SSCC Dataset
A synchronized multimodal 7-channel dataset for fault detection and fault classification on a single-speed chain conveyor system.
What’s included
Components provided in the SSCC v1.0 release.
Key features
Core characteristics of the SSCC dataset.
- Synchronized multimodal recordings enabling cross-modal representation learning.
- Multi-condition operating settings across speed, load, fault type, and noise.
- Benchmark protocols for fault analysis under distribution shift.
Overview
The SSCC dataset contains synchronized audio and vibration recordings collected from a laboratory single-speed chain conveyor system under diverse operating conditions. Each clip corresponds to a condition defined by conveying speed, load level, operating category (normal or fault type), and noise setting. SSCC is designed to support both fault detection (normal vs abnormal) and multi-class fault classification, enabling research on multimodal representation learning and fusion under realistic domain shifts.
Lab setup
Experimental platform and sensor configuration used for SSCC data acquisition.
Data specification
- Recording setup: Audio is captured using three devices: Zoom H5 (WAV, 44.1 kHz), iPhone 11 (MP4, 44.1 kHz), and Xiaomi Mi 9 SE (MP4, 48 kHz). Vibration signals are acquired using one single-axis sensor and one tri-axis sensor (four channels total) sampled at 100 kHz (CSV format). Two loudspeakers are used to replay real factory noise recordings to simulate realistic environmental noise conditions. Only audio tracks are used from mobile recordings.
- Segmentation: Data are collected under steady-state operating conditions and segmented into fixed-length clips (5 seconds per clip). No temporal overlap is applied between adjacent segments. All seven channels corresponding to the same time window are treated as synchronized modalities of a single sample.
- Labels/attributes: normal + 4 fault types; operating attributes (speed, load, noise).
- Splits: fault detection uses normal vs abnormal; fault classification uses separate multi-class splits.
Dataset composition (clips)
Clip counts by category, load, conveying speed, and noise setting.
| Category | Load | Velocity=20 | Velocity=40 | Velocity=60 | Velocity=80 | Velocity=100 | Total | |||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| clean | noise | clean | noise | clean | noise | clean | noise | clean | noise | |||
| normal | heavy | 69 | – | 68 | – | 69 | 96 | 77 | 97 | 68 | 96 | 640 |
| normal | med | 73 | – | 66 | – | 78 | 96 | 67 | 94 | 66 | 99 | 639 |
| normal | light | 68 | – | 86 | – | 69 | 97 | 66 | 97 | 67 | 97 | 647 |
| normal | Total | 210 | – | 220 | – | 216 | 289 | 210 | 288 | 201 | 292 | 1,926 |
| dry | heavy | – | – | – | – | 57 | 57 | 57 | 57 | 56 | 58 | 342 |
| dry | med | – | – | – | – | 69 | 57 | 56 | 57 | 64 | 56 | 359 |
| dry | light | – | – | – | – | 62 | 58 | 49 | 56 | 57 | 57 | 339 |
| dry | Total | – | – | – | – | 188 | 172 | 162 | 170 | 177 | 171 | 1,040 |
| lean | heavy | – | – | – | – | 57 | 57 | 57 | 57 | 57 | 57 | 342 |
| lean | med | – | – | – | – | 58 | 56 | 57 | 57 | 56 | – | 284 |
| lean | light | – | – | – | – | 57 | 56 | 69 | 60 | 56 | 56 | 354 |
| lean | Total | – | – | – | – | 172 | 169 | 183 | 174 | 169 | 113 | 980 |
| loose | heavy | – | – | – | – | 96 | 99 | 97 | 97 | 97 | 96 | 582 |
| loose | med | – | – | – | – | 73 | 96 | 96 | 96 | 96 | 99 | 556 |
| loose | light | – | – | – | – | 96 | 96 | 96 | 96 | 96 | 97 | 577 |
| loose | Total | – | – | – | – | 265 | 291 | 289 | 289 | 289 | 292 | 1,715 |
| screwdrop | heavy | – | – | – | – | 57 | 56 | 57 | 55 | 56 | 54 | 335 |
| screwdrop | med | – | – | – | – | 57 | 47 | 52 | 51 | 87 | 54 | 348 |
| screwdrop | light | – | – | – | – | 56 | 56 | 56 | 54 | 45 | 56 | 325 |
| screwdrop | Total | – | – | – | – | 170 | 159 | 165 | 162 | 188 | 164 | 1,008 |
| Overall | Total | 210 | – | 220 | – | 1,011 | 1,080 | 1,009 | 1,083 | 1,024 | 1,032 | 6,669 |
Dataset Directory (clips)
- ASDdataset/
- normal/
- heavy_vel40_clean/
- android/
- ios/
- audio/
- vibration/
- heavy_vel60_noise/
- ...
- heavy_vel40_clean/
- abnormal/
- dry/
- heavy_vel60_clean/
- heavy_vel60_noise/
- ...
- lean/
- loose/
- screwdrop/
- dry/
- normal/
Dataset Examples
Each example corresponds to the same condition and the same sample index, shown across sensing modalities: recorder, iOS device, Android device, and vibration sensors (4 channels).
Normal
Abnormal
Fault type notes: loose indicates that one of the two conveyor chains becomes excessively slack; screwdrop indicates that a small screw drops into the chain conveyor during operation.
Fault type: loose
Fault type: screwdrop
Benchmark
We provide two benchmark tracks with task-specific condition-level splits and unified evaluation protocols. All partitions are performed at the condition level.
The source code is available at: SSCC-Fault-Benchmark.
Fault Detection
The fault detection track follows a normal-only training paradigm. Models are trained exclusively on normal samples collected under a designated target velocity (vel=100), including all load levels (heavy, medium, light) and both clean and noise settings. No abnormal samples are included in training. Testing is conducted on a mixed set containing (i) normal samples from the remaining velocities (20, 40, 60, 80) and (ii) all abnormal samples across all velocities and load conditions.
Fault Classification
The fault classification track is formulated as a supervised multi-class classification task. All samples collected at velocity 80 are held out from training. Moreover, samples belonging to selected fault categories (dry and lean) under velocity 100 are removed from the training set and reassigned to the testing set.