guochen129/industrial-acoustic-event-classification

Industrial Acoustic Event Classification (5-class)

A 5-class industrial acoustic event dataset for sound event classification under severe class imbalance.

Classes

Class	Chinese	Files	Independent source events
Arc Fault	电弧	157	107
Arc Welding	电焊	142	16
Lighter Sound	打火机音频	247	9
Lithium Battery Explosion	锂电池爆炸	17	5
Background Noise	背景噪声	4000	4000
Total		4563	4137

Audio Format

• Sample rate: 8 kHz, mono, 16-bit PCM WAV
• Duration: 10 seconds per segment (segmented from longer recordings)
• Preprocessing for training: RMS normalization (target_rms=0.1), 128-bin log-mel spectrogram (n_fft=1024, hop_length=256)

Event-Aware Split

A critical property of this dataset is that multiple segments originate from the same recording session (source event). For example, the 247 lighter-sound segments come from only 9 recording sessions (one session alone produces 80 segments). Random (hash) splitting leaks same-event segments across train/test, inflating metrics.

manifest_event_split.csv provides the event-aware 5-fold grouping:

Column	Meaning
fname	audio filename
class_name	class label (Chinese)
class_label	integer label 0–4
event_id	source-event identifier (segments from the same recording share this)
fold	fold assignment 1–5 (same event always in the same fold)

Use event_id as the grouping key for any split — never split segments from the same event across train/val/test.

Directory Structure

.
├── 8k/
│   ├── 电弧/            (157 files, class_label=0)
│   ├── 电焊/            (142 files, class_label=1)
│   ├── 打火机音频/       (247 files, class_label=2)
│   └── 锂电池爆炸/       (17 files, class_label=3)
├── 背景/                (4000 files, class_label=4)
└── manifest_event_split.csv

Reproducibility

The companion code repository (industrial-acoustic-event-classification) resolves audio paths as:

• signal classes (label 0–3): <ROOT>/8k/<class_name>/<fname>
• background (label 4): <ROOT>/背景/<fname>

where <ROOT> points to this dataset. config.py variables FOUR_CLASS_ROOT and BACKGROUND_CLASS_DIR correspond to these paths.

Known Limitations

• Lithium Battery Explosion has only 5 events / 17 files. In a 5-fold event-aware split, each test fold contains ~1 event, making this class statistically unreliable (bootstrap 95% CI spans [0, 1]).
• The dataset is highly imbalanced: background noise dominates (4000/4563 = 88%).

Access

This dataset is gated: the dataset card, metadata and file list are public, but downloading the audio requires an approved access request on the Hugging Face Hub. No data-sharing agreement is in place yet.

License

To be determined (pending a data-sharing agreement with the data custodian).

Citation

@misc{industrial_acoustic_events,
  title = {Industrial Acoustic Event Classification (5-class)},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/guochen129/industrial-acoustic-event-classification}}
}