The camera saw the dent at 14:02:31.480. The PLC knew web tension was 12% low at that exact moment. The recipe had been modified fourteen minutes earlier, and an operator had acknowledged an override on the same zone. Four systems each held one quarter of the root cause — and in most plants, nobody joins them, because the clocks don\'t match, the exports take days, and the investigation runs on whoever remembers the shift. That is why complex defect RCAs stretch from days into weeks, while the documented fix — correlating alarms and process data with defects — has cut scrap by 15% just by putting the evidence in one place. iFactory's sensor fusion ledger joins vision detections, PLC tag history, recipe versions, and operator actions on one synchronized timeline — so Plant Copilot answers "why did this defect happen?" in seconds, with the evidence attached.
iFactory Sensor Fusion — Root-Cause Ledger
The Camera Sees the Dent. The PLC Knows Why. Fusion Tells You.
Vision detection timestamps, PLC tag history, recipe versions, and operator actions — joined on one millisecond-aligned ledger per defect, queryable in plain language through Plant Copilot.
4 streams
vision · PLC · recipe · operator, fused
Days to weeks
what complex RCAs cost today
15%
scrap cut by alarm-defect correlation
Seconds
to a ranked root-cause hypothesis
Four Systems, Four Clocks, Zero Joins
Every plant already records the evidence; almost none can connect it. The vision system stamps frames on its own clock, the PLC historian on another, the MES logs recipe downloads in minutes, and operator actions live in a shift log or nowhere. An RCA built on that starts with a week of exports and timestamp archaeology — and the standards literature is blunt about the alternative: an investigation based on operator memory is an investigation already compromised.
Clock Drift
Camera, PLC, and MES clocks disagree by seconds to minutes — enough to misattribute any defect to the wrong machine state.
Export Archaeology
Evidence assembly means CSV pulls, historian queries, and screenshots — days of work before a single "why" gets asked.
Memory-Based RCA
"The machine was acting up Saturday" is testimony, not telemetry — and it is what most 8D investigations actually run on.
Context Evaporates
Which recipe version was live, who overrode what — answerable today, gone by the time the customer claim arrives.
One Defect, One Ledger, Every Witness
Fusion turns each vision detection into a ledger entry that subpoenas every other system for the same moment: the tag values at that frame, the recipe version in force, and the operator actions in the window. This is the record an investigation starts from — already assembled, already aligned.
Root-Cause Ledger — Defect Event #D-3127
Dent · Camera 3 · Line 2
14:02:31.480
Vision
Dent class detected, zone B, severity 3 — frame and crop stored with the event.
14:02:31.470
PLC
Web tension TI-204 at 88% of setpoint — 12% low at the detection frame; low since 14:01:58.
13:48:12
Recipe
Recipe v4.7 downloaded to Line 2 — tension setpoint reduced 5% versus v4.6.
13:51:40
Operator
Manual override acknowledged on tension zone B by Shift A — comment: "compensating flutter."
Pick a defect from last month you never explained. Get a turnkey AI quote and we\'ll show you the ledger it would have had.
Ask the Ledger Like a Colleague
Plant Copilot sits on the fused timeline, so the investigation is a conversation instead of a query language. Anyone with a question gets the joined answer — engineer, shift lead, or the quality director writing the 8D.
Plant Copilot — Sensor Fusion
AI Manufacturing Lead
Why did camera 3 start flagging dents on Line 2 this afternoon?
Copilot
Dents started at 14:02 — 41 detections, all zone B. At every detection frame, web tension TI-204 reads 10 to 14% under setpoint. Two changes precede the onset: recipe v4.7 downloaded at 13:48 with a 5% lower tension setpoint, and a manual override on zone B at 13:51. The stack of the two explains the under-tension; dent rate tracks it almost exactly. Top-ranked cause: combined setpoint change plus override. Affected parts are tagged; a draft containment and a recipe-review task are ready for assignment.
AI Manufacturing Lead
Has v4.7 run anywhere else without dents?
Copilot
Yes — Line 4 ran v4.7 for six hours with zero dent detections and tension holding at setpoint, no overrides. That isolates the override as the differentiating factor on Line 2. Both ledgers are linked if you want the side-by-side for the 8D.
What Makes the Fusion Trustworthy
For an AI manufacturing lead, the interesting part isn\'t the demo — it\'s the joins. Fusion is only as good as its clock discipline and its honesty about confidence, so both are engineered, not assumed.
One Clock for Everything
Sources are time-synchronized at ingestion and drift-corrected continuously — frame-to-tag joins hold at millisecond resolution because the alignment is maintained, not hoped for.
Frame-to-Tag Joins
Each vision detection is joined to the full PLC tag vector at its frame time plus the preceding window — the machine state at the defect, not near it.
Versioned Context
Recipe versions, parameter diffs, and operator actions are first-class ledger entries — the slow-moving causes that pure signal correlation misses.
Ranked, Not Oracular
Hypotheses come ranked with confidence and the evidence behind each — counterexamples included — so the engineer judges, the model proposes.
Running your own data lake project for exactly this? Talk to our AI team and compare joins before you build another pipeline.
From Ledger to Closed 8D
The ledger feeds the frameworks you already use. D4 and D5 of an 8D — root cause and corrective action — stop being the weeks-long middle of the report and become its strongest pages, because every claim carries timestamps.
Evidence Pack per Defect
Frames, tag traces, recipe diffs, and operator log in one exportable record — the D4 attachment writes itself.
Containment by Timestamp
Affected parts are bounded by detection times and machine state — quarantines sized by data, not by caution.
Corrective Action, Tracked
The fix becomes a work order or recipe task linked to the ledger — and recurrence checks run against the same fusion automatically.
Pattern Library Grows
Closed cases teach the ranker — the recipe-plus-override pattern that took an afternoon today is flagged proactively next time.
Turnkey: Hardware, Software, Live in 6-12 Weeks
iFactory ships a pre-configured NVIDIA AI server — racked, software pre-loaded. Rack it, plug in power and Ethernet, and the AI is live inside your firewall. The engagement covers cabling, network, PLC/SCADA and vision integration, operator training, and 24×7 remote monitoring.
Phase 1 · Weeks 1-2
Sync & Ingest
Edge server on-prem; vision, PLC, recipe, and operator-log streams connected read-only and clock-aligned.
Phase 2 · Weeks 3-6
Fuse & Validate
Ledgers build on live defects; your engineers grade hypotheses against known cases until the ranking earns trust.
Phase 3 · Weeks 6-12
Copilot Live
Plain-language RCA for the whole team, 8D evidence packs on demand, 24×7 remote monitoring behind it.
1000+
clients running iFactory
6-12 wks
to live operation
On-prem
inside your firewall
What Fusion Gives Back
Joining the witnesses converts directly into investigations measured in minutes, quarantines measured in parts, and a defect pareto whose top entries finally have causes instead of theories.
Seconds
To first hypothesis
ranked causes with evidence replace the week of export archaeology
15%
Scrap reduction class
the documented payoff when defects and process data finally correlate
Bounded
Containments
affected parts tagged by timestamp and state, not quarantined by fear
Stronger
8D evidence
D4/D5 backed by synchronized telemetry instead of shift-log testimony
How long did your last serious RCA take, start to verified fix? Start a 6-week pilot and time the difference.
Frequently Asked Questions
How does the system keep four sources on one clock?
Sources are time-synchronized at ingestion against a common reference, and per-source drift is measured and corrected continuously rather than assumed away. Vision frames join to PLC tag vectors at millisecond resolution; slower sources like recipe downloads and operator actions are placed on the same timeline with their native precision, clearly marked.
Does this require replacing our vision system or PLCs?
No. Fusion reads what exists — detection events and frames from your vision systems, tag history from your PLCs and historians, recipe versions from MES, operator actions from HMI and log sources — all inbound-only. If a line lacks vision, iFactory can add AI vision as part of the deployment, but existing systems are first-class sources.
How are root-cause hypotheses generated and ranked?
By evidence, transparently. The engine looks for state changes and anomalies that precede and co-occur with the defect pattern — tag deviations, recipe diffs, overrides — and ranks candidates by how well each explains onset timing, location, and counterexamples (the same recipe running cleanly elsewhere, for instance). Every hypothesis ships with its evidence and its confidence; the engineer makes the call.
We have a data lake project for this. Why a ledger instead?
A lake stores the streams; the ledger joins them. The hard parts — clock discipline, frame-to-tag alignment, versioned recipe context, and a ranking layer engineers can interrogate — are exactly what generic lake projects spend years rebuilding. The ledger arrives with those solved, runs on-prem next to your lake, and can feed it.
What ships in the turnkey package, and where does the data live?
A pre-configured NVIDIA AI server with software pre-loaded — rack it, plug in power and Ethernet, and it\'s live. The engagement covers cabling, network, PLC/SCADA and vision integration, training, and 24×7 remote monitoring, with go-live in 6 to 12 weeks. Everything runs on-premise inside your firewall, read-only and inbound-only — frames and tag data never leave the plant.
Every Witness. One Timeline. Cause in Seconds.
See Your Defects on the Fusion Ledger
Bring one unexplained defect and the systems that watched it happen. We\'ll build its ledger live — frames joined to tags, recipe diffs and operator actions on the timeline, ranked hypotheses with evidence — and scope the 6-to-12-week turnkey deployment, on-prem, inside your firewall.
4 streams
one fused ledger
ms-aligned
frame-to-tag joins
1000+
clients · 99.9% uptime
