WorkflowShutdown, Turnaround & Outage

Turnaround Records Classification & Deduplication

Turnaround document classification that splits, sorts and deduplicates decades of records, and keeps the job history.

The problem

Turnaround document classification fails when one scan holds many documents

Job packs, schedule files, material take-offs, cost records, QA/QC documents, drawings and lessons learned from past turnarounds share the same folders and boxes. A scanning contractor often makes one PDF per box or job, so sorting by file gets it wrong.

Duplicates are usually found by file name or a checksum. Rescans get missed, and a recurring job pack looks like a copy. It carries the same pack number and nearly the same words, turnaround after turnaround, and it is the wear record of the equipment.

Built forTurnaround records leadsTurnaround engineersDocument controllersInternal auditors

The Problem Today

Decades of turnaround records sit in mixed files, and plain checks delete the job history planners need.

14–42K: hours to sort a sample estate by hand, at 1 to 3 minutes each our estimate
Every: inspection and test on process equipment must be documented published

Sorted by hand today

Job packs
Drawings
Schedules
Cost files

One file, many documentsA scanned box hides job packs, tests and drawings
Holes nobody can seeA list never shows which records are missing

The Solution

Split by pageEvery file read page by page; one scan becomes the documents inside it
Sorted and citedSeven types, with job pack number, tag, date and cost center per page
Same job or duplicate?Exact and near copies set aside; recurring job packs kept as history
Doubts go to engineersPossibly obsolete, incomplete or unsure, routed by unit and reason

Built-in controls

Second person countersigns each decision
Agents never delete; retired under your rules
Agents never rule a document obsolete

What You Get

Planners get a clean record to build the next turnaround on, and engineers judge only what needs them.

Counts that add upThe whole estate reconciled at every stage
Readings across turnaroundsTube wear on one exchanger, turnaround after turnaround
Missing records foundA map shows which turnaround is short, and why

The run, stage by stage

Turnaround document classification, stage by stage, over 21 days

Six turnarounds of records reach a sample refinery’s team in batches. The agents take 21 days over 8,642,190 pages. People then decide what those agents would not.

demoThe refinery, its people and its counts are sample data.

01Intake
614,188 files, and a count for those nobody can open
Intake agent and document control
Five sources feed the estate, all read only: system exports, closeout shares, scanned boxes, contractor drives and a planning drive. 1,284 files could not be opened: 612 password protected, 418 damaged, 254 in old formats. Each kind has an action beside it, and the sum on the screen adds up.
02Split
A box that looks like one file holds 12 documents
Page splitter and document control
A scanning contractor often makes one PDF per box. The page splitter reads this one page by page and shows it as a strip: 186 pages, 12 documents, 6 of them job packs, each with its range. On the right, a drawing with no pack number might belong to the job pack before it. Below 85%, a person rules and another countersigns.
03Sort
Seven document types, the hardest calls first
Type sorter and turnaround engineer
Each document lands in one of seven types with a confidence and cited reasons. Below the 92% line it goes to an engineer. The hardest calls sit on the right: is this a material take-off or a cost record? Beside each type sits its sample result: 1,509 of 1,537 type calls correct.
04Deduplicate
380,658 proposed as copies, and none retired until a person signs its rule
Duplicate finder and records lead
Exact copies are found by fingerprint and near copies by text and page image. That is 45% of the estate. The sieve now reads 466,654 unique, and the map by turnaround year shows the shape of what remains. Amber squares mark turnarounds that look short on a type, and the lead can ask document control to find the missing boxes.
05Near copies
Three copies of one hydrotest record, and the one with a note wins
Duplicate finder and records lead
Copy 1 is clean, copy 2 is a skewed rescan, and copy 3 carries an inspector’s handwritten note about the retest. The kept-copy rule keeps the one with the note, even though it is less tidy than the first. The others point to it. Near duplicates are signed off rule by rule, with a second person countersigning.
06Same job
A plain check calls this pair of job packs a duplicate
Same-job checker and records lead
The packs for one exchanger match 96.4% with the same pack number. The same-job checker reads further: two turnarounds, two work orders, separate QA records, thinnest wall 0.083 in then 0.072 in, plugged tubes 12 then 19. It calls them the same job at a later turnaround, and each reason links to its place on the page.
07Same job
The lead agrees, and nothing changes yet
Records lead
The decision shows amber with the reason typed out: different turnarounds, work orders and readings. A countersign button sits beside it for a different person, and the lead’s own attempt is blocked. A disagreement takes a reason code and waits for a second person in the same way.
08Read
Pack number, tag, work order and date, each lit on the page
Field reader and turnaround engineer
Opening one job pack puts the page beside its type, its fields and its checks. Clicking a value lights its place on the page. A blank stays blank. A wrong value can be fixed, and the old value stands until a second person countersigns. The tag check confirms this exchanger is current in the equipment register.
09Flag
4.2% of the estate goes to engineers, in 42 packets
Review router, then engineers
The bar across the top splits 35,447 documents into possibly obsolete, incomplete and low confidence. They are grouped by unit and discipline in seven two-week waves. At 5 minutes each that is 2,954 hours, or 14 weeks for 22 engineers at 10 hours a week. The lead proposes the release, and document control countersigns.

What a plain duplicate check would have thrown away

One exchanger’s job packs, across the five turnarounds in its chain. Each looks like a copy of the last, and together they are its wear record.

Turnaround in the chain	Thinnest tube wall	Plugged tubes
First	0.098 in	3
Second	0.091 in	6
Third	0.083 in	12
Fourth	0.072 in	19
Fifth	0.064 in	27

demoSample data, from the working solution.

Around the sieve

Job pack fields, the holes, the program and the audit trail

What the records team and auditors open between stages. Select a screen, or open it large.

Fields read from job packs with fill rates and blanks, an ask box for a new field, and old equipment tags mapped to current ones — **Fields by document type**Fill rates and blanks by field, with old equipment tags mapped to current ones, and those with no match flagged.

Map of turnaround records going to engineers by type and year, 35,447 in all, with notes on holes in cost records and lessons learned — **The map, to engineers**What goes to engineers, by type and turnaround year, with the agents’ notes on where the records look short.

Program progress for a turnaround records migration: phases, queues and clocks, one overdue item, and a forecast marked as an estimate — **The road to the new system**Phases, queues and clocks against the target date, with a forecast the screen marks as an estimate.

Turnaround records audit trail of agent steps and people’s decisions, beside five count checks that reconcile the whole estate — **Audit trail**Agents and people on one trail, beside five sums that add the estate up from files received to documents kept.

Duplicates list on the same-job tab: recurring job packs a plain check calls duplicates, the agent’s same-job call, and rows waiting for a person — **Duplicates and look-alikes**Same job, near and exact duplicates in separate tabs, each row with what a plain check says and what the agent says.

Classification settings: type confidence line at 92%, page boundary line at 85%, near duplicate similarity at 90%, same-job check locked on — **Settings**Confidence lines, near duplicate similarity and the locked same-job check. The administrator proposes, and the records lead approves.

demoSample data, from the working solution.

Who rules on a job pack, and who countersigns

One person can never decide and countersign the same call.

Turnaround records lead

Sees: The whole estate, the holes in it and the chains waiting
Signs: Decides duplicates and same-job chains, and proposes each release to engineers

Document control supervisor

Sees: Every source and batch, and each unsure page boundary
Signs: Countersigns rulings and releases

Turnaround engineer

Sees: Only the documents that need a judgment, with the reason for each
Signs: Rules on what is flagged, in packets by unit and reason

Internal auditor

Sees: The trail, the decisions and the count checks
Signs: Nothing; reads and exports

The agents

Seven agents on one estate, and the line each must not cross

Each hands to the next, or to a person when it is unsure.

Intake agent

Opens each file read only, fingerprints it and holds what it cannot open.

Never changes, moves or deletes a source
Never skips a file silently

Page splitter

Finds where each document starts, page by page.

Never drops or reorders a page
Below its line, never decides

Type sorter

Sorts into seven types with a confidence and cited reasons.

Never accepts below the line

Duplicate finder

Finds exact copies by fingerprint and near copies by text and page image, and proposes the kept copy.

Never removes anything

Same-job checker

Tells a recurring job pack from a duplicate using turnaround, dates, work orders and readings.

Never calls two turnarounds duplicates on text alone

Field reader

Reads fields per type, ties each to its page and checks tags.

Never guesses; a blank stays blank

Review router

Flags possibly obsolete, incomplete and low-confidence documents, and groups them in packets by unit and reason.

Never rules on a document
Never releases without a countersignature

Retiring a record or calling it obsolete stays with your people

The agents sort, read and propose. A person makes these calls.

Nothing is deletedA retired document points to the kept copy and stays in the archive. Disposal is your own process, under your schedule and legal holds.

Obsolete is an engineer’s callThe review router flags what may be obsolete, and an engineer rules on it.

Two turnarounds are never duplicates on text aloneThe same-job checker needs the turnaround, dates, work orders and readings.

A blank stays a blankThe field reader ties each value to its page and does not guess a missing one.

An unsure page boundary goes to a personBelow its confidence line the page splitter does not decide, and no page is dropped or reordered.

Where decades of turnaround records actually sit

Sources stay read only. A file that cannot be opened is held and counted, so nothing goes missing without a note.

Document managementSystem exports

Closeout sharesTurnaround file shares

Boxes of paperScanned by the contractor, often one PDF per box

Contractor drivesHandover drives, read only

Planning driveThe planning shared drive

Equipment registerCurrent tags, to check old ones against

What records leads ask about turnaround document classification

What is turnaround document classification?

Turnaround document classification sorts mixed turnaround records by document type, so planners can use them. Here agents first split each file into the documents inside it, then read key fields with a citation for each value, find exact and near duplicates, and flag doubts. A person decides and a second person countersigns.

How does it tell a recurring job pack from a true duplicate?

A same-job check reads the turnaround, dates, work orders and readings, then cites the differences it found. When it is unsure, it goes to the records lead. A disagreement is recorded with a reason code and waits for a second person.

How do we know the sorting is right?

Type calls are sampled, and the result sits beside each type on the sort screen. Below the confidence line a document goes to an engineer. Duplicates are signed off rule by rule by a named person, with a second person countersigning.

Can it handle legacy engineering document migration?

It is the foundation for it. Documents leave this work typed, read and deduplicated, with old tags mapped to the equipment register where a match exists. Decisions, reasons and pointers to kept copies stay on an audit trail for an outside reviewer.

What does near duplicate document detection miss on its own?

Dates and event context. A text comparison sees two job packs with the same pack number and nearly the same words, and it calls them copies. Near copies still get grouped here, and the copy holding a handwritten note is the one kept, but a pack from a later turnaround goes to the same-job check first.

Do engineers have to review all of it?

No. They see only what is flagged as possibly obsolete, incomplete or low confidence, grouped into packets by unit and reason, and released in waves. Nothing reaches an engineer until document control countersigns the release.

What happens to a scanned box with many job packs in it?

It is read page by page. The page splitter finds where each document starts, with the page range for each. Below its line, a person decides the boundary with a reason and another person countersigns. No page is dropped or reordered.

What happens to old equipment tags?

The field reader checks tags against the equipment register and proposes a match for old ones. A proposed tag keeps the old value until someone else signs. A tag with no match stays flagged, and its documents count as incomplete.

Shutdown, Turnaround & Outage

Start with one box
of old job packs.

We’ll walk you through the sort, the duplicates and the same-job check on your own records.

Book a demo Browse solutions

Turnaround Records Classification & Deduplication

Turnaround document classification fails when one scan holds many documents

The Problem Today

The Solution

What You Get

Turnaround document classification, stage by stage, over 21 days

614,188 files, and a count for those nobody can open

A box that looks like one file holds 12 documents

Seven document types, the hardest calls first

380,658 proposed as copies, and none retired until a person signs its rule

Three copies of one hydrotest record, and the one with a note wins

A plain check calls this pair of job packs a duplicate

The lead agrees, and nothing changes yet

Pack number, tag, work order and date, each lit on the page

4.2% of the estate goes to engineers, in 42 packets