GearXAI
Industrial gearbox with sensor traces and explainability heatmaps
IJCAI-ECAI 2026 Tuesday, August 18 · 11:30-12:30

Accepted to the IJCAI-ECAI 2026 Competitions and Challenges Track

GearXAI

An Explainable Neuro-Symbolic Gearbox Fault Diagnosis Challenge

Bremen, Germany · August 18, 2026 · Room 4.1 A

Read the abstract Session details
Task
9-class fault diagnosis
Input
8-channel vibration windows
Artifact
ONNX, CPU-only
Ranking
Explainability first

IJCAI-ECAI 2026 competition session

Bremen session details

The GearXAI competition session will take place on Tuesday, August 18, 2026, from 11:30 to 12:30 in Room 4.1 A. The room can accommodate up to 220 participants.

Presentation schedule

All times are local to Bremen (CEST)

  1. Organizers' opening remarks and competition results11:30-11:35 · 5 min
  2. Team presentation block11:35-12:27 · 8 min per team
  3. Closing remarks12:27-12:30 · 3 min

Presentation order · 8 minutes per team

  1. First PlaceTeam J2W
  2. Second PlaceTeam hakimi
  3. Third PlaceTeam CUFE
  4. Honorable Mention: Fourth PlaceTeam JS
  5. Honorable Mention: Fifth PlaceTeam AQiDA
  6. Honorable Mention: Best Macro-F1 ScoreTeam Sorndus

Q&A: Presenting teams should reserve 2-3 minutes of their allocated slot for questions and discussion.

Attendance format

In person only

Under IJCAI-ECAI policy, Zoom, Google Meet, and all other forms of live remote participation are not permitted.

Visa-related absence

Prerecorded presentation

Winning teams unable to attend because of visa issues may submit a presentation recording of approximately 8 minutes. Send the recording, or a download link, directly to eduard.hogea00@e-uvt.ro.

Submission deadline Calculating... Days · Hours · Minutes · Seconds
Presentation files

No advance submission required

Teams presenting in person do not need to send their presentation files in advance.

Competition abstract

Transparent diagnostics for rotating machinery

Title
GearXAI: An Explainable Neuro-Symbolic Gearbox Fault Diagnosis Challenge

GearXAI is a single-track IJCAI-ECAI 2026 competition focused on explainable and neuro-symbolic approaches for multiclass gearbox fault diagnosis from vibration time series. Participants submit ONNX models that are evaluated in CPU-only, containerized runs for reproducible inference. Submissions must meet a minimum predictive-performance requirement, after which ranking is determined by an automated explainability score that quantifies faithfulness under controlled perturbations and stability under small input changes. The challenge releases a realistic drivetrain dataset and a unified devkit with ready ONNX baselines, validation, leaderboard metric reporting, and submission packaging to enable transparent comparison and follow-up research.

Competition task

Predict the fault. Explain the evidence.

01

Drivetrain data

Participants work with synchronized multichannel signals from the Drivetrain Dynamics Simulator testbed, sampled at 5120 Hz across multiple speed-load regimes and fault states.

02

Shared label space

Each window is classified into one of nine classes: healthy operation plus eight fault types, aggregated across planetary gearbox operating regimes.

03

Standardized tensors

The official loader outputs windows shaped [N, 8, 100]. Models return class probabilities [N, 9] and relevance maps [N, 8, 100] for XAI scoring.

Evaluation

Accuracy is reported. Explanations decide the rank.

XAI

Explainability Score

Faithfulness
40%
Mechanistic relevance
40%
Simplicity
20%

Every valid submission is ranked by an automated Explainability Score combining deletion/insertion faithfulness, physically plausible relevance, stability under small perturbations, and compactness. The leaderboard also reports whether each model reaches 80% macro-F1 on the hidden test set.

  • Final submissions are single ONNX models evaluated with ONNX Runtime.
  • Organizer evaluation runs offline in a sealed, CPU-only container with no internet access.
  • Determinism, runtime, memory, and explanation sanity checks are part of the evaluation process.
  • Teams may include up to five members; each participant may join only one team.

Official standings

GearXAI leaderboard

Submissions
--
Eligible
--
Next update
--:--:--

Loading official standings...

Recomputing with the patched faithfulness evaluator. The table below shows the latest published snapshot until the full re-evaluation finishes.

Mechanical scoring was upgraded on June 9, 2026 to improve frequency resolution and perturbation-stability measurement. The new formula checks whether explanations emphasize mechanically meaningful signal regions. All submissions were re-evaluated identically.

Faithfulness Evaluation Update - July 1, 2026. We found and fixed an issue in the faithfulness evaluation. Faithfulness measures how the predicted-class confidence changes when the input signal is progressively deleted or inserted according to the submitted relevance map. As stated in the model-output requirements, submitted models are expected to return probabilities, with confidence values in the range [0, 1]. This requirement was already checked by the devkit evaluator on the original input samples. However, the previous faithfulness implementation did not account for the fact that some submissions could produce out-of-range confidence values on the perturbed inputs used during deletion/insertion evaluation. Without bounding these values, the faithfulness score could be artificially increased. This has now been patched: the predicted-class confidence used for deletion/insertion AUC is clipped to [0, 1] before computing faithfulness. To keep the comparison fair, all submitted models are evaluated with the same patched codebase. The devkit has also been updated to reflect this fix.

Only each team's best submission is considered for ranking. Earlier submissions may be shown for progress reference to support development; this is subject to change if evaluator load becomes too high.

Rank Team submission Final score Macro-F1 Faithfulness Mechanical Simplicity Status
Loading official standings...

Last published:

Submission

From dataset to final ZIP.

Participants train their own model, export it as ONNX, then run the devkit on that model to produce the final submission.zip for manual upload.

  1. 01 Start with the dataset

    Use the released DDS-SEU PGB windows and prepare the public validation split once with the devkit.

    gearxai prepare-data --windows-dir data/windows_100 --out prepared Open dataset
  2. 02 Download the devkit

    Install the submission packager, ONNX validator, and local metric report writer.

    Download devkit
  3. 03 Train and export your model

    Your model must accept [N, 8, 100] windows and return probabilities plus relevance maps.

    model.onnx
  4. 04 Run the devkit on your ONNX

    The command validates the model, computes public validation metrics, and creates the artifact.

    Creates submission.zip
  5. 05 Manually upload the ZIP

    Upload submission.zip through the participant form.

    Open upload form
Core command gearxai package --model model.onnx --data-dir prepared --split validation --out submission.zip

Awards and recognition

Recognition for top GearXAI teams

Top-ranked teams will receive official competition certificates, be listed on the GearXAI website, and be invited to present their methods during the GearXAI session at IJCAI-ECAI 2026 in Bremen. Subject to conference space, selected teams may also be invited to present posters.

🥇 1st Place

GearXAI Winner

Winner Certificate, featured method presentation at IJCAI-ECAI 2026, featured speaker invitation to the Neuro-Symbolic Artificial Intelligence special session at IEEE SYNASC 2026 in Timișoara, and first priority for any confirmed compute award.

🥈 2nd Place

GearXAI Runner-Up

Runner-Up Certificate, method presentation at IJCAI-ECAI 2026, website recognition, and access to any confirmed compute award subject to availability.

🥉 3rd Place

GearXAI Third Place

Third-Place Certificate, method presentation at IJCAI-ECAI 2026, website recognition, and access to any confirmed compute award subject to availability.

Mentions

Honorable Mentions

Website recognition and optional short presentation or poster invitation, depending on submission quality and available session space.

The IJCAI-ECAI competition session is in person only; live remote participation is not permitted. Winning teams unable to attend because of visa issues may provide a prerecorded presentation for playback in the room.

The organizers are working to provide a compute award of up to 100 GPU hours on the GPU@Timișoara high-performance computing infrastructure at West University of Timișoara, including H200 GPU resources. The final allocation will follow the progress and availability of the university infrastructure and will be confirmed before the final evaluation phase.

Timeline

Competition milestones

Final submissions close in Calculating...
Days
--
Hours
--
Minutes
--
Seconds
--
  1. Official competition announcement submitted to IJCAI-ECAI
  2. Call for Participation, dataset details, and unified devkit release
  3. Development submission window and organizer validation feedback
  4. Final ONNX submission deadline, 23:59 AoE
  5. Winner notifications
  6. GearXAI competition session in Room 4.1 A at IJCAI-ECAI 2026 in Bremen

Resources

Competition resources

Dataset

Public DDS-SEU PGB windows, metadata, and evaluator-ready preparation workflow.

Open on Hugging Face

Devkit

Strict participant package with the ONNX validator, metrics report, and submission ZIP builder.

Download devkit (.zip)

Upload

Manual upload page for the final submission.zip generated by gearxai package.

Open upload page

Organizers

Organizing team

GearXAI is organized by researchers affiliated with UVT, HRIA, and Xi'an Jiaotong University.

West University of Timișoara Timișoara, Romania HRIA Romanian Hub for Artificial Intelligence Xi'an Jiaotong University Xi'an, China

Cite us

Referencing the GearXAI challenge

If GearXAI is useful for your work, please cite the challenge website until the official competition proceedings or archival report is available. 

@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/}
}