Table of Contents
A knowledge graph for plant equipment is a database that stores each tag, piece of equipment, drawing, document and work order as a node, and the relationships between them as links. It lets engineers, and AI assistants, answer questions that cross systems, such as which procedures and past failures apply to the pump now installed at a given position.
Key takeaways
- A knowledge graph is "a graph of data intended to accumulate and convey knowledge of the real world", with nodes for entities and edges for relations (Hogan et al., 2021).
- In a plant, the hard part is identity: the same pump appears as a functional location, an equipment number, a P&ID tag and a phrase in a work order. Matching them is most of the work.
- Industry standards already define much of the vocabulary: ISO 14224 for equipment taxonomy and failure modes, CFIHOS for tag and equipment classes, DEXPI for P&IDs.
- Vector search does well when the answer resembles the question, but has no sense of the breadth of a dataset for global questions (Microsoft Research, 2024). A graph adds structure that search alone lacks.
- Build from structured sources first, then add links extracted from text, with engineers reviewing them.
What is a knowledge graph, in plant terms?
A survey in ACM Computing Surveys defines a knowledge graph as "a graph of data intended to accumulate and convey knowledge of the real world, whose nodes represent entities of interest and whose edges represent relations between these entities" (Hogan et al., 2021).
In a plant, the entities are things engineers already talk about: pump P-101, its motor, the P&ID sheet that shows them, the OEM manual, and the work orders raised on them. The relations are just as familiar: installed at, connected to, shown on, described by, reported on. Written as statements:
- Equipment 10045678 is installed at position P-101.
- P-101 is shown on P&ID sheet 101-PID-003, revision D.
- Equipment 10045678 is an instance of model XYZ-200, which is described by OEM manual 4711.
Each statement is one fact with a subject, a relation and an object. The W3C's RDF standard calls this a triple and defines an RDF graph simply as "a set of RDF triples" (W3C, RDF 1.1 Concepts).
Why is a CMMS hierarchy not enough?
A CMMS already structures equipment. In SAP, functional locations are "hierarchical structures with a systematic order" representing technical systems, while equipment is "an individual physical object that you would maintain as an autonomous unit", installed at a functional location (SAP Learning).
That distinction between the position (the tag) and the physical object (the equipment) is essential, because pumps get swapped. The manual and the failure history follow the physical object; the process function, the P&ID and the operating limits stay with the position.
The CMMS tree, however, has one parent per object. It does not hold the process connections from the P&ID, the link from a model to its manual, or the equipment named only in work order text. Today those links are joined by an experienced engineer's memory. A graph pays off on questions that cross systems:
- Which procedures and manuals apply to the motor installed at P-101 today, not the one removed last year?
- Which other pumps of the same model, on any site, have had seal failures?
- If valve XV-2003 is isolated, which equipment downstream is affected, and which open work orders sit on it?
What goes into a plant equipment graph?
| Node type | Typical source | Example | Main relations |
|---|---|---|---|
| Functional location or tag | CMMS, tag register | P-101 | part of unit, shown on P&ID, connected to line |
| Equipment (physical item) | CMMS equipment master | Serial 10045678 | installed at tag, instance of model |
| Model or equipment class | CMMS, OEM data, CFIHOS or ISO 14224 classes | Centrifugal pump model XYZ-200 | described by manual, member of class |
| Drawing | Document management, DEXPI files | P&ID 101-PID-003 rev D | shows tags and lines |
| Document | Document management | OEM manual, procedure, datasheet | describes model or tag, superseded by |
| Work order and notification | CMMS | WO 2026-1182 | reported on tag or equipment, failure mode |
| Spare part | CMMS bill of materials, stores | Mechanical seal kit | fits model |
Work order text deserves its own step. Stewart and colleagues describe maintenance work order text as "unstructured, terse and jargon-rich". Their MWO2KG tool combines what deep learning extracts from that text with structured fields such as dates, functional location and asset make and model, to build a graph reliability engineers can query (Stewart et al., 2024).
Which industry standards should you reuse?
You do not need to invent the vocabulary.
| Standard | What it gives you | Use in the graph |
|---|---|---|
| ISO 14224 | Reliability and maintenance data for petroleum, petrochemical and natural gas equipment, including equipment taxonomy and failure modes usable as a "reliability thesaurus" | Equipment classes, failure modes and failure causes |
| ISO 15926 | A representation of information on the engineering, construction and operation of process plants, including oil and gas production facilities | A reference for life-cycle data integration and shared reference data |
| CFIHOS | IOGP's handover specification, with tag and equipment class hierarchies, their properties and a reference data library; version 2.0 is aligned with ISO 14224 | Class names and properties for tags and equipment, and what contractors should hand over |
| DEXPI | A data exchange standard for the process industry; DEXPI 2.0 covers BFDs, PFDs and P&IDs in DEXPI XML | Process connections between tags, lines and instruments |
CFIHOS documents are free to download (JIP36). If new projects require CFIHOS handover from contractors, the data arrives close to graph shape. Old P&IDs held as scans need digitising first; see our guide to OCR for scanned technical documents and P&IDs.
RDF or property graph: which model?
| RDF triple store | Property graph database | |
|---|---|---|
| Data model | Triples of subject, predicate and object, identified by IRIs (W3C RDF 1.1) | Nodes and relationships, both carrying key-value properties (Neo4j documentation) |
| Query language | SPARQL, a W3C Recommendation since 2013 (W3C) | Cypher, and GQL, published as ISO/IEC 39075 in April 2024 (IEC webstore) |
| Strengths | Globally unique identifiers, suited to exchange between organisations | Properties on relationships (such as installation date), close to how application developers model data |
| Typical fit | Sharing reference data and classes across companies and contractors | An internal graph behind applications and assistants |
Either can hold a plant equipment graph. Whatever the choice, keep the CMMS and the document management system as the systems of record, and treat the graph as a connected view rebuilt from them.
How does a knowledge graph help an LLM assistant?
Microsoft researchers note that vector RAG "excels for local queries where the answer resembles the query", but has no sense of the breadth of the dataset to consider for global questions (Microsoft Research, 2024). Our guide to RAG for maintenance manuals covers retrieval in detail.
A plant equipment graph helps in three ways:
- It resolves identity. "The pump at P-101" becomes an equipment number, a model and a list of current documents before any text is searched.
- It filters retrieval. The assistant searches only the manuals linked to that model and the procedures that are current for that tag.
- It answers structural questions that are not written in any single document, such as which equipment is downstream of a valve.
Microsoft's GraphRAG takes another route: a language model extracts an entity graph from the documents and pre-generates summaries of groups of closely related entities. On global sensemaking questions over datasets of about one million tokens, it gave substantial improvements in comprehensiveness and diversity of answers over a conventional RAG baseline (Edge et al., 2024). Its project page warns that indexing "can be an expensive operation" and that out-of-the-box results may need prompt tuning (GraphRAG on GitHub). Microsoft's later LazyGraphRAG variant cuts indexing cost to 0.1% of full GraphRAG (Microsoft Research, 2024).
For plant equipment, most of the graph already exists in structured systems, so text extraction is a complement, not the foundation. An AI agent can query the graph as one of its tools, alongside the CMMS tools described in connecting an AI agent to a CMMS and the database queries discussed in text-to-SQL on production databases.
How do you build one, step by step?
- Write the questions first. Collect 20 to 30 real questions from engineers that today need several systems to answer.
- Pick one unit or one equipment family, not the whole site.
- Load the CMMS hierarchy. Functional locations, equipment, installation history and models, keeping CMMS identifiers as keys.
- Define a small model with only the node and relationship types your questions need, named after ISO 14224 or CFIHOS classes where they fit.
- Match identifiers. Write normalisation rules for tag formats (P-101, P101, 10-P-101) and record each match with its source.
- Add documents and drawings. Link manuals to models, procedures to tags and P&ID sheets to the tags they show, with revision status.
- Add work order text last, and have an engineer review extracted links before they are trusted.
- Test with the questions against what an experienced engineer would answer, and rebuild the graph from the source systems on a schedule.
What usually goes wrong?
- Identifiers do not match. The same tag is written differently across systems, and in Algerian plants free text may mix French, Arabic and English; see documents that mix Arabic, French and English.
- The graph goes stale. If it is not refreshed when equipment is swapped or a document is revised, it gives confident wrong answers.
- The model is too ambitious. Months go into an ontology before a single question is answered.
- Extracted links are trusted blindly. A language model will sometimes link the wrong pump to a work order. Mark extracted links until reviewed.
- Access rights are lost. A graph that joins documents must respect the permissions of each source document.
Poor source data is the root of most of these; see common industrial data quality problems. A systematic review of knowledge graphs in manufacturing found that "empirical research and industrial application are still missing to a large extent" (Buchgeher et al., 2021), so expect to learn on your own data.
What we have seen at KABAS AI
DALILE, our sovereign multi-agent assistant, runs on the client's own servers, searches across documents with exact sources, combines documents, spreadsheets and databases in one query, and connects to ERP and document management systems. These are the same sources a plant equipment graph draws on. We have not published a knowledge graph project, so what follows is guidance, not a reported result.
The problem is familiar. At ADFC (Sonatrach Group), decades of formulation experience were "dispersed in various formats and locations", and ADFC is exploring integration with DALILE for knowledge discovery and technical documentation. Equipment knowledge in most plants is scattered the same way. When scoping a graph, we recommend:
- Start from questions, not from a data model, and keep the first scope to one unit.
- Use the CMMS as the backbone and reuse ISO 14224 or CFIHOS classes rather than inventing names.
- Keep everything on-premise, since the graph concentrates equipment, drawings and maintenance history in one place.
- Plan how agents will use it; see multi-agent systems for industrial workflows.
Our agentic AI team builds agents and workflows on DALILE, and our data science team prepares and structures industrial data for this kind of work.
Sources
- Hogan et al., Knowledge Graphs, ACM Computing Surveys 54(4), 2021 (arXiv 2003.02320)
- W3C, RDF 1.1 Concepts and Abstract Syntax (Recommendation, 2014)
- W3C, SPARQL 1.1 Query Language (Recommendation, 2013)
- IEC webstore, ISO/IEC 39075:2024, Information technology, Database languages, GQL
- Neo4j documentation, What is a graph database
- SAP Learning, Explaining Technical Asset Structures
- ISO 14224:2016, Collection and exchange of reliability and maintenance data for equipment, scope (GSO standards store)
- ISO 15926-1:2004, Integration of life-cycle data for process plants including oil and gas production facilities, Part 1 (GSO standards store)
- IOGP JIP36, CFIHOS standard
- DEXPI, Specifications
- Stewart, Hodkiewicz, Liu and French, MWO2KG and Echidna: Constructing and exploring knowledge graphs from maintenance data, Journal of Risk and Reliability 238(5), 2024
- Edge et al., From Local to Global: A Graph RAG Approach to Query-Focused Summarization (arXiv 2404.16130, 2024)
- Microsoft, GraphRAG on GitHub
- Microsoft Research, LazyGraphRAG: Setting a new standard for quality and cost (November 2024)
- Buchgeher et al., Knowledge Graphs in Manufacturing and Production: A Systematic Literature Review, IEEE Access, 2021 (arXiv 2012.09049)
Frequently asked questions
What is the difference between a knowledge graph and the asset hierarchy in our CMMS?
The CMMS hierarchy is a tree: each functional location or piece of equipment has one parent. A knowledge graph also holds the links a tree cannot, such as which line feeds which vessel, which documents describe a model of pump, and which work orders mention a tag. The CMMS hierarchy is usually the first thing loaded into the graph.
Do we need a full ontology before we start?
No. Start with a small set of node and relationship types that answer real questions, such as tag, equipment, document, work order and their links. Reuse names and classes from standards such as ISO 14224 or CFIHOS where they fit, and extend the model when a new question needs it.
Is GraphRAG the same thing as a knowledge graph?
No. GraphRAG is a method published by Microsoft researchers in which a language model extracts entities and relationships from text to build a graph index for retrieval. A plant equipment graph is usually built mostly from structured sources such as the CMMS, tag lists and P&ID data, with text extraction used for the parts that exist only in documents.
Can a language model build the graph automatically from our documents?
It can propose nodes and links, but it makes mistakes, and Microsoft's GraphRAG documentation warns that indexing can be expensive and that prompts need tuning. Load what already exists in structured systems first, and have engineers review extracted links before they are used for maintenance or safety decisions.
Should we use an RDF triple store or a property graph database?
Both work. RDF and SPARQL are W3C standards built for shared identifiers and exchange between organisations, which suits reference data such as ISO 15926 or CFIHOS classes. Property graph databases with Cypher or the ISO GQL language are often easier for application developers. Choose based on the questions, the team's skills and the tools you must exchange data with.



