Synchronised ingestion
Camera streams, PLC and SCADA tags, MES batch events and operator terminal actions ingested with a common time base.
AI intelligence system
LineTrace
From defect detection to defect causality.
A multimodal system that correlates line video, machine telemetry, batch records and operator actions to explain why a defect happened, not merely that it did.
- Sector
- Manufacturing & Production Quality
- Year
- 2025
- Duration
- 10 months
- Team
- 8 people
- Platforms
- Edge · Web console
- Status
- In production
12s
Defect-to-event window
Median correlation lookback
4 streams
Aligned on one clock
Video, PLC, batch, operator
Ranked
Hypotheses, not verdicts
Each with its own evidence set
Days to minutes
Investigation time
For recurring defect classes
01 — The problem
The plant already had vision inspection. It reliably flagged defective units and stopped there. Every flag opened an investigation that a quality engineer ran by hand: pull the video, pull the PLC history, find the batch sheet, interview the shift.
Investigations took days and usually ended in a plausible story rather than a demonstrated cause. By the time anyone concluded that an extruder temperature excursion was responsible, four more hours of product had gone through the same excursion.
02 — What we did
The core insight is that causality lives in time alignment. We built a temporal correlation engine that puts video frames, PLC tags, batch events and operator actions on one clock, so a defect at 14:32:07 can be walked backwards across every modality simultaneously.
Rather than a black-box classifier, the system produces ranked hypotheses with the evidence attached: the deviation window, the frames, the tag traces and the historical precedents. A quality engineer confirms or rejects a hypothesis, and that judgement feeds back into ranking.
Detection tells you a unit is bad. Causality tells you what to change.
Interface
What it looks like in use
Two of the screens that carry the most weight in daily use, rebuilt here from the production design system.
Defect 48213 · 14:32:07T−12s
Extruder 3 · temp deviation +6.4°CRANK 1
Tool change · 41 min priorRANK 2
Material lot 88-CRANK 3
PLC tagVideo · Batch · Operator aligned
One clock, four streamsVideo, PLC tags, batch events and operator actions share a time base, so a defect can be walked backwards across all of them at once.
Hypothesis 1 · Extruder 3EVIDENCE
Frames 14:31:55 – 14:32:0914 CLIPS
Deviation window12.4s
Historical precedents7 MATCHES
Engineer decisionPENDING
Association score 0.81
Hypotheses, not verdictsEach candidate cause opens with its frames, tag traces and historical precedents already assembled for an engineer to confirm or reject.
Capabilities
What the system does
Grouped by the job each set of capabilities exists to do, rather than by which team built it.
Multimodal monitoring
01 · Everything on one clockClock drift correction
Continuous alignment against a reference source, because a two-second drift destroys causal reasoning silently.
Edge preprocessing
Frame selection and feature extraction at the line so bandwidth carries signal rather than raw footage.
Retention windows
Full-fidelity retention around flagged events and downsampled retention elsewhere.
Causal analysis
02 · The reason the system existsDefect-to-event correlation
Each defect is walked backwards across every modality within a configurable window to find coincident deviations.
Temporal root-cause ranking
Candidate causes scored on temporal proximity, historical association and deviation magnitude.
Machine behaviour modelling
Per-asset normal envelopes learned from history, so deviation is judged against that machine rather than a spec sheet.
Batch intelligence
Defect clustering by material lot, shift, tool change and recipe version to separate systemic from incidental causes.
Investigation workflow
03 · Built for a quality engineerEvidence bundles
Every hypothesis opens with its frames, tag traces, batch context and precedents already assembled.
Explainable review
The reasoning path is inspectable at each step rather than delivered as a score.
Engineer feedback loop
Confirmed and rejected hypotheses adjust ranking for future occurrences.
Recurrence tracking
Once a cause is confirmed, its signature is watched for and flagged on first recurrence.
Architecture
How it is put together
Layer by layer, with the reason each one exists — because the reason is usually the interesting part.
01Edge layer
Industrial gateways performing frame selection, feature extraction and buffering, resilient to plant network interruption.
- NVIDIA Jetson
- GStreamer
- OPC UA
02Time alignment
A synchronisation service that normalises timestamps across sources and continuously corrects drift.
- PTP
- Drift correction
03Vision models
Defect detection and segmentation trained per product family, with frame-level embeddings retained for similarity search.
- PyTorch
- Detection
- Embeddings
04Correlation engine
Windowed multi-stream analysis producing candidate causes with proximity, magnitude and historical-association scores.
- Time-series analysis
- Association mining
05Console
Investigation workspace with synchronised video scrubbing against tag traces and batch context.
- React
- WebCodecs
- TimescaleDB
Technology
Edge
- NVIDIA Jetson
- GStreamer
- OPC UA
- MQTT
AI & Vision
- PyTorch
- ONNX Runtime
- Vector search
Backend
- Python
- FastAPI
- TimescaleDB
- Kafka
Frontend
- React
- TypeScript
- WebCodecs
- Visx
Outcome
What changed
Measured against how the operation ran before, not against a benchmark chosen after the fact.
-
Investigations moved from days to minutes. Evidence assembly, previously the bulk of the work, is automatic.
-
Causes are demonstrated, not argued. Every hypothesis carries the frames and traces that support it.
-
The system learns from its engineers. Confirmations and rejections change how future hypotheses rank.
-
Recurrence is caught on the first repeat. Confirmed signatures become monitored patterns.
How it ran
-
Months 1-2
Instrumentation audit
Catalogued every available signal and, critically, measured the clock drift between them.
-
Months 3-4
Time alignment
Built and validated synchronisation before any modelling, because nothing downstream works without it.
-
Months 5-7
Vision and correlation
Per-family detection models and the windowed correlation engine.
-
Months 8-9
Investigation console
Synchronised scrubbing, evidence bundles and the feedback loop.
-
Month 10
Line rollout
Deployed line by line with a shadow period against manual investigation on each.
Related work
From the same practice
Telecommunications & Network Operations
NetCause
Not alarm detection. Failure hypothesis reduction.
A correlation engine that turns an alarm storm into a short ranked list of probable root causes by reasoning over network topology and fault propagation rather than alarm counts..
2025 · 8 months
Oil & Gas · Asset Integrity
PipeIntegrity
From a single inspection to longitudinal asset integrity.
A vision system that detects and localises pipeline deterioration, then tracks each defect across inspection cycles so maintenance is prioritised by progression rather than by appearance..
2024 · 9 months
Pharmaceutical & Regulated Manufacturing
BatchResolve
From deviation to evidence, with the audit trail intact.
A deviation investigation assistant that assembles batch records, equipment telemetry, inspection imagery, SOPs and historical deviations into an evidence-linked case file a quality unit can defend to a regulator..
2025 · 11 months
Something similar?
Running into the same problem LineTrace solved?
Tell us what you run. We will reply within two business days with what we would build for your situation — and, just as usefully, what we would leave out.