Dataset Overview

Dataset Description

Summary

• Name: Enriched DCASE 2023 Challenge Task 2 Dataset
• Category: Audio Classification
• Size: 1K < n < 10K
• Tags: Anomaly detection, abnormal sound detection, acoustic condition monitoring, machine fault diagnosis, machine learning, unsupervised learning, acoustic scene classification, acoustic event detection, acoustic signal processing, audio domain transfer, domain generalization
• License: CC‑BY‑4.0

Structure

Data Instances

• Audio: Mono, 10 s duration
• Path: Audio file path
• Section: Integer indicating section
• d1p: Parameter name
• d1v: Parameter value
• Domain: Integer (0 = source, 1 = target)
• Class: Integer indicating machine type
• Label: Integer (0 = normal, 1 = abnormal)
• Anomaly Index: Integer from local outlier factor algorithm
• Anomaly Score: Float from local outlier factor algorithm
• Embedding: Audio embedding generated by an audio spectrogram transformer

Data Splits

• Development Set: Train and test splits
  • Train: 7,000 instances
  • Test: 1,400 instances
• Additional Training Set: Train only, 7,000 instances
• Evaluation Set: Test only, 1,400 instances

Creation

Source Data

• Includes normal and abnormal sounds from seven machine types, each providing a section with training and testing data.
• Recordings contain machine operation sounds and environmental noise.

Supported Tasks & Leaderboard

• Task: Abnormal sound detection for machine condition monitoring
• Requirements: Unsupervised learning, domain generalization, training on new machine types, training on single machine type data

Considerations

Social Impact

• To be added

Bias Discussion

• To be added

Known Limitations

• To be added

Additional Information

Baseline Systems

• Baseline code is available on GitHub, providing a reasonable performance starting point for novice researchers.

License Information

• The original data was created by Hitachi, Ltd. and NTT Corporation and is released under the Creative Commons Attribution‑NonCommercial‑ShareAlike 4.0 International (CC BY‑NC‑SA 4.0) license.