Skip to content

AI uses for records

Engineering repair replacement decisions: what 6,169 log entries omit

By SourceX Editorial · Updated

10 sections

Short answer

A dataset of engineering repair replacement decisions needs the finding, the options, the cost of each, the choice and whether it held. A 6,169-entry aviation maintenance log shows how little of that an operational record keeps: every entry names the action, 1.6 percent cite a procedure, 0.2 percent mention a limit, none records a cost, and nothing links a repair to its later failure.

Record timeline

Engineering repair

This page covers engineering repair replacement decisions: what 6,169 log entries omit.Illustrative

Key takeaways

  • Of 3,846 replacements in a 6,169-entry aviation log, 3,386 renewed consumables such as gaskets; only 329 replaced a component, where a real choice existed.
  • Cracks split: 284 entries patched or stop-drilled them and 71 replaced the component, with no crack size, limit or price recorded in either group.
  • No entry records a cost or labor figure; six say FOR BILLING PURPOSES ONLY, which places the money in another system.
  • The richest decisions appear as sequences: in 17 episodes a repair attempt, often with a compression reading, came before the replacement.
  • Without dates and asset identifiers, a log cannot show whether a repair held, so outcome labels need a join to the asset register.

What must a dataset of engineering repair replacement decisions hold?

A record of an engineering repair-or-replace decision has to show the condition found, the options that were open, what each option would cost in money, labor and downtime, the constraint that tipped the choice, the choice itself, and whether it held. Remove any one and the record still looks complete, but a model trained on it can only learn which action usually follows which complaint.

That is a narrower skill than the one buyers want. Maintenance planners, reliability engineers and platform teams make the same call on pumps, engine cylinders, servers and software services: patch the existing unit and accept the risk it fails again, or replace it and pay more now. The interesting cases are the ones where the same finding leads to different choices, because only the missing fields can explain the difference.

This article tests how much of that decision a real maintenance log captures. The evidence is the aviation logbook file published with MaintNet, an open research collection of maintenance text, which pairs each reported problem with the action taken. Counting it shows what an operational log keeps, what it drops, and which joins a dataset needs to put the decision back together.

What 6,169 aviation log entries record about repair and replacement

The log records the action reliably and labels most entries as a replacement, but most of those replacements are routine parts rather than decisions. The file holds 6,169 entries, each an identifier, a problem description and an action; 5,900 of them mention an engine part such as a cylinder, baffle, intake, exhaust or oil system. Classifying each action with fixed whole-word rules gives 3,846 replacements (62.3 percent), 1,161 repairs such as tightening, cleaning, adjusting, patching or stop-drilling a crack (18.8 percent), 546 checks with no repair (8.9 percent) and 616 entries that fit none of those (10.0 percent).

Splitting the replacements by what was replaced changes the picture. Of the 3,846, 3,386 renewed a consumable such as a gasket, seal, screw, rivet or plug, 329 replaced a component such as a cylinder, valve, baffle, exhaust part or the engine itself, and 131 do not say, often because the action field stops at 55 characters. The actions that look like a genuine choice between repairing and replacing a component number in the hundreds, not the thousands.

The file also lacks two fields any outcome analysis depends on: it carries no date and no aircraft or engine identifier. A sibling dataset in the same repository, built from FAA accident and incident reports, is described by its maintainers as containing structured data such as details of the aircraft, failure codes and dates, which shows the gap is a property of this export rather than of maintenance records in general.

Same finding, different call: where the log splits

For several kinds of finding, the same description leads sometimes to a repair and sometimes to a component replacement, and the log never says why. Grouping entries by a condition word in the problem text makes the split visible. Loose parts are repaired far more often than a component is replaced (315 to 2), broken parts mostly lead to a component replacement (58 to 8), and cracks, wear, chafing and leaks fall in between.

Cracks are the clearest contested case. Of 406 entries whose problem mentions a crack, 284 record a repair, often stop-drilling the crack or riveting a patch of like material, and 71 record replacing a component. None of the 406 gives a crack size, and none states a price, though crack length against a limit and the cost of the part are the facts a technician weighs. A model trained on these pairs can learn that cracks are usually patched, but not when patching is the wrong call. The table and chart below list the six largest condition words:

Same finding, different call: where the log splits
Condition word in the problemEntriesRepairedComponent replacedWhat would explain the split
Loose4033152Torque value found and specified; thread condition
Crack40628471Crack length and location against the repair limit; part price and lead time
Leak3,0929446Leak rate; whether a gasket renewal was tried first (2,838 more entries replaced a consumable)
Worn1595512Measured wear against the manual limit
Chafed44179Depth of chafe; whether the cause was rerouted
Broken170858Whether a repair scheme exists for the part at all
A dumbbell chart of six condition words in an aviation maintenance log, each row joining the number of repairs to the number of component replacements: loose 315 and 2, crack 284 and 71, leak 94 and 46, worn 55 and 12, chafed 17 and 9, broken 8 and 58.
Repairs versus component replacements by condition word, MaintNet aviation log. Source: SourceX count of the file in nd-crane/trusted_ke, October 2026.

Where the reasons and the costs went

The log keeps almost none of the evidence behind a choice: reasons, limits and costs appear in a few percent of entries or not at all. A whole-word search finds a reference to a procedure, such as IAW followed by a manual, in 97 entries (1.6 percent), a compression reading such as 20/80 in 85 (1.4 percent), a limit word such as limits, beyond or exceeds in 10 (0.2 percent) and a deferral in 1. No entry contains a cost, labor, quote, price, invoice or warranty word, or a dollar sign. Six entries read FOR BILLING PURPOSES ONLY, which says plainly that the money lives in another system.

That is consistent with what the rules ask a maintenance record to contain. Under 14 CFR 43.9, an entry for maintenance work must describe the work performed, give the date it was completed and identify the person who did it, with the signature and certificate of the person approving it for return to service. Options considered and cost are not part of that list, so a log that meets the rule can still hold none of the decision.

Retention adds a second gap. The owner and operator rules in 14 CFR part 91 set the minimum retention for records of routine work as until the work is repeated or superseded, or one year after it is performed, so earlier repair attempts on a component may be gone by the time a dataset is assembled. This is general information about record contents, not legal advice; what a given operator must keep depends on its certificate and its own program.

Repair first, then replace: what the escalation trail shows

Where the log does show a decision, it shows it as a sequence: a repair attempt, a measurement, then a replacement. Grouping consecutive entries that share an identical problem description into episodes gives 5,620 episodes, of which 301 have more than one entry. Of those, 31 contain both a repair and a replacement; in 17 the repair came first and in 14 the replacement did.

The escalation episodes are the richest records in the file because they carry the evidence the single entries lack. In one, an engine cylinder with a compression reading of 20/80 had its exhaust valve staked "W/ NO AFFECT", and the next entry records the cylinder removed and replaced. In another, a cylinder at 45/80 was found with a burnt exhaust valve and worn guide, staking left it at 45/80, and a new cylinder assembly was installed. Five of the 17 escalation episodes include a compression reading.

Repairs that did not last show up too, but only by accident. Four entries describe a crack extending beyond an earlier stop-drill, the record of a repair that failed, and all four were repaired again rather than replaced. Without dates or an aircraft identifier, nothing links those entries to the original repair, so the file cannot say how long the first repair held.

The repair-or-replace decision trace and outcome rubric

A usable repair-or-replace record joins seven fields that live in at least three systems: the log or work order, the parts and finance system, and the asset register. The table compares what each field settles with what the MaintNet file carries, so a buyer can use the same rows as an acceptance checklist.

With those fields joined, every decision gets one verdict from a fixed set:

  • Repair held: a repair with no recurrence of the same finding on the same asset within the outcome window.
  • Repair failed, then replaced: a repair followed by a replacement for the same finding on the same asset.
  • Repair repeated: the same finding repaired again, such as a crack that ran past an earlier stop-drill.
  • Replaced on evidence: a replacement with a measurement against a limit, a quote comparison or a stated constraint.
  • Replaced without a stated reason: a component replacement with no measurement, limit, option or cost recorded.
  • Routine renewal: a consumable replaced as standard practice, kept out of the decision set.
  • Undetermined: no asset identifier or dates, so the outcome cannot be checked.
The repair-or-replace decision trace and outcome rubric
FieldWhat it settlesWhere a company keeps itIn the MaintNet file
Asset identity, age and hoursWhich unit, and how worn it already wasAsset register, logbook header, CMMSAbsent
Condition with a measurementHow far the finding sits from its limitInspection record, work orderText in every entry; reading in 85, limit word in 10
Options consideredWhat else was possibleEstimate, work order notes, approval threadOnly as a sequence, in 31 episodes
Cost and effort per optionLabor hours, parts price, downtimeERP, parts system, time tracking, quotesNone; 6 entries point to billing
Procedure and authorityWhich manual or approval allowed a repairManual reference, signoffReference in 97 entries
ChoiceRepair, replace or deferLogbook action, work order closeEvery entry
Outcome windowWhether the repair heldLater entries on the same assetCannot be joined; 4 repeat repairs noted

The same decision for software components and systems

Software teams make the same decision about components and whole systems, and their records split the same way. Repair means a patch, a configuration change or a refactor; replacement means a major upgrade, a new library or vendor, or a rewrite. Tickets and incident histories record the condition and the action, design documents or architecture decision records record the options, and the cost lives in cloud bills, license invoices and time tracking that engineering datasets rarely include.

Public-sector reporting shows how hard the system-level version is to document. The U.S. Government Accountability Office's 2025 review, Agencies Need to Plan for Modernizing Critical Decades-Old Legacy Systems, scored legacy systems on attributes that included age, vendor support and operating cost, and found agencies' modernization plans often incomplete. The replace decision was on record; the plan that would let anyone judge it later often was not.

For a dataset, the lesson carries over directly. Join each component's incident and ticket history to the decision document and to the spend records, and keep the recurrence window after the choice. The architecture decision article on this site covers linking a decision record to what happened next; this page covers which of those links a repair-or-replace label needs.

The strongest objection: most replacements here are not decisions

The strongest objection is that most replacements in a maintenance log are not decisions at all, so counting them overstates how much choice the log records. That is right, and it is why the rules above separate consumables from components and keep routine renewals out of the decision set. Of 3,846 replacements, only 329 name a component.

Two further limits apply. The log is a single published file, released as a research language resource rather than as a complete maintenance history, so its proportions should not be read as typical of airlines, factories or data centers. And the action text is cut at 55 characters in 945 entries, which hides some objects and some second steps.

The classification is rule-based rather than read by hand, so a few entries land in the wrong class: an action that installs a repaired exchange unit counts as a replacement, which is right at the aircraft level and wrong at the shop level. The rules are listed in full in the methods note so anyone can change one and recount.

How a buyer would use repair-or-replace records, and what to request

Buyers would use repair-or-replace records to train and test planning judgment: given a finding, a measurement, the asset's history and the cost of each option, what should happen, and did it work. An evaluation can hide the action and score a model's recommendation against the choice and its outcome; a training set can teach when the usual answer is wrong, which is exactly the slice a log without costs cannot supply.

The acceptance test follows from the field table. For a sample of decisions, a reviewer should be able to find the asset and its age, a measurement or limit, at least two priced options or a stated reason only one was viable, the choice, and the asset's next 180 days of entries. Home-service companies face a customer-facing version of this choice, covered in the HVAC repair-versus-replace article; this page covers engineering components inside an operator's own fleet or platform.

SourceX licenses business data directly from companies, and it looks for sources after a buyer describes the need instead of keeping datasets on hand. A request for repair-or-replace records is most useful when it says which asset classes matter, which of the seven fields are required, how long the outcome window must be, and what the records will be used to train or test.

Methods note: data file, classification rules and limits

Counts come from a script that reads OMIn_dataset/data/MaintNet_data/Aircraft_Annotation_DataFile.csv in github.com/nd-crane/trusted_ke at commit 135462a; the folder's README says the MaintNet data was downloaded and pushed there on September 17, 2023. No repository code was run. Text is upper-cased and every word test is whole-word: a match may not touch a letter or digit on either side. The first matching class wins. Replacement: REPLACED, REPLACE, REPLACEMENT, NEW, SERVICEABLE, OVERHAULED, R&R, R/R or INSTALLED REPAIRED. Repair: TIGHTENED, RETIGHTENED, SECURED, RESECURED, ADJUSTED, READJUSTED, CLEANED, REPAIRED, PATCH, PATCHED, PATCHES, SEALED, RESEALED, STAKED, TORQUED, RETORQUED, REATTACHED, SAFETIED, RESAFETIED, SAFETY WIRED, DRILLED, RIVETED, LUBRICATED, LUBED, REINSTALLED, RESEATED, RTV, SILICONE, DOUBLER, REPOSITIONED, ALIGNED, REALIGNED, STRAIGHTENED, SPLICED, REWORKED, DRESSED, BLENDED, SERVICED, FILLED, GLUED, BONDED or FABRICATED. Check: RAN, RUN, OPS, CK, CKD, CHECKED, INSPECTED, VERIFIED, FOUND, TESTED, NO DEFECT, COULD NOT DUPLICATE or CND. Anything else is other.

A replacement is consumable if the action has GASKET(S), SEAL(S), SCREW(S), NUT(S), BOLT(S), RIVET(S), PLUG(S), GROMMET(S), FASTENER(S), O-RING(S), WASHER(S), CLAMP(S), HARDWARE, SAFETY WIRE or COTTER; otherwise component if it has CYL, CYLINDER(S), ENGINE, INJECTOR(S), SERVO, MAGNETO, MAG, VALVE(S), PUMP, PISTON, ALTERNATOR, STARTER, MUFFLER, EXHAUST, CARBURETOR, CARB, PROP, PROPELLER, TURBO, LIFTER(S), PUSH ROD(S), PUSHROD(S), ROCKER, TUBE(S), HOSE(S), LINE(S), BAFFLE(S), SHROUD, COWL, COWLING, BRACKET, SUPPORT, GUIDE(S), NOZZLE(S) or BLADE(S); otherwise unspecified. Condition words in the problem: LOOSE; LEAK, LEAKS, LEAKING; CRACK, CRACKED, CRACKS; WORN; BROKEN; MISSING; TORN; CHAFED, CHAFING; FOULED. An entry is engine-related if problem or action has ENGINE, ENG, ENGINES, CYL, CYLINDER(S), BAFFLE(S), BAFFLING, INTAKE, EXHAUST, ROCKER, MAG, MAGNETO, IDLE, MIXTURE, OIL, PROP, COMPRESSION, INDUCTION, MUFFLER, INJECTOR, SPARK or PLUG(S).

An episode is a run of consecutive identifiers with identical problem text; an escalation is a multi-entry episode with a repair entry before its first replacement. Signals, in problem or action: reference IAW, 43.13, MANUAL, MM or AMM; limit LIMIT, LIMITS, BEYOND, EXCEEDS or OUT OF TOLERANCE; deferral DEFER, DEFERRED or MEL; money or labor COST, COSTS, LABOR, MANHOURS, QUOTE, PRICE, INVOICE, WARRANTY or a dollar sign; billing BILLING; compression reading one to three digits followed by /80; repeat repair BEYOND STOP DRILL, with EXISTING or PREVIOUS allowed before STOP DRILL; crack size a number or fraction followed by an inch mark, IN, INCH, INCHES or MM. Percentages are rounded once, half up. The saved output holds identifiers and classes only; the file has no personal data fields. The MaintNet paper describes the collection as maintenance logbook data from aviation, automotive and facilities domains.

Frequently asked questions

What information should a repair or replace decision record include?

At minimum: the asset and its age or hours, the finding with a measurement against a limit, the options considered, the cost or effort of each, the choice and who made it, and later entries showing whether it held. Most operational logs keep only the finding and the action.

Do maintenance logs record why a part was replaced instead of repaired?

Rarely. In a 6,169-entry aviation log, 1.6 percent of entries cite a procedure, 1.4 percent record a compression reading, 0.2 percent mention a limit, and none records a cost. The reason usually has to be rebuilt from inspection records, estimates and the parts system.

How do you tell a routine part renewal from a real repair or replace decision?

Separate consumables such as gaskets, seals and fasteners from components such as cylinders, valves and pumps. In the aviation log studied here, 3,386 of 3,846 replacements were consumables, so counting every replacement as a decision would credit the log with 3,846 choices where only 329 component replacements are on record.

Can a repair or replace dataset be built without cost data?

It can teach which action usually follows a finding, but not when the usual action is wrong, because cost, downtime and parts availability are often what tips the choice. Joining the log to the parts, invoice and time-tracking systems is what turns an action log into decision data.

How long should the outcome window be for a repair decision?

Long enough to see a failed repair recur on the same asset; this article's acceptance test uses 180 days of later entries. That requires asset identifiers and dates, which the published aviation file lacks, so its four repeat crack repairs cannot be traced back to the first repair.

Sources

  • The MaintNet aviation maintenance log file (6,169 entries with identifier, problem and action fields) counted and classified in this article. Source
  • The repository README for the MaintNet data folder: the data was downloaded and pushed to the repository on September 17, 2023, with links to the MaintNet paper and site. Source
  • The OMIn repository README: its FAA incident data also contains structured data, such as details of the aircraft, failure codes and dates. Source
  • Akhbardeh, Desell and Zampieri, MaintNet (2020): a collaborative open-source library of maintenance logbook data from aviation, automotive and facilities domains with preprocessing tools (paraphrased from a search excerpt). Source
  • 14 CFR 43.9: a maintenance record entry contains a description of the work performed, the completion date, the name of the person who performed it and the approving person's signature and certificate (paraphrased from a search excerpt). Source
  • 14 CFR part 91 subpart E (section 91.417): records of maintenance work are kept until the work is repeated or superseded, or for one year after it is performed (paraphrased from a search excerpt). Source
  • GAO-25-107795 (2025), Agencies Need to Plan for Modernizing Critical Decades-Old Legacy Systems: systems were scored on attributes including age, vendor support and operating cost, and modernization plans were often incomplete (paraphrased from a search excerpt). Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies