Skip to content

feat(connectors): add fraunhofer-iee/care-to-compare - #792

Open
ShaoSK wants to merge 3 commits into
OpenTSLM:mainfrom
ShaoSK:feature/care-to-compare
Open

ShaoSK wants to merge 3 commits into
OpenTSLM:mainfrom
ShaoSK:feature/care-to-compare

Conversation

@ShaoSK

@ShaoSK ShaoSK commented Nov 17, 2026

Copy link
Copy Markdown

Adds a connector for CARE to Compare (Fraunhofer IEE), a real-world benchmark for early fault detection in wind turbine SCADA data. It is the first energy/industrial dataset in the catalog.

Source

What the dataset holds

  • 95 Records, one per event file: 45 anomaly events (leading up to a documented turbine fault) and 50 normal events, across 36 turbines in 3 wind farms (A: 5 onshore, B: 9 and C: 22 offshore).
  • Per Record: one Signal per sensor statistic at 10-minute resolution (A 81, B 252, C 952) on a RegularAxis, plus the turbine's operating status (status_type_id). Skipped grid steps and empty cells are nulls. Units map to pint where unambiguous; otherwise the release string is kept as a source_unit annotation.
  • Annotations: wind farm, asset, event label, root-cause note, training_period, prediction_period, and the event_window interval.
  • Tasks:
    • ClassificationTask (event_label: anomaly / normal) scoped to the prediction period, for every Record.
    • TemporalLocalizationTask targeting the event_window annotation, for every anomaly Record.
    • ClassificationTask (fault_component) for the 18 anomaly events of farms A and B, derived from the release's root-cause notes (documented on the card as TimeNet-derived).
  • No split across events: the release defines a training and a prediction period inside each Record.

Implementation notes

  • source.py parses and validates the release: header, row IDs, the 10-minute grid, a single asset per file, training rows before prediction rows, and event windows on the grid. It has no TimeNet types, so it is tested on its own.
  • Wind farm A names its eight energy counters (sensor_44 to sensor_51) without a statistic suffix. They read as their sensor's single statistic. All three farms' headers were checked against their feature descriptions.
  • connector.py caches each event's values as a feature-major .npy next to the extracted CSV and builds Signals with Signal.from_loader, so a build holds one event in memory at a time (wind farm C alone is about 17.8 GB of CSV).
  • Tests use a synthetic wind farm in the release's shape (gaps, empty cells, a bare counter column, both asset column names, quoted descriptions) and include a storage round trip.

Verification

  • pytest on the connector tests: 13 passed. ruff format, ruff check and ty check are clean on the new package.
  • Real data, wind farm A converted and written with TimeFDataset.write, then read back with TimeFReader.verify(): 22 Records × 82 Signals, 34 classification and 12 localization tasks. The labels (12 anomaly / 10 normal; 6 hydraulic, 3 gearbox, 2 generator bearing, 1 transformer) match the release's event table. Conversion took 76 s on a 2-core Codespace, and the written version is 84 MB.
  • I have not run a full timenet-build build fraunhofer-iee/care-to-compare over all three farms. Wind farm C needs roughly 40 GB of free disk for extraction and cache.

ShaoSK and others added 3 commits November 17, 2026 09:42
…connector API

Main changed the connector API. A label set carries its name as an
annotation, a Hugging Face connector passes its Hub settings as class
keywords and reads a mapping from part to path, BaseConnector names the
whole download result, and Annotation takes its value by keyword.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants