This repository contains the public dataset and participant devkit for GearXAI: An Explainable Neuro-Symbolic Gearbox Fault Diagnosis Challenge, accepted to the IJCAI-ECAI 2026 Competitions and Challenges Track.

The release provides processed planetary gearbox (PGB) vibration data from the DDS-SEU drivetrain setup, packaged for direct use with Hugging Face Datasets and the GearXAI evaluator. The competition task is 9-class gearbox fault diagnosis from fixed 100 x 8 multichannel vibration windows, with leaderboard submissions packaged as CPU-only ONNX models.

Competition website: https://gearxai-ijcai-ecai2026.pages.dev/

What To Use

If you are participating in GearXAI, start with the default windows_100 config.

Config	Contents	Recommended use
`windows_100`	fixed `100 x 8` vibration windows	main competition training and validation
`samples_scaled`	scaled per-timestep channel rows	inspection, plotting, and custom preprocessing

The public dataset exposes train and validation splits. The official leaderboard evaluation split is held out by the organizers and is not published.

Quick Start

from datasets import load_dataset

train_windows = load_dataset("edi45/gearxai-dds-seu", "windows_100", split="train")
validation_windows = load_dataset("edi45/gearxai-dds-seu", "windows_100", split="validation")

Participant Devkit

Download the current participant devkit:

gearxai-devkit-v1.0.1.zip

The devkit contains:

gearxai prepare-data for converting public Parquet windows to evaluator-ready NPY files
gearxai package for creating the final submission.zip
gearxai inspect-package for checking the ZIP before upload
ready ONNX baselines that participants can package as examples

Recommended local workflow:

python -m pip install .
gearxai prepare-data --windows-dir data/windows_100 --out prepared
gearxai package --model model.onnx --data-dir prepared --split validation --out submission.zip
gearxai inspect-package submission.zip --data-dir prepared --split validation

The generated submission.zip is the model artifact for leaderboard upload. Team name, email, institution, country, and team members are collected separately through the participant form on the competition website.

Release Summary

Sampling rate: 5120 Hz
Channels: 8
Window length: 100
Fault classes: 9
Operating conditions: 19

View	Train	Validation	Public total
`samples_scaled`	754,281 rows	100,719 rows	855,000 rows
`windows_100`	737,352 windows	83,790 windows	821,142 windows

Each operating condition is stored as its own Parquet shard to keep the structure transparent and easy to inspect.

Data Schema

`windows_100`

signal: nested list with shape 100 x 8
fault_code
fault_name
condition_id
speed_hz
load_nm
regime
experiment_id
window_index

`samples_scaled`

channel_1 to channel_8
fault_code
fault_name
condition_id
speed_hz
load_nm
regime
experiment_id
sample_index

DDS Testbed

The data come from the drivetrain dynamics simulator (DDS) testbed. The setup includes a controller, motor, planetary gearbox, parallel/reduction gearbox, acceleration sensors, and brake. During acquisition, three-axis accelerometers are mounted at the input end of the gearboxes and signals are sampled at 5120 Hz.

This release focuses on the processed planetary gearbox task used by GearXAI.

Channel Layout

Each sample has 8 channels:

channel_1: motor vibration
channel_2, channel_3, channel_4: RGB vibration (y, x, z)
channel_5: torque
channel_6, channel_7, channel_8: PGB vibration (y, x, z)

Fault Labels

Fault code	Fault name	Source label
`HEA`	healthy	`0Health`
`CTF`	chipped tooth fault	`1Chipped`
`MTF`	missing tooth fault	`2Miss`
`RCF`	root crack fault	`3Root`
`SWF`	surface wear fault	`4Surface`
`BWF`	ball fault	`5Ball`
`CWF`	combination fault	`6Combination`
`IRF`	inner race fault	`7Inner`
`ORF`	outer race fault	`8Outer`

Operating Conditions

This release covers:

fixed speed/load: 20_0, 30_0, 30_1, 30_2, 30_3, 30_4, 30_5, 40_0, 50_0
variable speed: Experiment1 to Experiment10

The machine-readable condition table is included in metadata/conditions.parquet.

Metadata Files

metadata/conditions.parquet: condition definitions and per-condition counts
metadata/fault_map.parquet: fault-code mapping
metadata/release_summary.json: top-level counts and release metadata
metadata/split_policy.json: public split policy

Scope

This is a processed competition release, not a raw-data mirror.

Included:

processed planetary gearbox data
public train/validation splits
fixed-window benchmark tensors
condition and label metadata
participant devkit and ready ONNX baseline models

Not included:

raw TXT signal dumps
the original raw archive
the RGB / parallel gearbox task
labeled leaderboard evaluation data

Acknowledgment

This release was made possible thanks to the efforts of Professor Dr. Ruqiang Yan. We gratefully acknowledge his work on the DDS platform and dataset, and his role both as a coauthor of the underlying work and as a member of the GearXAI organizing team.

Terms and License

The Hugging Face license tag is other because the repository combines a custom processed-dataset release with MIT-licensed devkit code.

The license boundary is:

processed Parquet data, metadata, and dataset documentation: custom GearXAI processed dataset terms in TERMS.md
participant devkit code inside downloads/gearxai-devkit-v1.0.1.zip: MIT license as declared by the devkit package and reproduced in DEVKIT_LICENSE.md

In practical terms, you may use, copy, and redistribute this processed Hugging Face release for GearXAI participation, academic/nonprofit/industry research, benchmarking, teaching, and model development, provided attribution and these terms are preserved.

This repository does not transfer ownership of upstream DDS materials and does not grant new rights to redistribute or relicense the original raw DDS archive, raw TXT files, or other upstream source materials.

This release is provided as is, without warranties of any kind.

Citation

If this release is useful in your work, please cite the GearXAI challenge and the supporting papers below.

GearXAI Challenge

@misc{hogea2026gearxai,
  title = {GearXAI: An Explainable Neuro-Symbolic Gearbox Fault Diagnosis Challenge},
  author = {Hogea, E. and Onchis, D. M. and Ivascu, T. and Yan, R.},
  year = {2026},
  note = {IJCAI-ECAI 2026 Competitions and Challenges Track},
  url = {https://gearxai-ijcai-ecai2026.pages.dev/}
}

LogicLSTM

@article{hogea2024logiclstm,
  title = {LogicLSTM: Logically-driven long short-term memory model for fault diagnosis in gearboxes},
  author = {Hogea, Eduard and Onchis, Darian M. and Yan, Ruqiang and Zhou, Zheng},
  journal = {Journal of Manufacturing Systems},
  volume = {77},
  pages = {892--902},
  year = {2024},
  doi = {10.1016/j.jmsy.2024.10.003},
  url = {https://www.sciencedirect.com/science/article/pii/S0278612524002280}
}

Rule-Guided Transformer

@article{hogea2026ruleguided,
  title = {Rule guided transformers for dynamic knowledge adaptation in rotating machinery fault diagnosis},
  author = {Hogea, Eduard and Onchis, Darian M. and Yan, Ruqiang and Zhou, Zheng},
  journal = {Advanced Engineering Informatics},
  volume = {72},
  pages = {104444},
  year = {2026},
  doi = {10.1016/j.aei.2026.104444}
}

Downloads last month: 169

Total file size:

103 MB