Discord

Checklist AICC 2026 Round 9 · Task 2

Buried Fault

Classify the fault type of multichannel vibration recordings and localise the fault event in time, although training labels give only the fault type.

  • Time series
  • Weakly supervised time-series classification and temporal localisation
Read the official statement

The task

A turbine plant records six vibration sensors on every machine. Occasionally a machine enters one of six fault modes for a second or two. Archived recordings are labelled with the fault type, but the time of the fault was never recorded. An early-warning tool must both identify the fault type and locate the event, on machines it has never seen.

Each recording has six channels sampled at 64 Hz for 32 seconds (2048 time steps per channel), with missing samples stored as NaN; missingness differs by channel and machine. Each recording contains exactly one fault event occupying between 2 and 15 percent of the time steps, anywhere in the recording, alongside normal operation, noise and other non-fault disturbances. The 12 training machines come from 6 sites; test machines come from sites that do not appear in training.

For each test recording the contestant predicts the fault type (0–5) and the integer start and end sample indices of the event (0 <= start < end <= 2048). Event times are never provided for training recordings.

Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.

At a glance

You get
train.npy (float32, shape (2400, 6, 2048), contains NaN), train_meta.csv (recording_id, machine_id, site_id, label), test.npy (float32, shape (1800, 6, 2048)), test_meta.csv (recording_id, machine_id, site_id), sample_submission.csv.
You submit
submission.csv with header recording_id,label,start,end; one row per test recording.
Scoring
score = 0.5 × macro_F1 + 0.5 × mean_IoU, where macro_F1 is over the six fault types and IoU is the intersection over union of predicted and true intervals, counted as zero when the predicted fault type is wrong. Baseline 0.10, reference solution 0.70.
Rules
  • Individual participation (maximum team size 1); at most 60 submissions per day.
  • AICC contest rules (stated on each Kaggle rules page, not enforceable): no use of LLMs for writing code or getting task ideas; no internet use other than official library documentation and the contest platform; no communication with anyone during the contest; clarifications only via the #clarification-requests channel on the AICC Discord server.
Format
AICC Round 9, online on Kaggle, 24 Jul 2026 18:06 UTC – 26 Jul 2026 18:00 UTC.

Notebook

Competition

Solution

Details

Year
2026, Online (Kaggle)
Round
Round 9 · Task 2
Language
English
License
Varies by task: Word Lookups — MIT; Buried Fault and Shuffled — CC BY-NC-ND 4.0 (Kaggle competition licences). Solutions repository: MIT., as stated by the source