Blog

2026.11.09

AI Tag Verification for Factory P&IDs: Link Drawings to the Asset Register

AI Tag Verification for Factory P&IDs: Link Drawings to the Asset Register

Scanning a piping and instrumentation diagram (P&ID) does not tell a maintenance engineer whether pump P-101 on the drawing is the same asset as P101 in the CMMS. Nor does it prove that a valve absent from a revised drawing has been removed from the plant. A searchable PDF solves only the first part of the problem. For a factory planning AI-assisted P&ID tag verification, the real project is to extract evidence, compare it with an asset register, route uncertain cases to an engineer, and control the resulting changes.

This guide explains the practical sequence for a Thai manufacturing site with legacy drawings: define the authoritative revision, extract text, symbols and connections separately, normalize tags without losing the originals, compare them with asset records, then specify acceptance criteria for the request for proposal (RFP), factory acceptance test (FAT) and site acceptance test (SAT). AI output is a candidate for engineering review, not an authorization to change a safety-relevant asset record.

Why digitizing a P&ID is more than OCR

A P&ID depicts equipment, piping, valves, instruments and control relationships. Old scans force engineers to transcribe tags manually; even CAD-derived drawings can disagree with the plant after modifications. “The text was read correctly” and “the record describes the installed equipment” are separate findings. The ISA 5.1 overview describes a uniform identification and symbol approach, but an existing plant may also have local conventions and legacy aliases. Do not silently rewrite its established tags merely to fit a model’s preferred format.

AWS’s published P&ID digitization guidance separates text detection, symbol detection and line detection, then combines their positions and relationships into structured output. It is an architecture example, not evidence that every low-quality drawing can have its topology verified automatically. OCR can identify a string while attaching it to the wrong valve or the wrong line. The review workflow must expose those distinctions.

Three layers of verification

Layer one is the literal tag and its page coordinates. Layer two is the object to which the tag belongs: pump, valve, instrument or pipe line. Layer three is the corresponding asset-register or CMMS record. Measure each layer separately. Reading XV-201 as XV-207 is an OCR error; attaching a correctly read XV-201 to the adjacent valve is an association error; selecting a retired CMMS row for XV-201 is a master-data error. A single headline “accuracy” score conceals the source of the problem.

AI Tag Verification for Factory P&IDs: Link Drawings to the Asset Register - figure 1

Classify results as exact match, naming variant, drawing only, register only, multiple candidates, unreadable, or topology review required. “Drawing only” does not establish that the plant contains an unregistered device; the register extract may be incomplete. “Register only” does not establish removal; the asset may appear on another sheet or a revision may be missing. Use these labels to start an investigation, never to announce its conclusion.

Establish the source documents first

Build a drawing inventory with drawing ID, revision, issue date, approval state, area, format, scan quality and original location. Keep redlines, drafts and approved as-built drawings distinguishable. Otherwise a PoC can report hundreds of “AI mismatches” that are simply differences between old and new revisions. Freeze a CMMS or asset-register snapshot at the same time and record its extract date, columns, filters and known aliases. If a plant has separate equipment, instrument and line lists, document which list governs each field.

A representative sample must include faint scans, rotated text, folded pages, hand annotations and crowded diagrams, as well as clean drawings. More pixels cannot restore a tag that is missing from the source paper. Return such a case as unreadable. Include drawing type and quality in test reports so a strong result on new CAD exports does not hide poor results on the old backlog.

Split extraction into auditable stages

First check whether the PDF already contains embedded text. Use OCR for raster pages. Correct orientation, identify title blocks, revision tables, legends and the drawing area, and avoid treating legend examples as real equipment. Store text candidates, symbol candidates and line or signal candidates separately. Join them using position and connectivity while retaining coordinates and a link to the source crop. A reviewer needs to see the evidence behind every candidate.

Common character confusions include O/0, I/1, missing suffixes and different dash characters. A tag rule may narrow candidates but must not discard valid local conventions. The interface should display the image crop next to the interpreted string. For symbols and lines, distinguish a detected line from an approved engineering conclusion about connectivity; crossing lines, off-page connectors and hand-drawn modifications need special attention.

If the plant has structured CAD data, inspect it before commissioning image recognition for every sheet. Autodesk’s Plant 3D tag documentation shows tag fields accessible through a Data Manager. The cost of extracting those records differs from reconstructing tags in a raster scan, although both still require a check against actual equipment and the register.

Normalize tags without losing information

Normalize harmless differences such as space, case and dash shape for candidate matching, but retain the original string. P-101A and P-101-A may be variants at one site; P-101A and P-101B may represent two parallel machines. Removing a plant-area prefix may merge distinct assets. Document every normalization rule, its scope, exceptions, owner and version. Score candidate matches using area, system, object type, adjacent tags and a verified alias table as well as text similarity.

Never force a single match from two plausible candidates. If a duplicate tag exists in the CMMS, fix or adjudicate the master records; a higher AI confidence score cannot establish the correct asset. Record old and new numbers as aliases with validity dates where possible. Keep source text, inferred candidate and approved value in separate fields. This also prevents an AI guess from being mistaken later for something printed on the drawing.

AI Tag Verification for Factory P&IDs: Link Drawings to the Asset Register - figure 2

Design the human review and priority queue

Sort exceptions by plant risk as well as model confidence. Critical loops or high-consequence equipment may require engineering review even when text is clear. Display the source crop, nearby piping, detected symbol, candidate asset row, normalization rule and revision in one view. Let reviewers choose match, different asset, update drawing, update register, field check or hold. Save the reviewer, timestamp and reason; decide whether the next action is a CMMS update or a formal change request.

The NIST AI RMF Core calls for human oversight processes to be defined, assessed and documented. Here that translates into explicit decision rights and an audit trail. A disputed case should become an investigation ticket with owner and deadline, not disappear into a generic unprocessed queue. Corrections made by reviewers should be quality checked before they become training examples.

Drawings, register and field equipment may all disagree

When the drawing and register disagree, neither automatically wins. Compare the approved revision, modification handover records, maintenance history and the installed asset. A nameplate can provide another piece of evidence, as discussed in our equipment nameplate AI-OCR guide, but it cannot by itself establish a piping connection. Field verification must follow the site’s access and safety procedures, and unresolved records should remain explicitly marked “not field verified.”

Version-to-version drawing comparison is a different task, covered in our drawing revision AI comparison guide. A changed symbol on two sheets does not show whether the physical modification was installed or whether the CMMS was updated. This article focuses on linking a selected, approved drawing revision to current asset records.

Specify the output model and ownership

Retain drawing ID, revision, page, source tag, normalized tag, object type, coordinates, connection candidates, register ID, match state, review decision, file hash and processing version. A flat CSV with tags alone loses the evidence required to investigate a later discrepancy. Specify the CMMS import format early and identify which system owns each field. For example, the approved P&ID may govern documented connectivity, the CMMS may govern maintenance status, and a modification record may govern the reason for a change.

DEXPI’s specification page describes vendor-neutral models for process information and now identifies version 2.0.1 as the corrective release replacing 2.0. DEXPI may be an appropriate exchange option, but “export to DEXPI” is not proof that connections were reconstructed correctly. State required attributes, missing-value behavior and an import test in the RFP. ISO 15926-1 addresses life-cycle information integration for process plants; in day-to-day deployment, stable identifiers and change histories are the practical foundation for that continuity.

What to put in the RFP

Provide representative input files, drawing versions, languages, tag conventions, security constraints and the register schema. Require not just a tag list but source locations, associated symbols, connection candidates, match states, review history and output to the CMMS. Ask bidders how they create ground truth, score each stage, flag low-confidence or ambiguous cases and disclose failure examples. Do not copy a generic vendor accuracy percentage into the contract without a site-specific denominator.

RequirementSpecifyEvidence at acceptance
InputsRevisions, formats, quality, areasInput inventory and exclusion reasons
ExtractionOriginal tag, coordinates, object and line candidatesReview screen linked to source
MatchingAsset candidates, state and reasoningComparison with approved truth set
ReviewRights, decisions and historyAudit log and change record
IntegrationCSV/API or exchange modelCMMS import result
SecurityStorage, access, retention and deletionConfiguration and operating procedure

The price comparison should include inventory preparation, ground-truth labeling, register cleanup, reviewer time, system integration, FAT/SAT and ongoing correction. Per-page OCR pricing alone does not represent the completed business process. If the backlog is small and good CAD data exists, structured-data extraction plus a reviewable difference list may be enough. If updates are frequent and numerous, exception routing and repeatable processing deserve more investment.

Separate PoC, FAT and SAT

The proof of concept tests feasibility on a stratified set of real drawings: old and new, dense and sparse, high and low scan quality, with and without structured CAD data. Engineers and register owners should agree on the truth set; mark disputed source cases separately. In FAT, test text errors, incorrect tag-to-symbol links, wrong register selection, ambiguous holds, permissions, audit logs, reprocessing and recovery. Include high-confidence wrong answers. Re-run the same cases after a fix and check that previously correct cases remain correct.

SAT tests the actual plant network, user roles, drawing repository and CMMS. Include resubmitted files, a new revision, simultaneous register updates, outages and reviewer absence. Measure whether a real reviewer can resolve an exception, not simply whether the algorithm returns a result. A method that passes in a vendor environment can still fail when the site’s permissions or document names differ.

AI Tag Verification for Factory P&IDs: Link Drawings to the Asset Register - figure 3

Measure completed work and control revisions

Track missed tags, false tags, correct object associations, correct asset rows, wrong automatic approvals, human review minutes and unresolved differences. State each denominator. Sample some automatically approved records for independent review; reviewing only flagged cases would miss confident mistakes. Compare the total pre- and post-deployment time through CMMS update, including investigation and rework, rather than reporting model runtime alone.

Treat each new drawing issue as a new revision rather than overwriting an old result. Re-run matching, prioritize changed candidates, and retain the previous decisions. An unchanged tag may still have different connections. Link every change to drawing revision, asset ID, modification project, approver and update time. Define rollback and avoid uncontrolled two-way synchronization among drawing storage, the register and CMMS.

A practical sequence for a Thai plant

Start with one system and one equipment group. Bring design, instrumentation, maintenance, operations and IT/security together to agree on the same drawing revision and register snapshot. Japanese, Thai and English names may differ; do not use a translated display name as the sole identifier. Inventory documents and naming rules, run a representative PoC, build the review and exception process, test FAT and SAT, then expand in stages. The expansion gate is whether approved, traceable matches can support maintenance work, not how many scans the AI processed.

Frequently asked questions

Is a searchable PDF enough?

It helps locate tags, but does not establish which object each tag identifies or which CMMS row corresponds to it. Verification needs the source location, object candidate, register candidate and decision history.

Can AI automatically verify all piping topology?

Models can propose text, symbol and line relationships. Crossings, off-page connectors, annotations and damaged scans can be ambiguous. Require engineering approval where a connection could affect maintenance or safety; choose any automatic scope from your own test evidence.

Do we need OCR if CAD data exists?

Extract available structured tag data first. OCR is useful for raster scans or unstructured drawings. Matching the resulting records to the installed plant remains necessary in either case.

Which record should be corrected when the drawing and register differ?

Check the approved drawing, modification history, field evidence and register history, then have the responsible owners decide. Keep the difference open until that decision is documented.

What should we prepare for a quotation?

Representative drawings including poor scans, a revision inventory, tag rules, register field definitions, aliases, CMMS import constraints and information-security conditions. State who can review uncertain matches and what acceptance evidence you expect.

Conclusion

Successful AI-assisted P&ID tag verification is an evidence and change-control workflow. Confirm the source revision, extract text and objects separately, match using reversible rules, send uncertain and important cases to engineers, and verify the outcome in PoC, FAT and SAT. Keep the link from each approved asset record back to the drawing and reviewer.

TOMAS TECH can help Thai plants scope representative drawings, define PoC acceptance criteria, design exception review and plan integration with an asset register or CMMS. Share sample P&ID sheets and an anonymized register schema to discuss what can be reused and which differences need engineering confirmation. Contact us.

Sources