Plane-text

Overview

Aircraft maintenance runs on text: the regulations that govern it, the airworthiness directives that order inspections and repairs, the bulletins and advisory circulars that explain them, and the reports mechanics and investigators write when something goes wrong. Airlines hold large internal collections of this kind of text, but little of it is available for open research.

Plane-text gathers the public side of it from the FAA, the eCFR and the NTSB, cleans it, and publishes it as versioned datasets: the same kinds of documents a maintenance organization works with, ready for search, analysis and language models.

Explore

Query the FAA airworthiness directives with SQL, right here. The query runs in your browser with DuckDB-WASM, which reads only the parts of the parquet file it needs. The table is ads, one row per directive.

Datasets

Each dataset is published as dated snapshots of parquet files.

14 CFR, Aeronautics and Space ecfr-title14

6,357 sections and appendices. Source: eCFR.

Airworthiness directives faa-ads

20,184 final rules, 1946 onward. Source: FAA Dynamic Regulatory System.

Proposed rules faa-proposed-rules

8,290 proposed ADs and 548 rulemaking notices. Source: FAA Dynamic Regulatory System, Federal Register.

Special Airworthiness Information Bulletins faa-saibs

1,339 bulletins. Source: FAA Dynamic Regulatory System.

Advisory circulars faa-advisory-circulars

1,686 circulars. Source: FAA Dynamic Regulatory System. Some scanned pages still await OCR.

Aviation Maintenance Alerts faa-maintenance-alerts

109 monthly issues, 2004–2012. Source: FAA Dynamic Regulatory System.

Service Difficulty Reports faa-sdrs

1.77 million reports, 1995 onward. Source: FAA SDRS.

Accident and incident narratives ntsb-narratives

90,108 narratives, 1982 onward. Source: NTSB aviation databases.

Accident findings and event sequences ntsb-findings

Findings, events and causes for the same accidents. Source: NTSB aviation databases.

How it is built

Every dataset comes in two layers. Raw is the source as delivered, converted to parquet with its own field names and nothing changed. Processed is built only from raw: cleaned, typed, one row per document or report, with coded fields decoded into readable labels next to their codes. Each file records where it came from and when.

The pipelines are Python, and they run on Cloudflare: each run starts a container that fetches its source and writes both layers to storage. The code is on GitHub.

Status

Early and changing. The datasets above are built and stored; for now you can query the airworthiness directives here. Downloads, scheduled updates, more sources (EASA, NASA ASRS and Transport Canada among them) and a data license are next. All sources are US government works or other public records.