The future of industry starts with semantic modeling

Everyone in manufacturing is talking about AI. Predictive maintenance. Autonomous quality inspection. Self-optimizing production lines. The vision is compelling. But most industrial AI projects stall — not because the algorithms are wrong, but because the data is a mess.

The uncomfortable truth: AI cannot reason about data it doesn't understand. And right now, most industrial data is unlabeled, unstructured, and inconsistent across machines. A temperature reading from Line 1 looks nothing like a temperature reading from Line 3, even though they measure the same thing on identical machines.

The fix isn't better AI. The fix is better data — specifically, semantically modeled data. And the technology to do that already exists. It's called OPC UA information modeling.

What is semantic modeling?

Semantic modeling means giving data a meaning that machines can interpret automatically — not just a value, but a context.

Consider a raw data point: 42.7. What is it? A temperature? A pressure? A speed? In what unit? Is it normal? Without context, it's meaningless.

Now consider the same data point in a semantic model:

Object: Pump_3 (type: CentrifugalPump)
Variable: DischargeTemperature
Value: 42.7
Unit: °C (UNECE code: CEL)
Range: 20.0 – 65.0
Status: Good
Timestamp: 2025-04-22T08:15:32Z

Suddenly, any system — human or AI — can understand what this data means, whether it's within normal bounds, and how it relates to other variables on the same pump.

That's semantic modeling: turning numbers into knowledge.

OPC UA: the semantic backbone

OPC UA is the only industrial communication standard that includes a full information modeling framework. It doesn't just transport data — it describes what the data is.

In OPC UA, every piece of data lives in an address space — a structured graph of objects, variables, methods, and events. Objects have types. Variables have engineering units and value ranges. Methods have input and output arguments. Events carry structured payloads.

This isn't metadata bolted on after the fact. It's the foundation of the protocol. When an OPC UA client connects to a server, it can browse the entire data model at runtime — discovering what data exists, what it means, and how it's organized. No documentation required. No custom mapping.

Companion Specifications: industry-standard models

But semantic modeling only works if everyone agrees on the model. If every machine vendor defines "temperature" differently, you're back to the mapping spreadsheet.

This is where Companion Specifications come in. A Companion Specification is a standardized OPC UA information model for a specific industry or machine type. It defines the objects, variables, and relationships that all machines of that type should expose.

Key examples:

PackML (ISA-88 / ISA-TR88). Defines a state machine for packaging machines — states like Execute, Idle, Held, Aborted. Any machine that implements PackML exposes the same state model, regardless of vendor.

Euromap 77 / 83. Defines information models for injection molding machines and extrusion lines. A machine from Engel and a machine from KraussMaffei expose the same variables for cycle time, shot weight, barrel temperature.

PADIM (Process Automation Device Information Model). Defines models for process instrumentation — pressure transmitters, flow meters, temperature sensors. A Siemens transmitter and an Endress+Hauser transmitter expose the same diagnostic and measurement variables.

PLCopen. Defines a model for motion control systems — axes, programs, status.

MTConnect (via OPC UA). Defines models for CNC machines — spindle speed, feed rate, tool wear.

There are over 80 published Companion Specifications today, covering industries from robotics to commercial kitchen equipment.

Why this matters for AI

Here's the connection that most AI discussions miss:

1. AI needs consistent features

Machine learning models require consistent, labeled features. If "discharge temperature" is called DT_03 on one machine, temp_discharge on another, and TI_342 on a third, you spend 80% of your time on data wrangling and 20% on actual modeling.

Companion Specifications eliminate this. Every machine of the same type exposes the same variables with the same names, types, and units. Your feature engineering becomes trivial — or automatic.

2. AI needs context, not just values

An anomaly detection model that sees a value of 42.7 knows nothing. An anomaly detection model that knows the value is a discharge temperature on a centrifugal pump, that the normal range is 20–65°C, and that the pump has been running for 3,200 hours since last maintenance — that model can make a meaningful prediction.

OPC UA's information model provides this context as structured, machine-readable metadata.

3. AI needs to transfer across assets

The real value of industrial AI isn't training one model for one machine. It's training a model on 50 machines of the same type and deploying it to the 51st without retraining.

This only works if all 50 machines expose the same data model — the same variables, in the same structure, with the same units. That's exactly what a Companion Specification guarantees.

4. AI needs to explain its decisions

A model that says "Alert: anomaly detected in variable X17A" is useless to an operator. A model that says "Alert: discharge temperature on Pump_3 exceeds the 95th percentile for this operating state" is actionable.

Semantic models give AI the vocabulary to explain itself in terms that operators understand.

The modeling gap

If semantic modeling is so powerful, why isn't everyone doing it?

Modeling is hard. Creating a good OPC UA information model requires deep domain knowledge — understanding the equipment, the process, and the relevant Companion Specifications. It's a skill that sits at the intersection of automation engineering, software architecture, and domain expertise.

Tooling has been scarce. Until recently, building an OPC UA information model meant writing XML NodeSet files by hand — a tedious, error-prone process that scared off most teams.

The benefits are invisible at first. The value of a semantic model only becomes apparent when you try to scale — when you add the 10th machine, deploy the 3rd analytics tool, or onboard a new vendor. By then, retrofitting a model is expensive.

This is changing. Tools like Sterfive's OPC UA Modeler let you define information models in YAML, validate them in real time, generate NodeSet files automatically, and version-control them in Git — the same workflow that software developers have used for decades.

The path forward

The factories that will lead in industrial AI are not the ones with the best algorithms. They are the ones with the best-modeled data.

The path is clear:

Model your processes using OPC UA information models and relevant Companion Specifications. Start with one machine type. Get the model right.
Deploy OPC UA servers that expose this model at the edge. Every machine of the same type exposes the same address space.
Build applications on top of the model, not on top of raw tag lists. Analytics, AI, dashboards, MES connectors — all consume the same semantic data.
Scale horizontally. Add machines, add lines, add plants. The model is the contract. Applications work everywhere the model is deployed.
Let AI do what AI does best — find patterns, predict failures, optimize processes — on data that is clean, consistent, and self-describing.

The future of industry isn't AI. It's semantically modeled data — with AI on top.

References

OPC Foundation — OPC UA Part 5: Information Model. opcfoundation.org
OPC Foundation — Companion Specifications. opcfoundation.org/developer-tools/documents
ISA — ISA-88 / PackML: Packaging Machine Language. isa.org
Euromap — Euromap 77/83: OPC UA for Injection Molding and Extrusion. euromap.org
FieldComm Group — PA-DIM: Process Automation Device Information Model. fieldcommgroup.org

Sterfive builds the tools that make semantic modeling practical: OPC UA Modeler for defining information models, node-opcua for implementing them in software, and Omni-Edge for deploying them at scale. If you're ready to model your processes, talk to our team.