Industries / Transport and vehicles / Aviation
Maintenance and repair
This is the work of keeping airliners fit to fly: inspecting, servicing, repairing and overhauling aircraft and engines on a planned schedule, and keeping the records and parts paperwork that prove each job was done right.
6
Milestones achieved, from 1979 to 2023.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1979
A shortcut cracked a pylon
On 25 May 1979 the left engine and its pylon, the strut that holds the engine to the wing, came off an American Airlines DC-10 as it took off from Chicago O'Hare. All 271 people aboard and two on the ground were killed. To save time on a manufacturer's service bulletin, the airline had taken the engine and pylon off as one unit, held on a forklift. That method was not in the maintenance manual and McDonnell Douglas had not agreed to it, and putting the unit back cracked the very fitting that had just been inspected. The FAA revoked the DC-10's type certificate for 37 days, and aircraft certification now formally covers certain "vital" maintenance tasks.
LessonContinental Airlines had already damaged pylons with the same forklift method, and the FAA says the accident might have been avoided had that been shared with the manufacturer or other airlines.
Credit
NTSB investigation; FAA
1980
Plans built on failure effects
Airliner maintenance once meant overhauling parts at fixed intervals, a method called "hard time." For the Boeing 747, a Maintenance Steering Group wrote MSG-1 in 1968, a handbook of minimum scheduled checks, and MSG-2 followed in 1970 for the DC-10 and L-1011. In 1974 the US Department of Defense asked United Airlines to study how airlines build these programs. United engineers Stan Nowlan and Howard Heap found that preventive maintenance sometimes raised the failure rate, and their 1978 report, "Reliability-Centered Maintenance," judged each task by what a failure would cost. MSG-3, first published in 1980 by the Air Transport Association (now Airlines for America), works from the top down: it sorts the consequences of each failure into safety and economic ones and drops tasks that add no reliability.
LessonAccording to an article by engineer Gilles Zwingelstein, United's records showed that only about 11% of components had a dependable service life, so replacing parts on a fixed clock did little for most of them.
Credit
the Maintenance Steering Group of airlines and manufacturers; Stan Nowlan and Howard Heap of United Airlines
1988
Small cracks joined up
On 28 April 1988 an 18-foot (about 5.5 m) section of the upper fuselage tore away from Aloha Airlines Flight 243, a Boeing 737-200, at 24,000 feet over Hawaii. A flight attendant was swept overboard and killed, and the crew landed on Maui. The aircraft had flown 89,680 flight cycles (one cycle is a takeoff, a pressurization and a landing). Its short island hops had piled up cycles about twice as fast as Boeing's maintenance advice assumed, and many small fatigue cracks along a poorly bonded, corroded skin joint had linked into one long crack. The NTSB found that the airline's maintenance program had failed to detect the damage. The FAA then required corrosion-control programs and later rules on widespread fatigue damage in aging aircraft.
LessonCracks in a pressurized fuselage grow with the number of flights, not flying hours, so inspection schedules have to count cycles.
Credit
NTSB investigation; FAA
2018
Faults predicted from fleet data
In March 2018 easyJet agreed to use Skywise, Airbus's data platform, to predict technical faults across its A320-family fleet so parts could be changed before they broke. In the trial, which covered three kinds of technical problem on 85 aircraft, the software predicted 31 failures. Airbus's FOMAX data unit, fitted across the fleet by summer 2019, can record up to 24,000 parameters. In 2024 Swaran Sidhu, easyJet's head of fleet technical management, said the predictions helped the airline avoid 44 cancellations in July and 35 in August that year. Airbus counts about 11,600 aircraft connected to Skywise in late 2024.
Credit
Airbus (Skywise); easyJet
2021
Drones check lightning damage
After a lightning strike, mechanics must check an aircraft's skin for burn marks, usually from lifts and scaffolding. By 2021 Airbus had listed Donecle's drone in its A320-family Aircraft Maintenance Manual for lightning-strike inspections. The drone flies around the aircraft in the hangar and photographs the skin for people to review. Regional Jet Center, a repair shop, said its average lightning check fell from about eight hours to 1.3 hours. By 2025 Airbus had also approved drones from Mainblades, Boeing had added drones to its 737 maintenance manual, and the FAA had accepted Delta Air Lines' use of them, after a data package Delta submitted in 2023. Delta says a narrowbody takes under 90 minutes, against about 16 hours by hand.
Credit
Donecle (Toulouse) and Airbus; Mainblades; Delta Air Lines
2023
Forged papers on engine parts
Every spare part travels with a release certificate, such as the FAA's Form 8130-3 or EASA's Form 1, stating that it is fit to fly. In June 2023 staff at TAP Air Portugal asked Safran about CFM56 engine parts that looked older than their paperwork said, and Safran found the documents falsified. On 4 August 2023 EASA and the UK Civil Aviation Authority warned the industry about parts sold by AOG Technics, a UK broker, and airlines grounded aircraft to find and replace them. The company's director, Jose Alejandro Zamora Yrala, pleaded guilty to fraudulent trading in December 2025 and in February 2026 was jailed for four years and eight months. The Serious Fraud Office said more than 60,000 parts worth £6.9 million were sold with falsified certificates between January 2019 and July 2023.
LessonThe FAA and GE Aerospace spotted the fakes through small slips, including wrong description references, missing standard wording and wrong layout, so checking the paper itself is a real defense.
Credit
TAP Air Portugal and Safran, who caught it; EASA, the UK CAA and the FAA; the UK Serious Fraud Office
Watch for
AI checks parts paperwork
DUST Identity sells software that applies "AI-assisted verification" to FAA 8130-3s, EASA Form 1s and other airworthiness documents. It says the software flags reused documents, fields that do not agree, and odd signatures, stamps and layouts. It names no customers and publishes no accuracy figures. After the AOG Technics case, an industry group, the Aviation Supply Chain Integrity Coalition, recommended shared software to check the key fields on every release certificate.
Achieved whena regulator or another independent party runs the software on a known set of genuine and forged certificates and publishes how many fakes it caught and how many genuine ones it wrongly flagged.
Claimed by
DUST Identity, April 2026
Autonomous drones at JAL
Japan Airlines and Donecle have started a project that uses fully autonomous drones to inspect the outside of aircraft in JAL hangars. They call it the first in Japan to use drones approved in manufacturers' maintenance manuals, and say it aims to make maintenance safer and quicker. No figures have been given.
Achieved whenJAL or Japan's civil aviation regulator reports the drones in routine use on its fleet, with measured inspection times and findings compared against inspections done by hand.
Claimed by
Japan Airlines and Donecle, September 2026
Open
01
Logbook pages anyone can test
A public set of scanned aircraft maintenance-record pages taken from US accident dockets, each transcribed by hand into fixed fields (date, aircraft hours, work done, inspection or directive cited, whether a signature is present), with names and certificate numbers blacked out. Add a scoring script that measures how accurately any software reads the pages. The set is for testing reading software, never for judging whether an aircraft is airworthy.
The briefCredit · unclaimed02
Maintenance errors coded across countries
A public table of accidents and serious incidents in which investigators found a maintenance error, drawn from at least four investigation bodies (for example the US NTSB, the UK AAIB, Australia's ATSB and Brazil's CENIPA). Code each one against the FAA's "Dirty Dozen," its list of the twelve most common causes of maintenance error (lack of communication, complacency, lack of knowledge, distraction, lack of teamwork, fatigue, lack of resources, pressure, lack of assertiveness, stress, lack of awareness and norms), and link each row to its report.
The briefCredit · unclaimed
Open 01 · the brief
Logbook pages anyone can test
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of scanned aircraft maintenance-record pages taken from US accident dockets, each transcribed by hand into fixed fields (date, aircraft hours, work done, inspection or directive cited, whether a signature is present), with names and certificate numbers blacked out. Add a scoring script that measures how accurately any software reads the pages. The set is for testing reading software, never for judging whether an aircraft is airworthy.
Why it is still open
The public aviation maintenance-log sets found are typed text, not page images. A 2025 set on Zenodo holds 6,169 aircraft-engine problem-and-action records with no images, and a 2020 study of handwriting recognition on maintenance logs calls its data "very limited and noisy" and links no released data. Meanwhile the NTSB already publishes maintenance records in its accident dockets, such as 15 pages for a 2017 Cessna 421B accident. A search found no public benchmark of logbook page images.
Built with what exists
The NTSB's public docket system; PyMuPDF to split the PDFs into page images; Label Studio for transcription; AI models that read page images, such as Claude reading each page as an image the way Enveda's document reviewer does (see the Note "Seven more flagged"), plus an open-weights model for comparison; Hugging Face to host the set.
First step
Open 30 light-aircraft dockets that list a "Maintenance Records" item, note which hold handwritten scans and which hold typed text, and transcribe 50 pages to settle the field list.
Done looks like
At least 1,000 pages from at least 200 dockets under an open license, every field transcribed by two people, and a published table of field-by-field accuracy for at least two models, with dates and hour readings scored separately.
Open 02 · the brief
Maintenance errors coded across countries
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of accidents and serious incidents in which investigators found a maintenance error, drawn from at least four investigation bodies (for example the US NTSB, the UK AAIB, Australia's ATSB and Brazil's CENIPA). Code each one against the FAA's "Dirty Dozen," its list of the twelve most common causes of maintenance error (lack of communication, complacency, lack of knowledge, distraction, lack of teamwork, fatigue, lack of resources, pressure, lack of assertiveness, stress, lack of awareness and norms), and link each row to its report.
Why it is still open
Researchers code these reports by hand, one country or fleet at a time. A 2024 study, for example, coded maintenance-related helicopter accidents in Brazil from 2013 to 2023 with HFACS-ME, a maintenance version of a human-factors coding scheme, and named a larger sample as the main need. A search found no public table of coded maintenance-error accidents covering several countries.
Built with what exists
Final reports published by the investigation bodies, including the NTSB's CAROL database; Python; an AI model such as Claude filling a fixed form for each report, with a second pass that flags doubtful codes for people to check (see the Note "Seven more flagged"); a shared spreadsheet; GitHub to publish.
First step
Code the American 191 and Aloha 243 reports and 28 more NTSB airline cases whose probable cause names maintenance, by hand against the twelve causes, and publish the coding rules.
Done looks like
At least 300 reports from four or more bodies, 2000 to 2025, each coded with a link. Two people code a random 60 independently, and their agreement and the model's agreement with them are both published as Cohen's kappa, a standard agreement score, with a target of at least 0.6.