In-situ process monitoring & layer-wise defect detection in metal additive manufacturing

Metal parts built by laser powder-bed fusion (L-PBF) or directed energy deposition (DED) can hide pores, lack-of-fusion zones and recoater damage that only show up in expensive CT after the build. In-situ monitoring uses cameras, optical tomography and melt-pool sensors during the build, with ML models that localise anomalies layer by layer. Operators can stop a doomed build early, and quality teams get a per-part evidence trail that reduces (but does not yet fully replace) CT.

Problem & value

• Today: quality is verified after the build by CT, destructive tests of witness coupons and process-parameter freezing. Builds run for many hours or days, so a recoater crash or powder shortage found late wastes machine time, powder and expensive alloys.
• AI value: early abort or repair decisions, targeted CT instead of blanket CT, traceable per-part quality data, and faster process development. Defect detection must be tied to the 3D position in the part, which is why the signals are stacked into volumetric data.
• Evidence: OEMs and research institutes (e.g. Fraunhofer ILT) report correlations between OT or melt-pool signatures and CT-detected porosity, but correlation strength depends on material, machine and defect type; there is no universal replacement of CT today [1][2][3].
• Qualification role: aerospace, medtech and energy customers need process monitoring as part of the qualification and certification evidence, alongside standards from ISO/ASTM TC 261 / F42 [4].

Variants

• Powder-bed (L-PBF) layer imaging: high-resolution images per layer to detect recoater streaks, short feeds, part warping and super-elevation.
• Optical tomography / NIR layer integration: time-integrated emission per layer to detect hot spots and cold zones linked to porosity.
• Melt-pool monitoring: coaxial photodiodes or cameras at very high sampling rates for spatter, keyhole and lack-of-fusion signatures (time-series plus image models).
• DED / wire-arc (WAAM): melt-pool geometry and thermal camera tracking, with closed-loop control of power or travel speed.
• Multi-machine fleets: cross-machine drift detection and build-to-build comparison.
• Research / university labs: open datasets and custom sensor retrofits.

How it works

1. sensors (layer camera, OT/NIR, melt-pool photodiode/camera, machine log)
2. synchronisation & registration to slice geometry (x, y, layer z)
3. preprocessing (flat-field, powder-bed mask, normalisation per machine)
4. models: image anomaly detection / segmentation + time-series classifier
5. voxel-level fusion into 3D defect-probability volume
6. rules & thresholds (zone-specific: critical vs non-critical regions)
7. action: alarm, pause/abort, (optional) parameter correction, CT sampling plan
8. archive: build report linked to serial number, powder batch, machine state

• Unsupervised/semi-supervised anomaly detection (autoencoders, PatchCore-style feature distance) handles rare defects; supervised segmentation (U-Net/YOLO-family) is used where CT labels aligned to layers exist.
• Melt-pool signals: 1D CNN, transformer or gradient-boosted classifiers on photodiode streams.
• Closed loop: real-time correction exists mostly in research and vendor pilots; production use is typically alert-only with a human decision.
• Human-in-the-loop: process engineer confirms abort/continue; quality engineer approves release and CT sampling; model drift reviewed after each recalibration of the machine.

Data preparation

Process data is large, fast and rarely labelled; plan for it from the start.

• Data and location: layer images, melt-pool photodiode and camera streams, and machine logs sit on the machine controller or a local server; part, material batch and build-file data sit in MES, ERP or CAD/PLM; CT scans and test results are in the quality department's file shares.
• Typical gaps: proprietary machine formats, no time alignment between sensor streams and layers, no link between a flagged layer and the final CT or tensile result, too few defect examples, and unclear ownership of build data between operator and customer.
• Preparation steps: export raw data via machine APIs or interfaces, align everything to a common layer and coordinate index, link builds to batch and post-process results, define a defect scheme (lack of fusion, keyholing, spatter), harvest labels from CT and metallography, and hold back complete builds as test data. Store data with retention rules for qualification evidence.
• Tools: Python with NumPy, OpenCV and a time-series store, a labelling tool such as CVAT, object storage.
• Share of effort: high, typically 50-60 percent.

Related preparation use cases: PREP-005, PREP-019, PREP-020, PREP-018.

Implementation paths

Which path for whom

• SMB (service bureaus, small manufacturers): Path A, because machine makers and specialist vendors offer mature, integrated melt-pool and layer-monitoring packages with EU support, and an SMB rarely has the data science capacity to build its own.
• Enterprise (aerospace, automotive, energy, medical): Path C, because build data and qualification records are trade secrets, fleets of machines make licence costs high, and one internal platform can serve several processes.
• Developer (monitoring vendors, research institutes): Path C, because open-weight vision and time-series models on HW2 edge hardware give full control of sensor data and certification evidence; Path D adds custom training.

Path A — Buy

• Machine-integrated monitoring: EOS (EU, Germany) EOSTATE suite (Exposure OT, MeltPool, PowderBed, Contour) with EOSTATE software; Nikon SLM Solutions (EU, Germany) Melt Pool Monitoring and Layer Control; Trumpf (EU) TruPrint with melt-pool and powder-bed monitoring; Renishaw InfiniAM Spectral; Velo3D Sapphire (Assure); Additive Industries (EU, NL) and Colibrium Additive (formerly GE Additive) with in-machine monitoring [1][2].
• Third-party analysis software: Materialise (EU, Belgium) Control Platform / Build Processor modules for traceability, Authentise, Oqton MES, Siemens, Hexagon, Dyndrite (data-to-build).
• Choose A when: you already own an OEM machine; vendor models are calibrated to that machine and material, and warranty/qualification needs vendor support.

Path B — Build on cloud APIs

• Fit is limited: raw sensor data is large and process IP is sensitive. Use cloud (Azure ML, Vertex AI, AWS SageMaker in EU regions, or EU clouds such as OVHcloud, STACKIT) for training, fleet analytics and dashboards only; run inference at the edge.
• Data plumbing: machine data via OPC UA / MTConnect / vendor APIs, time-series stores (InfluxDB, TimescaleDB), MLOps (MLflow).
• VLMs are of limited use for the raw signal, but helpful to summarise build reports for non-experts.

Path C — Self-host (open source)

• Models/tools: anomalib (PatchCore, PaDiM), Ultralytics YOLO or segmentation models, PyTorch/Lightning, Segment Anything for annotation, tsai/sktime for melt-pool signals, ParaView and 3D Slicer for volume visualisation, Open3D for registration [5].
• Serving: ONNX Runtime / TensorRT on an industrial PC with GPU next to the machine; Kafka or MQTT for streaming; NVIDIA Triton for multi-machine servers.
• Hardware: HW1 for per-layer inference on a single machine, HW2 for retraining and fleet analytics.

Path D — Custom training

Worth it when vendor tools do not cover your alloy, machine or geometry (e.g. refractory alloys, large DED). Needs build campaigns with deliberately seeded defects, CT ground truth registered to layers, and a validated labelling protocol. Expect months of data generation; the data is the bottleneck, not the model.

State of the art (2026)

• Hybrid sensor fusion (layer images + OT + melt-pool) with 3D registration to CT is the research and product direction; vision foundation models (DINOv2-class features) improve few-shot anomaly detection on small datasets.
• Synthetic data and physics-informed models (melt-pool simulation) augment rare defects.
• Closed-loop control remains largely research / pilot; most plants use monitoring to inform decisions rather than to autonomously change parameters.
• Limits: weak signal-to-defect link for sub-100 µm pores in some materials, machine-specific models that do not transfer, false alarms from powder-bed artefacts, long labelling cycles and limited public benchmark datasets. Certification bodies accept monitoring as supporting evidence, not yet as sole proof of quality for critical parts [4].

Shared resources & combinations

• Same vision stack: VIS-001 (surface inspection), VIS-056 (wafer-map classification, similar anomaly pipeline), VIS-026 (measurement).
• Downstream: VIS-053 (CT/X-ray evaluation, ground truth and sampling), PRED-014 (virtual quality sensors), ANOM-005 (process anomalies from sensors), OPT-009 (offline parameter optimisation using monitoring data).
• Hardware: edge GPU PC (HW1/HW2) shared with other line-side inspection tasks.

Effort, cost & KPIs

• Path A: weeks to enable on a new machine; licences and sensors are priced by OEM; calibration campaign per material.
• Path B/C: 6–12 months to a validated model including CT-aligned datasets; main cost is data generation, CT time and process-engineering effort.
• Path D: 12+ months for qualification-grade evidence.
• KPIs: probability of detection vs CT (per defect class and size), false-alarm rate per build, scrap cost avoided, share of parts released with reduced CT, time-to-detect, model drift across machines/builds.

People & change

Management and the head of additive manufacturing or quality set the goal (scrap reduction, qualification evidence). Machine operators and process engineers are the end users; involve them early because they judge alarms and supply labels. IT and IT security integrate machine networks and protect build data, data owners (operations, quality, customers) regulate access, and legal or compliance check customer contracts and qualification rules in aerospace, medical or energy.

Works council: depends. The system observes the machine and part, not people, but alarm acknowledgements and operator interventions are logged per user, which makes the system objectively suited to monitor behaviour; §87(1) no. 6 BetrVG then applies, so include the works council before the pilot and inform it under §90(1) no. 3. Public-sector labs involve the Personalrat. Data protection: usually no personal data beyond operator IDs; a short DPO check suffices.

AI literacy: support operators with practical training on reading alerts, false positives and the limits of the model (AI Act Art. 4). Run a pilot on one machine in advisory mode next to normal inspection, compare against CT results, and only then link alerts to process actions. Typical concerns: loss of operator authority and fear of blame; address them by keeping the operator able to ignore or confirm alerts and by documenting that final quality decisions stay with people.

See GUIDE-009, GUIDE-011, GUIDE-008 and GUIDE-004.

Risks, regulation & compliance

• EU AI Act: minimal risk for monitoring of production. If the model is a safety component of a regulated product (e.g. medical device manufacturing) the product regulation applies; the monitoring tool itself is usually process equipment. AI literacy (Art. 4) applies [6].
• GDPR: no personal data in the sensor streams; operator logs and user accounts may be personal data.
• Sector rules: aerospace (EASA Part 21 and customer specifications), medical devices (MDR, ISO 13485 process validation), energy (PED, ISO 17025 labs). Process changes need revalidation.
• IP and data security: build data reveals proprietary parameters; keep it on-premises or in EU clouds; consider NIS2 and Cyber Resilience Act for connected machines.
• Risks/mitigations: model drift after machine maintenance (recalibration and golden builds), over-trust in anomaly maps (keep CT sampling), traceability of model versions in the quality record.

Getting started

1. Inventory machines, materials and part criticality; decide which monitoring signals are available.
2. Run a seeded-defect build with CT and register layers to CT volumes.
3. Start with the OEM monitoring software in alert-only mode and compare alarms with CT findings.
4. Train a lightweight anomaly model on your own good builds; evaluate per-defect POD.
5. Define governance: thresholds, abort authority, CT sampling policy and model change control.

Sources

1. EOS — EOSTATE monitoring suite — https://www.eos.info
2. Nikon SLM Solutions — Melt Pool Monitoring — https://www.nikon-slm-solutions.com
3. Fraunhofer ILT — Additive manufacturing and process monitoring — https://www.ilt.fraunhofer.de
4. ISO/TC 261 Additive manufacturing standards — https://www.iso.org/committee/629086.html
5. anomalib (open-source anomaly detection) — https://github.com/open-edge-platform/anomalib
6. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
7. ASTM F42 Committee on Additive Manufacturing Technologies — https://www.astm.org/membership-participation/technical-committees/committee-f42