Table of Contents
Data contextualisation means linking each historian tag, the time series a sensor produces, to the equipment it measures, that equipment's place in the plant hierarchy, and what was happening at the time. In practice it is a maintained mapping from each tag to a functional location, with units, a measurement role and the dates each link is valid.
Key takeaways
- One vendor defines contextualisation as linking "time series, files, and diagram tags to assets (functional locations)" (Cognite). Without it, a model or an assistant sees thousands of columns with no meaning.
- Tag names alone are unreliable. A Honeywell patent notes there are "often multiple distinct naming conventions used within a specific plant" and calls tag mapping "the single largest effort" (US 11017008 B2).
- Map tags to the functional location, not the serial number, and use the CMMS installation history to know which equipment was in place on a given date.
- Do the matching in layers: curated mappings, then rules, then machine learning suggestions, with an engineer confirming what is uncertain.
- Templates per equipment type let one mapping, one calculation and one model be reused across every similar asset.
What does contextualisation add to a time series?
A historian tag holds a timestamp, a value and a status. What an engineer knows about it, such as which pump, which unit and whether the pump was running, sits elsewhere. OSIsoft's best-practice guidance describes a "smart element", an asset in its model, as combining time series such as flow, pressure and vibration; asset details such as name, model and manufacturer; external data such as last maintenance date and design documents; and events such as downtime, startup and failure (Batke, OSIsoft, 2019).
That gives three kinds of context:
| Kind of context | Question it answers | Typical source |
|---|---|---|
| Asset | Which equipment and position does this tag measure? | CMMS functional locations and equipment, tag register, P&IDs |
| Process | What is upstream and downstream, and which tags belong to the same loop? | P&IDs, control system configuration |
| Time | What was the equipment doing: running, starting up, in maintenance, on which grade? | Event logs, CMMS work orders, batch or shift records |
Why is a tag name not enough?
Because "tag" means three different things. An instrument tag on a P&ID follows a convention such as ISA-5.1, whose purpose is "a uniform means of designating instruments" for measurement and control (ISA). An equipment tag in the CMMS is a functional location, the position a pump or vessel occupies. A historian tag is neither: it is the name of a data stream, often inherited from the controller that produced it.
The historian name may carry the instrument tag, a controller address or something else. In an OSIsoft training lab, the same "Sealer Temperature" is stored as FFS01SealTemp on one machine and 02ST3000 on the next. The lab explains that this "often occurs whenever the PI Points are named after the original data source items", which follow the naming conventions of different PLC and SCADA manufacturers (OSIsoft, PI World 2020 lab).
Equipment also moves. In SAP, a functional location is one of the "hierarchical structures with a systematic order", while equipment is "an individual physical object" that "can be installed at functional locations"; installations and dismantlings can be analysed from both sides (SAP Learning). The transmitter on a pump's discharge measures the position, so map the tag to the functional location and let the installation history say which serial number was there on each date. Sensors that travel with the machine, such as some vibration probes, are the exception and should be flagged.
Which hierarchy should tags hang from?
Use one that already exists and that maintenance trusts. OSIsoft's best-practice guidance says the hierarchy should follow a standard such as ISA-95, and that the ERP hierarchy "can be good source to start" (Batke, OSIsoft, 2019).
The ISA-95 equipment levels, as listed in the OPC UA companion specification, run from enterprise, site and area down to process cell and unit for process plants, and include equipment module and control module levels below them (OPC 10030). For equipment classes, ISO 14224 provides a basis for collecting reliability and maintenance data in a standard format, including an equipment taxonomy, for petroleum, petrochemical and natural gas industries (ISO 14224:2016). CFIHOS adds tag and equipment class hierarchies and a reference data library, and all its documents are free to download (IOGP JIP36).
Keep it shallow. The OSIsoft lab warns that a hierarchy with, for example, eight levels to drill down to the equipment of interest frustrates users (OSIsoft, PI World 2020 lab). Knowledge graphs for plant equipment build on the same CMMS backbone.
What should a tag-to-equipment mapping record?
Each row should be checkable by a person and usable by a pipeline.
| Field | Example | Why it matters |
|---|---|---|
| Historian tag | 21PI1043.PV | The key used to pull data |
| Functional location | DZ01-U21-P-101A | Ties process data to maintenance history |
| Equipment class | Centrifugal pump | Selects the template and the comparable assets |
| Measurement role | Discharge pressure | Lets a model ask for "discharge pressure" on any pump |
| Engineering unit | bar(g) | Prevents mixing units across sites |
| Valid from / to | 2024-03-01 / open | Handles re-instrumentation and moved tags |
| Match method | Rule, model suggestion, manual | Shows how much to trust it |
| Confirmed by, date | Process engineer, 2026-06-12 | Makes the mapping auditable |
The tag and location values above are invented examples. For units, OPC UA's EUInformation structure maps engineering units by default to UN/CEFACT codes, so a client can interpret them programmatically (OPC 10000-8).
How do you match thousands of tags to equipment?
In layers, from the most certain to the least. Cognite's matching pipeline "applies manual mappings, then regex rules, then machine learning", and its documentation advises preferring "deterministic rules and curated mappings before you retrain a model" (Cognite, entity matching).
| Method | How it works | Strength | Limit |
|---|---|---|---|
| Curated mapping | Engineers fill a table, often from the instrument index | Reliable | Slow; does not scale to every tag |
| Naming rules | Regular expressions extract unit, loop and equipment codes from the tag name | Fast and explainable | Breaks wherever conventions differ |
| Similarity model | Splits names into tokens and scores candidate pairs | Covers what rules miss | Needs some similarity between the strings; scores are not probabilities |
| Engineer review | Draft matches confirmed against the P&ID | Catches errors | Takes time from scarce people |
Cognite's model "can only find matches when there are some similarities between the strings", and its feature scores "can't be interpreted as a probability". An unsupervised model suits a start with few known matches; a supervised one can learn from confirmed matches once you have some (Cognite, entity matching). In the user interface, a person confirms each suggested match and reviews the drafts before they are saved, and confirmed regular-expression rules are saved with the pipeline (Cognite, match entities).
Research points the same way. A study on about 29,000 asset tags from 50 upstream oil and gas facilities trained an LSTM on tag characters to predict hierarchy relationships. It reported good results on relationships but limitations in predicting some numeric elements of tags (Pronk, 2024). Those numbers often tell two identical pumps apart, so a person still checks.
How do templates save work?
Once tags are mapped to roles on one pump, the same structure should apply to every pump. In AVEVA PI Asset Framework, element templates make "every boiler, every pump, or every transformer" share a consistent view, and a new calculation built on the template can be applied automatically to every asset that uses it (OSIsoft, PI World 2020 lab). Templates can also create new historian points with consistent names (Batke, OSIsoft, 2019).
Where old tags follow no pattern, the same lab describes associating each attribute with its point by hand, using PI Builder, an Excel add-in, to edit configurations in bulk. A unified namespace does the same job at publish time: the topic path places each value in the hierarchy before any consumer reads it.
Step by step: from tag list to contextualised data
- Pick one unit and one use case, such as a pump fleet for anomaly detection. OSIsoft's advice is to "only model what you have understood" (Batke, OSIsoft, 2019).
- Export the tag list with descriptor, engineering unit, range and source address. Our guide to extracting historian data covers the access path.
- Export the CMMS hierarchy: functional locations, equipment, classes and installation history.
- Define a template per equipment class, with the measurement roles a model or report needs.
- Apply curated mappings and naming rules from the instrument index and tag conventions.
- Run similarity matching on the rest and queue the suggestions for review.
- Confirm against the P&ID, with a process or instrument engineer.
- Load the mapping into the asset model, the namespace or a governed table, with validity dates.
- Measure coverage. Cognite suggests checking "in-scope coverage, orphans, and confidence" before tuning models (Cognite).
How do you keep the mapping current?
Contextualisation decays. Pumps are swapped, instruments are added, units are revamped. Three habits help:
- Tie changes to change management. OSIsoft's governance guidance includes naming conventions for historian tags and asset objects and an approval process for changes (Batke, OSIsoft, 2019). Add "update the tag mapping" to the checklist for every modification.
- Keep confirmed links when you rerun matching. Cognite advises preserving "subject matter expert (SME)-confirmed links" on reruns (Cognite).
- Watch for deletions. The same OSIsoft presentation notes that importing an asset structure adds and modifies objects but does not remove those deleted on the master. Stale links survive unless someone removes them.
On ownership, see data governance for industrial companies.
Why does it matter for AI?
The Honeywell patent ties tag mapping directly to analytics: mapping tags to asset models reduces configuration effort and makes advanced analytics, such as predicting equipment failure, faster and easier to set up (US 11017008 B2). In practice:
- Models get the right inputs. A model for pump P-101 asks for its discharge pressure by role, not by a guessed tag name.
- Models scale across a fleet. A model built on a template can be deployed to other assets of that class, then checked on each.
- Assistants can answer plant questions. When an engineer asks about "the pump at P-101", an assistant or agent needs the mapping to find its tags, its work orders and its manuals; see root cause analysis with an LLM and maintenance history and connecting an AI agent to a CMMS.
What we have seen at KABAS AI
Our data science service covers data collection and preparation, predictive models and the MLOps pipelines that run them, and our information systems service covers integration of existing systems. DALILE, our on-premise assistant, combines documents, spreadsheets and databases in one query. We have not published a contextualisation project, so the points below are practical guidance, not reported results.
- Budget for mapping before modelling. The Honeywell patent calls it the single largest effort.
- Start from the CMMS functional locations, not from the historian tag list, so process and maintenance data share one key.
- Store how each link was made and who confirmed it. A model is only as trustworthy as the mapping that feeds it.
Sources
- Cognite, Contextualization (documentation)
- Cognite, Entity matching (documentation)
- Cognite, Match entities (documentation)
- Honeywell International, US 11017008 B2, Method and system for contextualizing process data, 2021
- OSIsoft, See the Light: Easy wins to improve your Asset Framework experience, PI World 2020 lab
- Batke, Best Practice: Asset Framework Deployment and Management, OSIsoft PI World, 2019
- SAP Learning, Explaining Technical Asset Structures
- ISA, ISA5.1 Instrumentation Symbols and Identification committee
- OPC 10030, OPC UA Common Object Model: ISA-95, ISA95EquipmentElementLevelEnum
- OPC 10000-8, OPC UA Part 8: Data Access, EUInformation
- ISO 14224:2016, Collection and exchange of reliability and maintenance data for equipment, scope (GSO standards store)
- IOGP JIP36, CFIHOS standard
- Pronk, Generative LSTM Models and Asset Hierarchy Creation in Industrial Facilities, arXiv 2403.06103, 2024
Frequently asked questions
Is data contextualisation the same as building a knowledge graph?
Not quite. Contextualisation is the work of attaching each time series to the equipment, hierarchy and events it belongs to, usually recorded in a mapping table or an asset model. A knowledge graph is one place to store and query those links together with documents, drawings and work orders. Most plants need the tag-to-equipment mapping first, whichever store they choose.
Should a tag be mapped to the functional location or to the physical equipment?
Map it to the functional location, the position in the plant, because the transmitter measures the process at that position whatever pump is installed there. Then use the CMMS installation history to know which physical equipment was in place at any date. This keeps the time series continuous when a pump is swapped and lets maintenance history follow the serial number.
Can machine learning or a language model do the mapping automatically?
They can propose matches, not decide them. Vendor tools such as Cognite's entity matching apply manual mappings and rules first and machine learning last, and their scores are not probabilities. A language model can suggest a match from a tag description in the same way. Treat every suggestion as a draft that an engineer confirms against the P&ID before models or reports depend on it.
Do we have to rename our historian tags?
No. Renaming tags breaks displays, reports and scripts that use the old names. Keep the source tag names and add context on top, either in an asset model such as AVEVA PI Asset Framework, in a mapping table or in a unified namespace topic structure. Apply a naming convention to new tags from now on.
Where should the tag-to-equipment mapping live?
In one governed place that other systems read from, not in each project's spreadsheet. Common choices are the historian's asset model, a small database maintained by the data team, or a knowledge graph. Whatever the choice, the CMMS stays the master for functional locations and equipment, and the mapping is rebuilt or checked whenever either side changes.



