01 · Vision
◫
Visual quality
Learns normal production appearance and localizes deviations on each part.
- Defect heatmaps
- Inspection ranking
- Surface, PCB, packaging or product
Industrial Predictive Quality
From plant data to prioritized decisions
Visual defect detection, machine-risk scoring and process-deviation alerts built from data the plant already produces.
Industrial visionSensors & SCADASelf-supervised learningLocal deployment
◫
Visual inspection
Heatmaps and part ranking
Heatmaps and part ranking
∿
Machine risk
Physical signals and context
Physical signals and context
↗
Process deviation
Expected vs observed evolution
Expected vs observed evolution
◎
Prioritization
Patch → part → cycle → lot → line
Patch → part → cycle → lot → line
The problem
Signals appear before the cost
The goal is not to generate more data or more alarms. It is to identify what changed, where it changed and what deserves attention first.
01 · Signals
Cameras, vibration, current, SCADA and test logs
Rich but fragmented information at different spatial and temporal scales.
→
02 · Hidden cost
Late defects, scrap, micro-stops and reactive maintenance
Impact becomes visible when the intervention window is already small.
→
03 · Decision
A prioritized, explainable and measurable queue
What to inspect, which asset to review and when the deviation started.
Modular offer
Three modules, one decision system
02 · Sensors
∿
Machine risk
Combines physical signals, process variables and calibrated models.
- Top assets/cycles
- Lead time and false alarms
- Engineered + learned features
03 · Process
↗
Temporal deviation
Compares expected process evolution with what is actually observed.
- Local surprise and persistence
- Window- or cycle-level change
- Lot- and line-level aggregation
Architecture
From raw data to an actionable alert
The system combines strong industrial baselines with self-supervised representation learning. Complexity is kept only when it adds measurable value.
Plant dataImage, sensor, SCADA, recipe, asset
→
RepresentationDINOv2, physical features, temporal tokens
→
ModelsPatchCore, PaDiM, GBT, world model
→
DecisionHeatmap, risk, top-k, lead time
Baseline first
Rules, physical features and classical models define the standard any advanced model must beat.
Self-supervision where useful
Learns structure from large volumes of unlabeled images and process sequences.
Operational hierarchy
Local scores become risk by part, cycle, tool, lot, shift or production line.
Visual module
Visual inspection that can start with normal images
◫
Dense features + normality memory
DINOv2/ResNet extract local descriptors. PatchCore and PaDiM compare each region with normal history to localize deviations.
- Image- or part-level score
- Patch-level heatmap
- Threshold calibrated on normal production
- Ranking to reduce manual inspection
Visual inspection queue
Part #A-1842
0.94
Part #A-1837
0.81
Part #A-1819
0.62
Part #A-1811
0.38
Pilot output: ranking, heatmap and comparison against normal references.
Sensor and process module
A learned signal that complements process engineering
The practical standard remains operational context + physical features. The token-level model is retained when it delivers a measurable incremental signal.
0.6609
Metadata-only
Contextual AUPRC baseline in the current diagnostic.
0.6766
Metadata + EWMA surprise
Temporal surprise adds a small but positive signal.
0.6960
Metadata + token world model
Best no-action combination; Δ AUPRC +0.0352.
Why self-supervision matters
Learn normal operation without waiting for defects
0
Defect labels required to begin learning normalityFor visual representation, anomaly detection and temporal deviation.
1
Unlabeled data already has value
Normal images, sensor sequences and production cycles can teach structure and expected evolution.
2
Labels improve measurement; they are not a universal prerequisite
Defects, scrap and failure events allow recall, lead-time and calibration analysis when available.
3
Two valid outputs
Without labels: anomaly and deviation. With historical events: supervised risk and economic validation.
Technical evidence
Results that justify validation on real plant data
1.00
Visual AUPRC
DINOv2 + PatchCore/PaDiM in a quick MVTec bottle run.
+0.0352
Sensor delta
Token world model over metadata-only in the current diagnostic.
4–6
Weeks
To audit, benchmark, model and deliver a go/no-go recommendation.
What this supports
The visual architecture produces a strong public-benchmark signal, and the dense sensor model adds incremental temporal information.
What we do not assume
A public result does not guarantee performance on another line. Value is confirmed on the client’s real distribution, defects and operating conditions.
Pilot
A short path to an investment decision
Week 1
Data map
Inventory, quality, traceability, temporal structure and use-case definition.
Weeks 1–2
Baselines
Rules, classical models and a strong visual foundation to set a realistic reference.
Weeks 2–5
Modeling
Heatmaps, risk scoring, temporal surprise, calibration and error analysis.
Week 6
Decision
Local demo, final benchmark, expected impact and integration roadmap.
Deliverables
A demonstration the team can see, measure and challenge
- Data manifest and quality audit.
- Benchmark against rules and strong baselines.
- Visual heatmaps and inspection ranking.
- Asset/cycle risk and temporal deviation curves.
- Operational metrics: precision@top-k, false alarms and lead time.
- Local demo, executive report and technical documentation.
- Go/no-go recommendation and integration architecture.
Top operational alertslocal demo
Line 2 · Lot 184
High
Machine 04 · Cycle 912
High
Part B-3381
Review
Night shift · Process 7
Watch
Starting data
Start with what exists; labels are optional
Minimum to begin
- Normal images or process time series.
- Timestamps, part/asset/cycle identifiers and temporal ordering.
- Available operating context: recipe, material, setpoints or machine.
- A technical owner who can interpret the process and its limits.
Increases value, but does not block the pilot
- Labeled defects or inspection outcomes.
- Failure, scrap, rework or maintenance history.
- Cost of false alarms and missed defects.
- Verified temporal commands for action-conditioned evaluation.
Self-supervision learns normality without labels. Historical events are used when a specific supervised target must be measured or calibrated.
Success criteria
The pilot ends with a quantitative answer
↓
Less unnecessary inspection
Focus manual review on the parts or cycles with the highest measured risk.
↑
Earlier warning
Detect deviation before the defect, stoppage or reactive maintenance event.
✓
Defensible go / no-go
Decide whether to integrate, collect more data or stop without relying on black-box promises.