MODULE 3
From Prototype to Production

The Factory That Knows Itself

Industrial IoT, digital twins & industrial AI
From Prototype to Production — FramtidR
Framing
A dashboard is not a decision
A dashboard shows you data; a data-driven factory acts on it.
  • Source. Where does the data come from — and can you trust it?
  • Action. Who or what acts on the insight, and how fast?
  • Failure. What breaks when a sensor, network, or cloud fails?
  • We follow data from a vibrating spindle to a maintenance work order.
Most “smart factory” pilots stall not on the analytics but on the data plumbing, sensor reliability, and the human/MES workflow around the insight. (Plumbing = how data moves and is stored.)
FramtidRThe Factory That Knows Itself2 / 30
Part 1

From the spindle to the cloud

Architecture, sensors, protocols & where data lives
FramtidRThe Factory That Knows Itself3 / 30
IIoT architecture
A three-floor building
Data volume shrinks and the time window grows as you move up.
  • Tier 1 — Sensors. Pick up raw signals from machines and robots.
  • Tier 2 — Edge. Clean, shrink and study data near the source; low delay.
  • Tier 3 — Cloud. Combine data, train models, plan logistics.
CloudcombinedEdgecleaned, low delaySensors / Actuatorsraw, high-frequencymore delay OK ↑ / less data ↓
Figure: the IIoT funnel — lots of data at the base, checked and chosen at the top.
FramtidRThe Factory That Knows Itself4 / 30
Sensing
Giving the machine senses
  • Vibration / accelerometers. Early warning of bearing, gear and imbalance faults.
  • Thermocouples / RTDs. Temperature of motors, tooling, process media.
  • Acoustic emission. Stress waves from cracks, friction, tool wear.
  • IO-Link (IEC 61131-9). Point-to-point link carrying values, parameters and diagnostics.
IO-Link turns a plain 4–20 mA sensor into a device that names itself and reports its own health — it is point-to-point, not a fieldbus.
FramtidRThe Factory That Knows Itself5 / 30
The edge
Deciding at the source
Edge ≠ PLC — keep exact, real-time control apart from analytics.
PLC
exact real-time control
  • Fixed real-time scan cycle
  • Safety-rated logic
  • No test-stage ML
  • Don’t change it weekly
Edge gateway
analytics & connectivity
  • Filters and thins out raw data
  • Spots odd readings locally, sub-second
  • Containerized apps (e.g., Docker)
  • Stores and forwards when cloud drops
A clean split: the edge node does analytics, buffering and connectivity; the PLC runs the exact, real-time control on a fixed timing.
FramtidRThe Factory That Knows Itself6 / 30
Protocols
A common language for the floor
OPC UA and MQTT work together, not against each other.
  • OPC UA (2008). Describes its own data; strong device-to-edge and machine-to-machine.
  • MQTT (1999). Light publish/subscribe; works well over slow links; edge-to-cloud.
  • Rule of thumb. OPC UA below the gateway, MQTT above it.
  • Always use TLS. Add client certificates; plain, unencrypted data is a risk.
MachinesEdgeCloudOPC UAMQTTTLS + client cert
Figure: Machines —[OPC UA]→ Edge —[MQTT/TLS]→ Cloud.
FramtidRThe Factory That Knows Itself7 / 30
ISA-95 / Purdue
The plant in layers
One map of where each system lives and who may talk to whom.
  • ISA-95 (IEC 62264) & Purdue share the same level hierarchy.
  • Levels 0–2. Process, sensors/control, and supervisory SCADA/HMI.
  • Levels 3–4. MES (site operations) and ERP / logistics (enterprise).
  • Level 3.5 DMZ. Separates OT (0–3) from IT (4); basis for IEC 62443 zoning.
L4 — Enterprise / ERPITL3.5 — DMZL3 — Site ops / MESL2 — SCADA / HMIL1 — Sensors / basic controlL0 — Physical processOTOT 0–3
Figure: Levels 0–4 with a highlighted 3.5 DMZ; OT below, IT above.
FramtidRThe Factory That Knows Itself8 / 30
Storage
Where the data comes to rest
Hot at the edge, warm in the historian, cold in the lake — check and tidy data before you store it.
Hot — edge buffer
seconds–minutes; fast to query
Warm — time-series historian
days–months; fast time-range queries (+ MES events / part history)
Cold — data lake / lakehouse
years, raw, shaped when read; feeds model training; cheaper at scale
FramtidRThe Factory That Knows Itself9 / 30
Part 2

A living copy of the machine

Digital model, shadow, twin — and the thread that links them
FramtidRThe Factory That Knows Itself10 / 30
Definitions
Model, shadow, twin — it’s the arrows
Told apart by how the data flows, not by how detailed the model is (Kritzinger et al., 2018).
Digital modelphysicaldigitalmanual, both waysDigital shadowphysicaldigitalauto, one-wayDigital twinphysicaldigitalauto, bidirectionalThe difference is the arrows, not the detail.Origin: Grieves’ 2002 “Conceptual Ideal for PLM”; the name “digital twin” later coined by NASA’s John Vickers.
Figure: model = manual both ways, shadow = auto one-way, twin = auto bidirectional.
FramtidRThe Factory That Knows Itself11 / 30
Another view
Same idea, three jobs
A second way to sort twins — by what the twin is of (its life stage). Design and production twins often start as models; only the live asset twin needs a real-time link.
1
Product (design)
Test performance on screen with FEA/CFD before cutting metal — a model.
▶
2
Process (production)
Simulate the line — reach, cycle times, bottlenecks; test PLC code — model → shadow.
▶
3
Asset (performance)
Use live IIoT data; estimate remaining useful life (RUL) — a true twin.
FramtidRThe Factory That Knows Itself12 / 30
Digital thread
One thread through the whole life
  • Digital thread. Continuous information flow across the lifecycle (NIST).
  • It connects the twins. Without it, twins are isolated islands.
  • PLM is the backbone. Config, BOM, engineering change orders, as-designed vs as-built.
  • Payoff. Full traceability, fewer redesigns, shorter cycle time.
DesignMakeInspectOperateServicePLMBOM / ECO / config
Figure: a lifecycle loop with one bold thread; PLM is the central system of record.
FramtidRThe Factory That Knows Itself13 / 30
Part 3

Hearing failure before it happens

Industrial AI: predictive maintenance, copilots & the metaverse question
FramtidRThe Factory That Knows Itself14 / 30
Predictive maintenance
Hearing failure before it happens
Watching machine condition is the top ML use case on the floor — faults show up in the vibration signal before the part fails. Strategy ladder: reactive → preventive → predictive → prescriptive.
1
Sensor
raw vibration
▶
2
Features
FFT / time-frequency
▶
3
ML model
fault class + RUL
▶
4
MES order
an action someone owns
FramtidRThe Factory That Knows Itself15 / 30
Worked example
A bearing, six weeks of warning
One predictive-maintenance loop — each step you can check, the RUL kept rough (no fake precision).
wk 0wk 3wk 6baselinepeak emergingouter-race faultnormalwatchopen work order —replace at next planned stopBaseline → detection → diagnosis → RUL window (“weeks, not days”) → scheduled action.
Figure: an FFT peak grows over six weeks; the state bar moves normal → watch → planned swap.
FramtidRThe Factory That Knows Itself16 / 30
Edge vs cloud ML
Train centrally, decide locally
Common pattern: train in the cloud, run the model at the edge — push models down via containers.
Edge ML
decide locally
  • Fast, low-delay answers
  • Works when the cloud link is down
  • Keeps sensitive raw data on site
  • Spots odd readings; simple classifiers
Cloud ML
train centrally
  • Heavy model training
  • Learns and retrains across the whole fleet
  • Compares sites
  • Needs lots of data from many machines
Trade-offs to weigh: delay, bandwidth/cost, where the data wants to live, privacy, and how it copes offline.
FramtidRThe Factory That Knows Itself17 / 30
Generative AI
The technician’s copilot
  • Industrial copilots. Plain-language assistants for technicians and engineers.
  • Real use cases. Find past records and schematics, draft PLC code, explain OEE numbers.
  • Productivity gains. Real but vendor-claimed — test them in your own setting.
  • Helps, does not act alone. A person stays in the loop, by a small majority in surveys.
Strong use cases are finding records, writing docs and drafting code — backed by manuals, past records and live sensors, with a check step. Key risk: a confident wrong answer (hallucination) in a safety job — which is exactly why copilots only assist.
FramtidRThe Factory That Knows Itself18 / 30
MES / ERP / OEE
Turning signals into output
Analytics matter only when they meet the systems that run the plant.
  • MES / MOM. Schedules work, tracks each part’s history, receives maintenance & quality actions.
  • ERP. Materials, orders, costing, logistics — the insight shows up as cost & delivery.
  • OEE = A × P × Q. Predictive → Availability; quality → Quality; cycle-time → Performance.
  • The trap. High OEE on parts nobody ordered is not progress.
Map each AI use case to the lever it moves — that is how you justify a data project to operations.
FramtidRThe Factory That Knows Itself19 / 30
Industrial metaverse
What’s real, what’s sold
Fund the left, question the right.
Real today
measurable ROI
  • Virtual prototyping
  • Line / production simulation
  • Training
  • Remote collaboration
Hype
mostly not proven yet
  • Always-on, full-immersion 24/7 factory
  • A headset for everyone
  • “Replace the physical plant”
  • Broad consumer-metaverse claims
FramtidR stance: adopt the proven simulation/training pieces; be wary of the all-or-nothing vision.
FramtidRThe Factory That Knows Itself20 / 30
Part 4

When the network reaches the machine

OT cybersecurity — and the 90% below the waterline
FramtidRThe Factory That Knows Itself21 / 30
IT/OT convergence
When the network reaches the machine
  • New cyber-physical risk. Industry 4.0 connects isolated OT to IT and the internet.
  • OT priorities differ. Uptime and safety first; decade-long lifecycles; no quick patching.
  • A hacked PLC is not a data breach — it is a physical event.
  • Connect safely. Connecting IT and OT is unavoidable; the job is to do it safely.
The flip: IT values keeping data secret, OT values uptime and safety — where IT and OT overlap is the new place attackers aim at.
FramtidRThe Factory That Knows Itself22 / 30
Zero trust for OT
Trust nothing by default
  • Segmentation. Split IT and OT with a Level 3.5 DMZ and industrial firewalls (Purdue zoning).
  • Zero trust. Give each user only the access they need; encrypt internal traffic; limit who can change PLC logic.
  • Safe onboarding. Device certificates and HSMs (secure key chips) check every node before it joins.
  • Frameworks. IEC 62443 (zones & conduits) and NIST SP 800-82 — complementary.
IT zoneDMZ3.5OT cell AOT cell Bfwdevice certs
Figure: IT — firewall — DMZ (3.5) — firewall — segmented OT cells, each a conduit-connected zone.
FramtidRThe Factory That Knows Itself23 / 30
Historical lesson
What Stuxnet taught the shop floor
  • 2010, Siemens S7 PLCs. Hit centrifuges at Natanz, Iran.
  • Crossed an air gap via an infected USB; used several unknown Windows flaws.
  • PLC rootkit. Changed motor speeds while playing back normal readings.
  • Lesson. Air gaps are not perfect; both the control logic and the data must be trusted.
USBeng.workstnS7 PLCVFDfreqcentri-fugeoperator HMI sees “normal”
Figure: USB → workstation → S7 PLC → VFD → damage, with a parallel “HMI sees normal” deception lane.
FramtidRThe Factory That Knows Itself24 / 30
Data quality
The feed is the product
  • Models are only as good as the data. Calibration, drift, missing samples, wrong labels.
  • Drift is silent. A sensor that slowly loses calibration quietly spoils every model.
  • What you need. A calibration schedule, checks, clean timestamps, clear ownership.
  • Garbage in, confident garbage out. Bad data plus ML gives wrong answers that look right.
Budget calibration and data validation as first-class engineering work — the feed is the product.
FramtidRThe Factory That Knows Itself25 / 30
Governance & first step
Who owns the data — and where to start
  • Ownership. Which team owns each data domain (sensor, MES, quality)?
  • Definitions. A shared dictionary so “downtime” means one thing.
  • Access & rights. Who can read/change/delete — and who owns vendor/cloud data?
  • Retention. How long and where raw vs curated data lives.
Start small: instrument one painful asset → prove one closed loop (sensor → insight → a real work order someone owns) → measure vs baseline → replicate. Settle “who owns the data” in the contract, not after.
FramtidRThe Factory That Knows Itself26 / 30
Prototype → production
The 90% below the waterline
A dashboard demo is ~10% of a production data system.
  • Data plumbing. Reliable ingestion, store-and-forward, schemas, time-sync.
  • Calibration & drift. Ongoing, not one-time.
  • Network reliability. Keeps working when links drop; edge buffering.
  • Security & integration. IEC 62443 / NIST SP 800-82; insight reaches MES/ERP.
  • Ownership. A named human or system that acts on the insight.
seen vs worksdashboard demodata plumbingcalibration & driftnetwork reliabilitysecurityintegration (MES/ERP)ownership
Figure: the demo is the tip; the unseen 90% sits below the waterline.
FramtidRThe Factory That Knows Itself27 / 30
Recap
What to carry out of the room
◢
IIoT is a funnel
Sensors → edge → cloud; right protocol per link; map on ISA-95/Purdue.
⇅
Say it precisely
model = manual, shadow = one-way auto, twin = bidirectional auto; PLM is the backbone.
⚛
AI’s best wins
Predictive maintenance and helper copilots; be wary of the all-or-nothing metaverse.
🛡
Security is physical
Segment (DMZ/Purdue), zero-trust, IEC 62443 / NIST SP 800-82; data quality is the foundation.
FramtidRThe Factory That Knows Itself28 / 30
Further Reading
IIoT & protocols
  • OPC UA Protocol & MQTT Synergy — EMQ Technologies
  • IO-Link (IEC 61131-9) — IO-Link Community / Wikipedia (CC BY-SA)
Digital twin & thread
  • A Literature Review of the Digital Thread — Systems, MDPI (CC BY 4.0)
  • Digital Thread for Smart Manufacturing — NIST (public domain)
  • Digital twin / shadow / model — Oxford Insights (2023); Michael Grieves, Wikipedia
AI & predictive
  • Vibration Monitoring for Predictive Maintenance — Eng. Proceedings, MDPI (CC BY 4.0)
  • Industrial metaverse: Hype or reality? — HCLTech
Security & standards
  • What is the Purdue Model? — Software Toolbox
  • IEC 62443: Zones, Conduits, Defense in Depth — IoT Worlds
  • Since Stuxnet: Critical Infrastructure Attacks — Forescout
FramtidRThe Factory That Knows Itself29 / 30
⚠

Before you rely on this

This series was made by FramtidR with the help of AI. AI can make mistakes, including about people. Please check this information before you use it.

Made with AI help • checked by FramtidR
FramtidRThe Factory That Knows Itself30 / 30