What Is a Digital Twin in Manufacturing, Really?
In manufacturing, a digital twin is a digital model of a physical asset, line or process that is kept up to date with data from the real equipment. What makes it a twin is that connection. A static drawing or a one-off simulation is not a twin. In practice, plants build twins at one of three levels: a live dashboard that mirrors current state, a simulation that uses plant data to test "what if" questions, or a closed loop that sends decisions back to the process. Each level depends on the one before it, and all of them depend on reliable plant data.
Why is the term so hard to pin down?
"Digital twin" covers a wide range of things. The same phrase can describe a 3D animation of a production line, a trend screen, a physics model of a heat exchanger or a scheduling optimiser. Some of these are useful, and some are only presentations.
A practical working definition has three parts:
- A model. It describes the asset or process: its structure, its states and, at higher levels, how it behaves.
- A data connection. Real measurements update the model at an agreed frequency.
- A purpose. The model answers a specific operational question that someone has to decide on.
If any of the three is missing, you have something else. Without a data connection, it is a design model. Without a purpose, it is a visualisation project that will be hard to justify maintaining.
What are the practical levels of a digital twin?
It is more useful to think in levels than to treat a twin as a single product. Each level adds capability and adds requirements.
| Level | What it does | Typical question it answers | Main dependency |
|---|---|---|---|
| 1. Monitoring twin (dashboard) | Mirrors current and historical state of an asset or line | "What is this machine doing now, and what did it do last shift?" | Trustworthy, contextualised live and historical data |
| 2. Simulation twin | Uses a behavioural model, fed and checked against plant data, to test scenarios | "What happens to throughput if we change the changeover sequence?" | A validated model plus enough history to calibrate it |
| 3. Closed-loop twin | Feeds model outputs back to setpoints, schedules or recipes, with defined limits and approvals | "Which setpoint should this line run at for the next order?" | Proven model accuracy, controlled write paths and change governance |
Level 1: the monitoring twin
This is where most plants should start, and it is often where most of the practical value sits. A monitoring twin presents live and historical data in the structure of the plant: lines, machines, units and states. It is closer to a well-designed dashboard than to anything exotic. If you are building one, the guidance on designing a production shift dashboard applies directly.
Level 2: the simulation twin
A simulation twin adds a model of behaviour: how material flows, how a machine responds to load, how a buffer fills. It earns its place only when it is checked against real results. A simulation that has never been compared with plant history is an assumption with a graphical interface.
Level 3: the closed-loop twin
At this level, the model's output changes what the plant does, either by writing to the control system or by issuing schedules and recipes. This is where risk rises sharply. Any write path into control needs defined limits, a fallback when the model or communications fail, and clear ownership of changes. Anything that touches a safety function must be reviewed with a qualified assessor. The article on preventing duplicate commands between MES and PLC covers one of the integration risks that appear here.
What data do you need before building a digital twin?
Every level rests on the same foundation. Before committing to a twin, work through this checklist. Gaps here will show up later as a model nobody trusts.
Data prerequisites checklist
- ] Signal access. You can read the required values from existing controllers or add sensors where none exist. See [collecting data from an existing PLC.
- ] A historian or equivalent store. Time-series data is retained long enough to show normal behaviour, variation and failures. See [when a plant needs a data historian.
- [ ] Consistent timestamps. Data from PLCs, SCADA, MES and other sources can be aligned in time.
- [ ] Standard engineering units and naming. The same quantity carries the same unit and a clear, consistent tag name everywhere.
- [ ] Appropriate sampling intervals. Values are captured often enough to show the behaviour the model needs, and deadbands do not hide events.
- ] Context. Machine data is linked to product, batch, order and shift so you can compare like with like. See [adding product and shift context to machine data.
- [ ] Data quality status. Bad, stale or substituted values are flagged rather than silently passed on.
- [ ] Measurement confidence. Critical instruments have calibration records and a way to detect drift.
- [ ] Equipment states. Running, idle, stopped, faulted and changeover states are defined and recorded the same way across machines.
- [ ] Network and security design. Data leaves the machine network through a controlled path, not a direct connection from office systems.
- [ ] Ownership. Someone is responsible for each data source, the model and the decisions it supports.
If several of these boxes are unchecked, the right next project is usually data cleanup rather than a twin. The guidance on cleaning plant data before building dashboards is a sensible starting point.
What does a sensible first project look like?
Illustrative example: a plant has a packaging line with recurring short stops and wants to "build a digital twin" of it.
A practical sequence might be:
- Define the question. "Where on the line do short stops start, and what conditions precede them?"
- Check the prerequisites. The team confirms which machine states and counts are available, adds context for product and shift, and aligns timestamps between the filler and the case packer.
- Build the monitoring twin. A line view shows each machine's state, buffer levels and stop history in one timeline.
- Review the evidence. After a period of operation, the team looks for patterns before deciding whether a simulation would add anything.
- Decide on the next level. Only if the question needs scenario testing, such as changing buffer sizes, does the team move to a simulation twin, and it checks the model against recorded history before relying on it.
In this example, the monitoring level may answer the original question on its own. That is a reasonable outcome, not a failure to reach a "real" twin.
When does a digital twin not make sense?
A twin is not the right tool in every case. Consider a simpler approach when:
- The question is already answered by existing reports. A better trend or a corrected report may be enough.
- The process changes faster than the model can be maintained. Frequent product, layout or equipment changes can make a behavioural model obsolete before it pays back the effort.
- The data foundation is weak. A twin built on unaligned, unlabelled or drifting data will produce confident-looking but unreliable answers.
- Nobody will act on the output. If no decision depends on it, the model becomes a maintenance burden.
What are the common mistakes?
Starting with the 3D model. Visual fidelity is easy to show and rarely the hard part. Start with the data and the question.
Skipping validation. Any behavioural model must be compared with recorded plant results before people rely on it, and checked again after process changes.
Jumping straight to closed loop. Writing to control systems without proven model accuracy, limits and a fallback adds operational risk.
Treating the twin as a one-off project. Models, tag mappings and context rules need an owner and a change process, just like PLC programs.
Ignoring data quality flags. A model that consumes bad or stale values without knowing it will mislead operators.
Frequently asked questions
Is a SCADA screen a digital twin?
It can serve as a basic monitoring twin if it reflects live state in a structured model of the plant and supports a defined purpose. Most SCADA screens are designed for operation rather than analysis, so a twin usually adds history, context and a model of the asset.
Do we need a historian before building a digital twin?
For almost every practical twin, yes, or an equivalent time-series store. Simulation and closed-loop levels in particular need history to calibrate and validate the model.
Can a digital twin work on older equipment?
Often, provided the required signals can be read from existing controllers or added with retrofit sensors. Older equipment usually needs more work on signal access, naming and context.
Which level should we aim for?
Aim for the lowest level that answers your operational question. Move up a level only when there is a specific decision the current level cannot support and the data prerequisites are in place.