Every vendor in industrial IoT claims to offer "digital twins." Most of them are dashboards with a 3D model bolted on. A real digital twin is something more specific and more useful: a synchronized software model that mirrors a physical asset's state, understands its behavior, and predicts its future. Here's what building one actually involves.

What it is and what it isn't

A dashboard showing live sensor readings? That's monitoring. A 3D model of a machine? That's visualization. Neither is a digital twin.

A digital twin needs three things. A state model — a structured representation of the asset's current condition. Every valve position, every temperature, every runtime hour. A behavior model — mathematical relationships between states. "When bearing temperature exceeds 85°C and vibration FFT peaks at 147Hz, the bearing is entering early failure mode." And a synchronization engine — bidirectional data flow keeping the digital and physical in lockstep.

Without the behavior model, you have a fancy dashboard. Without synchronization, you have a simulation. You need all three.

The data pipeline nobody wants to build

Industrial equipment generates messy, heterogeneous data at wildly different rates. A CNC machine might produce spindle vibration at 25.6 kHz, temperature readings every 10 seconds, tool position every 100ms, and event logs asynchronously.

You can't pipe all of this to the cloud. Not economically, and not at the rates required. So you tier it.

Edge processing handles the high-frequency stuff. Run the FFT on vibration data locally, send the frequency spectrum to the cloud — not the raw samples. Time-series telemetry flows through MQTT into a time-series database (we use TimescaleDB or InfluxDB depending on the project). Events get their own reliable, ordered channel.

The heuristic we use: anything needing sub-second response stays at the edge. Anything needing cross-machine correlation or historical analysis goes to the cloud. The edge does the heavy lifting so the cloud can do the thinking.

Keeping the twin in sync

The twin needs to reflect reality within a latency budget. Under 5 seconds for monitoring. Under 500ms for control. That sounds easy until you deal with: variable network latency, sensors reporting stale data, and the fact that the physical world doesn't have transactions.

We use event sourcing. Every state change is an immutable event with a timestamp. The twin reconstructs state by replaying events in order. When events arrive out of order — and they will — the twin re-applies them chronologically.

Conflict resolution is simple: physical reality wins. If the twin predicts one state and a sensor reports another, the sensor reading overrides. The twin adjusts its behavior model. Over time, the model gets better. But it never overrules the real world.

Predictive maintenance: the actual payoff

This is why you build a digital twin. Once the behavior model is calibrated against real operating data, it starts spotting patterns that lead to failures.

A compressor's twin learns that discharge temperature, suction pressure, and current draw follow a specific curve during healthy operation. When the real data deviates from the model — even slightly — that's an early warning.

The practical challenge is false positives. An uncalibrated model cries wolf constantly. Operators start ignoring alerts. The system becomes useless.

We solve this with confidence scoring. Alerts are classified as informational (log only), advisory (notify maintenance), or actionable (schedule intervention). New models start everything as informational. As the model accumulates validation data and proves its predictions, alerts get promoted. This isn't glamorous. It's patience. But it's how you build trust with the maintenance team — and trust is what determines whether the system actually gets used.

Key Takeaways

• A real digital twin = state model + behavior model + synchronization. Not a dashboard.
• Process high-frequency data at the edge. Send features, not raw samples.
• Event sourcing handles out-of-order telemetry. Physical reality always wins conflicts.
• Start predictions as informational. Earn trust with the operations team gradually.