives drive costly rework.
Core Concepts¶
NDT Method Taxonomy¶
| Method | Principle | Best For | Limitation |
|---|---|---|---|
| Ultrasonic Testing (UT) | Sound wave reflection/attenuation | Subsurface cracks, thickness | Couplant needed, rough surfaces |
| Phased Array UT (PAUT) | Electronically steered beam array | Complex geometries, rapid scanning | Cost, data volume |
| Eddy Current (ECT) | Electromagnetic induction | Surface/near-surface cracks in conductors | Depth limited, lift-off sensitivity |
| Radiographic (RT/DR) | X-ray / gamma-ray transmission | Porosity, inclusions, welds | Radiation safety, 2D only |
| Magnetic Particle (MT) | Ferromagnetic field disruption | Surface defects in ferrous materials | Ferrous only |
| Thermography (IRT) | Thermal anomaly from defects | Delaminations, moisture ingress | Emissivity issues |
| Visual (VT) | Direct or camera-aided inspection | Surface defects | Surface only |
Key Terminology¶
- PAUT (Phased Array UT): Array of piezoelectric elements fired with controlled delays to steer and focus the beam electronically. Produces B-scan (cross-section) and C-scan (top-view) images.
- FMC/TFM (Full Matrix Capture / Total Focusing Method): All elements fire and receive independently; TFM reconstructs high-resolution images in post-processing. Computationally heavy but gold standard.
- POD (Probability of Detection): Statistical measure of an NDT system's ability to detect defects of a given size. Critical for certification (e.g., ASME, EN standards).
- TOFD (Time of Flight Diffraction): Uses diffracted waves at crack tips for precise depth sizing.
AI in NDT — Current State (2026)¶
- Detection: CNNs and transformers applied to B-scan / C-scan images approach human expert accuracy on known defect types.
- Characterization: Harder — requires regression for defect size/shape, not just classification.
- Challenges: Domain shift (different materials, geometries, equipment), lack of standardized datasets, regulatory acceptance.
- Trend: Physics-informed neural networks (PINNs) incorporating wave propagation equations to reduce data requirements.
Standards & Certification Bodies¶
- ASNT (American Society for Nondestructive Testing) — NDT Level I/II/III certification
- ISO 9712 — International NDT personnel qualification
- ASME Section V — NDT procedures for pressure vessels and piping
- EN 4179 / NAS 410 — Aerospace NDT certification
Open Questions¶
- What is the minimum labeled dataset size needed to train a reliable PAUT defect classifier for a new material/geometry combination?
- How do PINNs compare to purely data-driven models when labeled data is scarce?
- Will regulators (FAA, EASA, ASME) accept ML-based NDT decisions for safety-critical components, and what validation framework will they require?
- How does multimodal sensor fusion handle conflicting indications across NDT modalities?
- Can PAT (Phased-Array Thermography) be miniaturized and embedded into structural panels for continuous in-flight SHM without adding prohibitive weight?
- CATT-S runs real-time on CPU — what's the performance floor? Can it maintain accuracy at full PAUT scan speeds (>100 mm/s mechanical scanning)?
- RFECT vs MFL for ferromagnetic pipelines: under what wall-thickness and defect-geometry conditions does RFECT outperform magnetic flux leakage?