WorkflowEngineering Data, Tags & Documents

Oil and Gas Document Extraction

Oil and gas document extraction that turns a messy handover package into one table, each cell tied to where it was read on the page, with people checking only where reading was hard.

Follow one handover package, from ask to sign-off ↓
The problem

A handover arrives as boxes of scans, and no number has a trail to its page

A package lands from a former operator, a vendor or a contractor's drive: well reports, datasheets, test certificates and isometrics, often one scan holding dozens of documents with no index. Files that will not open get chased by email weeks later.

An analyst pages through it and keys values into a spreadsheet, usually with no record of which page each came from. Stamps, faded ink and pen changes are judged on the spot, tables that break across pages get retyped, and when someone later doubts a figure, the search starts over.

Built forEngineering data analystsDiscipline leadsDocument controllersInternal auditors

The Problem Today

Old scans, pen marks and stamps defeat plain reading, so people retype numbers with no trail to the page.

70%
right on image-based well records, in one published study published
110 h
to key one table set by hand, at 45 seconds a value our estimate

Read and retyped by hand today

  • Datasheets
  • Well files
  • Drawings
  • Test certs
  • A pen changes the numberThe print says one rating, a pen beside it another
  • One scan, many documentsA single box holds dozens of documents and no index

The Solution

  1. Say what you needPlain words become a proposed table; you agree before it is read
  2. Each value pinned to the pagePrint, stamps, pen and faded scans read into one table, pinned to the page
  3. Shows where it struggledHard spots lit up, tables joined across pages, two languages compared
  4. People check only the doubtsA person decides with a reason; a second person checks and signs

Built-in controls

  • Each decision waits for a second person's check
  • Only signed rows leave, each with its page link
  • A value not on the page stays empty

What You Get

Engineers get a checked table, and the proof for any number is one click away.

28 h
analyst hours in place of 110: settling and checking hard values our estimate
  • Silent misses countedA second person samples values the agents were sure of
  • Asked once, reusedA saved table runs on the next package
  • Rows leave with their pageEach sent row carries a link back to where it was read
One package, one table

Oil and gas document extraction on a legacy well file handover of 9,640 pages

An operator has taken over an offshore field and its records arrive as 221 files in boxes and a file share. Nobody indexed them. The analyst needs the pressure limits for every well before the integrity team can plan.

demoField, former operator, wells and counts: all of it is sample data.

  1. 01The package

    214 files opened, 1,386 documents inside, and 2,528 pages hard to read

    Engineering data analyst

    The opener reports what came in and lists the 7 files it could not open, with why. Its map lays the package out by document type, one square for about ten pages, colored by the cause: faded, handwriting, stamp, a table across pages or two languages. People start with those pages, and a discipline lead's spot check tests the other 7,112.

  2. 02The package

    Hard spots light up the faded test pressure under a RECEIVED stamp

    Page reader

    Opening an end of well report and turning on hard spots greys the page and lights what was difficult: a faded number in the production casing row and the receipt stamp across it. The panel beside it lists every value read on the page with how sure the reader is. She sees where the trouble is before she asks for anything.

  3. 03The package

    A 1998 wellhead sheet: 3,000 struck through, 5,000 written in pen

    Page reader

    On a vendor datasheet from the old operator, the casing head rating is printed one way and changed by hand, with initials. The shell test is raised the same way. A plain read would take the print, or the pen, and the spreadsheet would never show which. The reader marks the page handwriting and keeps both readings.

  4. 04The table

    1,264 rows across 62 wells, and the pen-raised rating is held

    Agents, then the analyst

    She asked in plain words for pressure limits for every well and agreed six proposed columns before anything was read. The table fills, 8,848 values with 412 for a person. Clicking the casing head cell brings up a crop of the pen figure. A pen mark alone does not prove a rating, and no nameplate photo, as-built record or change record is in this package, so the value waits.

  5. 05The table

    Read 3,800 at 62% sure, and the daily report from the same day says 3,500

    Page reader, then the analyst

    Under the stamp sits the production casing test cell. Only a 3, then 00, can be made out, and the second digit could be a 3, a 5 or an 8. A cross-check found the daily drilling report written that day, which says 3,500, and cites it. The analyst opens it, with the casing test line marked as the second document.

  6. 06The table

    She takes 3,500 from the second document, and it waits for the lead

    Engineering data analyst

    She uses 3,500 with the reason confirmed on a second document. The cell now reads 3,500, was 3,800, decided by her, waiting for the discipline lead to check. She cannot check her own decision. The cross-checker offers the same move on 96 other flagged values where a second document says the same number, each to be looked through first.

  7. 07Sign-off

    He checks her decision first, and only then can rows be signed

    Discipline lead

    The sign-off page holds signing until every decision is agreed or returned. He opens the corrected test pressure with its page crop and agrees; it reads Checked. Now 870 rows are ready, every value in them settled and checked, and 394 stay held. The person who decided a value cannot check it, and a row leaves only when signed.

Where 2,528 of the 9,640 pages were hard to read, by cause

The package's hard pages, as the hard pages screen counts them.

Faded or poor scan1,042
Handwriting on the page517
Stamp over text386
Table across pages304
English and Arabic on one page209
Skewed or turned70
Hard pages in the package2,528

demoSample data, from the working solution.

Five values in one handover, and what each one needed

The values that tell the package's story, from the plan's own sources.

ValueWhat the page showsWhat happens
Casing head rating, 1998 datasheet3,000 printed, struck through; 5,000 in pen, initialedHeld: no nameplate, as-built or change record in the package
Production casing test, end of well reportFaded under a stamp, read 3,800 at 62%Corrected to 3,500 from the daily report; checked by the lead
Surface casing depth, second well2,531 under an approval stamp, 71%A faded well sketch also shows 2,531
Test pressure, certificate MT-22094English 7,500, Arabic 7,000Nothing prefilled; the analyst asks the lab
Certificate MT-22071Both languages agree on all 11 linesNeeded no person

demoSample data, from the working solution.

Around the table

Split scans, tables across pages and the rows that reach your systems

What document control, the lead and an auditor open around one table.

One scan, 43 documents612 pages from one box, split into documents. 37 documents are sure; 6 boundaries wait for document control, and values from those documents stay provisional until then.
One table, three pagesA vessel's nozzle schedule breaks across three sheets. Repeated headings are dropped, the cut row is joined, and a row added by pen is kept and marked.
PackagesSix open packages, 26,112 pages, from handovers, vendors and scanned boxes, each with its stage, owner and clock, and the files held with why.
Saved tablesA table asked for once runs on the next package. Each version records what changed and who approved it.
Send to your systemsEach column mapped to the well data store, the maintenance system or the document system. Nothing can be sent until rows are signed.

demoSample data, from the working solution.

The analyst decides a value; someone else checks it before a row is signed

Four roles around one package, and none of them both decides and checks.

Engineering data analyst
Sees
The table, the hard spots and each flagged value with its page
Signs
Decides each doubtful value with a reason
Discipline lead
Sees
Each decision with the crop of its page
Signs
Agrees or returns each decision, then signs the rows that are ready
Document control lead
Sees
Unsure boundaries between documents, held files and the send
Signs
Settles boundaries and sends signed rows with a receipt
Internal auditor
Sees
The trail from any value to its page
Signs
No signature: reads and exports the record
The agents

Eight agents open, split, read and cross-check, and none decides a value

Each works on the package and hands its doubts to a person.

Package opener

Opens every file, records its fingerprint and sets aside any file it cannot open.

  • Never drops a file without saying so
  • Leaves the source files untouched
Page sorter

Splits files into documents and sorts them by type.

  • Never merges two documents when unsure
Table planner

Turns a plain-words request into a proposed table.

  • Never starts filling before a person agrees
  • Never invents a column it cannot source
Page reader

Reads print, stamps, pen and faded scans, and ties each value to its place with a confidence.

  • Never guesses a value not on the page
  • Never hides a low-confidence value
Table stitcher

Rejoins tables that run across pages and removes repeated headings.

  • Never joins rows without showing the join
Two-language reader

Reads the English and the Arabic on a page, each in its own direction, and compares them.

  • Never picks one language over the other
Cross-checker

Looks for a second document that confirms or contradicts a value.

  • Never overrules a person
  • Never changes a value
System handoff

Maps signed rows to your systems, runs a dry run and sends with a receipt.

  • Never sends an unsigned row

Questions data teams ask about oil and gas document extraction

What is oil and gas document extraction?

Oil and gas document extraction pulls fields and tables out of well reports, vendor datasheets, test certificates and drawings into structured data. Here a person asks for a table in plain words, agrees the columns, and gets one table where each cell points to where it was read on the page, with the hard readings sent to a person.

Does it work on legacy well files and old scans?

Yes. Faded scans, stamps over numbers, handwriting and skewed pages are read, and each is marked on the page so people can see where reading was hard. A value the reader cannot see stays empty rather than guessed.

What happens when a pen has changed a printed number?

Both readings are shown. The cross-checker looks for a nameplate, an as-built record or a change record, and when none is in the package the value is held until a person says what confirms it.

Can it pull a table that breaks across several scanned pages?

The table stitcher rejoins it, drops repeated headings and shows each join, including rows added by hand. A person agrees the join before the rows count.

What about certificates in two languages?

Each language is read in its own direction and compared line by line. When they disagree, nothing is prefilled, and the analyst can ask the lab before deciding.

Who checks the values before they reach our systems?

A discipline lead agrees or returns each decision the analyst made, then signs the rows that are ready. Only signed rows leave, each with a link back to its page.

Can we reuse a table on the next package?

Yes. A saved table runs on the next package, and a change to it takes a second person's approval and is kept as a new version.

Send us one box
from your last handover.

Tell us in plain words what table you need from it, and we’ll show the columns first, then each cell on its page.