etl_version stringclasses 1
value | created_utc stringclasses 1
value | purpose stringclasses 1
value |
|---|---|---|
1.2.0 | 2026-09-09T07:21:52.627819+00:00 | private repository write probe |
SSCC compact balanced multimodal subset
This private derived dataset contains 288 synchronized SSCC clips from 15 medium-load, clean operating conditions at speeds 60, 80, and 100. It retains recorder FLAC audio, four anti-aliased 25 kHz vibration channels in compressed float32 NPZ, and five sparse frames from both iOS and Android videos. Five sample IDs retain both unchanged source MP4s for presentation and loader tests.
The subset is balanced between normal and fault states and balances fault-classification train/test roles per fault. It is intended for transfer learning or frozen representations, not training three deep encoders from scratch. It tests speed-domain transfer only; broader load and noise generalization require additional operating conditions.
The saved development pilot favored recorder-only audio for size and simplicity. The 25 kHz vibration choice is a size-driven research compromise: its grouped pilot metric was 0.716, but it did not pass the original strict 99% spectral-energy criterion. Retain this limitation when interpreting models.
Source: https://yucongzh.github.io/SSCC-Dataset/ (CC BY 4.0).
Policy hash: bf3c00d0367253c360b01042a271d96e618a73b110a6eb39d3b4a0e8ccee004d.
- Downloads last month
- 197