PI logo

LEA · Logbook Extractor & Analyzer

Structuring decades of helicopter-engine maintenance records into traceable data, with a human validating every step.

PI Project · Projects/ LEA · Logbook Extractor & Analyzer· All projects

§01 The problem

Helicopter engines outlive the rules their paperwork was written under

A helicopter engine can stay in service for forty or fifty years. Its paper record has to follow it that whole life, because in aviation the record is the airworthiness: a part whose history cannot be established is a part that cannot fly. And over those decades, everything around the record changed. Regulatory regimes succeeded one another (in Europe alone, from national authorities to JAA to EASA), certificate formats were replaced, engines were sold across borders into new jurisdictions and new languages, and the documents themselves aged from typewritten originals to imperfect scans.

The result sits in maintenance archives everywhere: records that are complete and legally valid, yet written against referentials that no longer exist. Turning them into data a modern maintenance operation can compute with is today largely manual, expert work. That translation step is the problem LEA addresses. Here is what it looks like for a single engine.

One engine · 1/5

1985: the file opens

The engine enters service with a certification file issued by the national authority of its day, typewritten, in French. Perfectly valid then; a historical artefact now. Nothing in it anticipates the referentials that will judge it decades later.

One engine · 2/5

The nineties: two regimes in one binder

By its first overhauls, European standardisation is arriving. The paperwork straddles the old national system and the new joint one, and the same physical engine now has documents that do not share a format, a numbering scheme or a vocabulary.

One engine · 3/5

2005: the record changes continent

The engine is sold abroad, so its history is converted into another authority's format and another language. There are more than 190 national aviation authorities, each with its own forms and requirements; an engine only needs to cross one border for its record to fragment.

One engine · 4/5

2010: current standards join the pile

Later maintenance is documented under today's European regime. The newest paperwork is clean and standardised; it simply sits on top of twenty-five years of paperwork that is neither.

One engine · 5/5

2025: someone has to read all of it

The engine finally comes back for a shop visit. In between, it lived far from its manufacturer, with independent shops, operators, a resale: a history written by many hands and seen only in fragments. Before any work, that forty-year file must be read, reconciled and trusted: four regulatory regimes, five formats, at least two languages, some of it scanned, some of it handwritten. Today that reading is expert manual work. And this is one engine; a fleet means thousands of files like it.

§02 Why it is hard

Six dimensions of variation, all at once

Any single dimension below would be a manageable engineering task. A real archive varies along all six simultaneously, and the dimensions multiply rather than add.

Time

Fifty years of records across five or six documentary eras, from typewritten national certificates to today's standardised forms, with every transition period mixing both.

Geography

More than 190 national aviation authorities, each with its own forms, required fields and numbering. An EASA Form 1 is not an FAA 8130-3, and neither matches what a third country issued in 1998.

Technical variants

A single widely flown engine family can count more than 28 certified variants, plus factory options, service bulletins and retrofits, each with its own documentary consequences.

Operator segments

Military, parapublic, offshore, medical, corporate and utility operators each keep records differently, from classified proprietary formats to minimal paperwork.

Physical condition

The input is whatever survived: photocopies of photocopies, faded stamps, handwritten entries, scans made at every quality level. The information is there; it is not clean.

Language

French, English, Spanish, German and more, often mixed within one file, with era-specific technical vocabulary that no general-purpose glossary carries.

The blocking difficulty is not any one dimension. It is that reading such a record correctly requires AI engineering, general airworthiness practice and regulatory history at the same time, a combination rarer than any of its parts. That intersection is where this project lives, and it is why our team's avionics background matters as much as its machine learning.

§03 What LEA is

A translation system for regulatory history, with a human in the loop

That variability cannot be patched downstream; it has to be designed for from the start. LEA reads a legacy record and carries it, step by step, into data a current maintenance system can use, keeping the trace from every extracted value back to the exact place on the page it came from. Four functions, in order:

1EXTRACT

Pull the information off the page, whatever the page is: printed form, stamp, handwritten entry, degraded scan.

vision-language models · OCR
2CLASSIFY

Identify what the document is: its era, its issuing authority, the referential it was written against.

document classification
3NORMALISE

Translate the content into the current referential: map old designations, units and regulatory references to their present-day equivalents. Matching a part across generations is identity resolution, not string comparison: the same object can carry different designations depending on era and variant.

mappings · domain rules
4VALIDATE

A qualified person reviews and signs off, with the scanned original and the extracted values synchronised side by side. In airworthiness, human validation is a design requirement, not a temporary limitation.

human in the loop · audit trail

And what it is not

  • Not an OCR project. Reading characters is one component; the value is in knowing what a 1985 certificate means under 2025 rules.
  • Not fully automatic. Every output is built to be reviewed and signed off by a qualified person, with its evidence attached.
  • Not generic document software. It is a specialised aeronautics instrument, narrow on purpose.

§04 What it unlocks

Structured records are the ground floor of value

Structuring is not the end goal: it lays the ground for the value built on top, and that value compounds with the volume of records treated. In order of proximity:

review & audit

The immediate value. Review time drops and audits gain evidence: every extracted value stays traced to the exact place on the page it came from, every correction is logged, and stamps and signatures are preserved as digital evidence.

integration

Records that flow instead of being retyped. Once validated, the data is machine-readable by design (API, structured export), so it can feed the systems that need it: maintenance information systems, ERP and MES on the industrial side, and emerging AI assistants on the analysis side.

prediction

The raw material of predictive maintenance. A prediction is only as good as the history it learns from, and for older fleets that history is locked in paper. Validated, structured records are what make fleet-level analytics credible. The leading helicopter-engine manufacturer alone counts more than 21,000 engines in service; each one trails a file like the one above.

extension

Beyond logbooks. The pipeline is modular: the same extract, classify, normalise and validate chain applies to other airworthiness documents, from service bulletins to release certificates.

§05 External validation

Winner of the Safran Helicopter Engines Challenge

Hack2Bridge · Aerospace Valley · November 2025

The LEA demonstrator was built during the Hack2Bridge hackathon for the challenge set by Safran Helicopter Engines, and won it. The jury's problem is the one described on this page: structuring legacy helicopter logbooks with full traceability, so downstream systems and operators can work on clean, unambiguous data.