Source-linked AI summary
ToyADMOS2: Another dataset of miniature-machine operating sounds for anomalous sound detection under domain shift conditions
Noboru Harada, Daisuke Niizumi, Daiki Takeuchi, Yasunori Ohishi, Masahiro Yasuda, Shoichiro Saito
TL;DR
Acoustic anomaly detection needs evaluation settings that better reflect domain shifts and controlled anomaly difficulty than basic challenge scenarios. ToyADMOS2 addresses this with two large miniature-machine sound datasets containing varied operating conditions and damage depths, and reports high Toy-car AUC values across damage levels while making higher-damage Toy-train anomalies easier in the clean source domain.
Problem
Existing acoustic anomaly-detection evaluations can be less realistic than applications involving shifts in machine models, speeds, microphones, and environmental conditions, while anomaly difficulty also needs control beyond SNR.
Method
The paper constructs ToyADMOS2 from normal and deliberately damaged toy-car and toy-train operating sounds with controlled domain shifts and anomaly damage depths.
Results
ToyADMOS2 provides two domain-shift sub-datasets with over 27 k normal and over 8 k anomalous sounds each; Toy-car AUC was around 0.99 across damage levels, while higher-damage clean-source Toy-train anomalies were easier to detect.
Takeaways & Limitations
ToyADMOS2 offers a freely downloadable benchmark for assessing acoustic anomaly-detection systems under varied machine, speed, microphone, noise, and damage conditions.
Takeaways & Limitations
The example evaluation assumes many source-domain normal samples but only four target-domain normal samples per target domain for training.
Abstract
from arXiv · showhide
This paper proposes a new large-scale dataset called "ToyADMOS2" for anomaly detection in machine operating sounds (ADMOS). As did for our previous ToyADMOS dataset, we collected a large number of operating sounds of miniature machines (toys) under normal and anomaly conditions by deliberately damaging them but extended with providing controlled depth of damages in anomaly samples. Since typical application scenarios of ADMOS often require robust performance under domain-shift conditions, the ToyADMOS2 dataset is designed for evaluating systems under such conditions. The released dataset consists of two sub-datasets for machine-condition inspection: fault diagnosis of machines with geometrically fixed tasks and fault diagnosis of machines with moving tasks. Domain shifts are represented by introducing several differences in operating conditions, such as the use of the same machine type but with different machine models and parts configurations, different operating speeds, microphone arrangements, etc. Each sub-dataset contains over 27 k samples of normal machine-operating sounds and over 8 k samples of anomalous sounds recorded with five to eight microphones. The dataset is freely available for download at https://github.com/nttcslab/ToyADMOS2-dataset and https://doi.org/10.5281/zenodo.4580270.
1. INTRODUCTION
ToyADMOS2 addresses the limited realism of prior acoustic anomaly-detection evaluations by introducing controlled domain shifts and damage depths. It provides two machine-sound inspection datasets designed for varied application conditions.
- 1. INTRODUCTION: Prior challenge settings were easier than realistic applications, where machine models, operating speeds, and other conditions can differ without target-domain training data.Few open datasets were available for evaluating this need, and ToyADMOS2 adds further test variation.
- 1. INTRODUCTION: ToyADMOS2 controls evaluation difficulty through both domain differences among normal samples and statistical differences between normal and anomalous sounds.This avoids relying solely on added-noise SNR, which noise-reduction methods can mitigate.
- 1. INTRODUCTION: ToyADMOS2 provides a large-scale ADMOS dataset specifically designed to evaluate systems under domain-shift conditions.The dataset extends prior ToyADMOS data with more varied test configurations.
- 1. INTRODUCTION: The dataset covers product inspection with a toy car and moving-machine fault diagnosis with a toy train.These are the two ADMOS task types represented in the dataset.
- 1. INTRODUCTION: Its controlled variations include machine models, part configurations, operating speeds, microphone arrangements, environmental noise, and anomaly damage depth.The previous ToyADMOS dataset can be combined with ToyADMOS2 for additional test-condition variety.
2. DATASET OVERVIEW
ToyADMOS2 organizes anomaly-sound inspection around fixed-task toy cars and moving-task toy trains. Its domain shifts vary machine, operating, microphone, and environmental conditions while anomaly severity is controlled through deliberate damage.
- 2. DATASET OVERVIEW: The dataset contains two sub-datasets: toy-car product inspection and toy-train fault diagnosis for a moving machine.Each sub-dataset includes three domain-shift conditions.
- 2. DATASET OVERVIEW: The toy car runs on an inspection device and is recorded with five microphones, while the toy train runs on a railway track and uses microphones surrounding the track perimeter.The toy-car configuration changes machine models, parts, operating speeds, and microphone arrangements relative to the earlier ToyADMOS setup.
- 2. DATASET OVERVIEW: Anomalous sounds are produced by deliberately damaging machine components or adding extraneous objects, with anomaly level controlled by damage depth.This provides different significance levels for the statistical difference between normal and anomalous samples.
- 2. DATASET OVERVIEW: Environmental noise is recorded at multiple locations with multiple microphones and reproduced through four loudspeakers using the recording configurations.Some noise samples are newly recorded and others come from ToyADMOS.
- 2. DATASET OVERVIEW: Domain shifts vary machine models and parts configurations, operating speeds, microphone types and arrangements, and environmental noise.Normal and anomalous data are recorded under the corresponding changed conditions.
3. DETAILS OF SUB-DATASETS
The sub-datasets provide detailed toy-car and toy-train recordings with multiple configurations, operating speeds, microphones, anomaly types, and damage depths. Tables document anomaly and variation settings, while example configurations support domain-shift evaluation.
- 3. DETAILS OF SUB-DATASETS: Toy-car recordings use five microphones around the inspection device across five machine configurations, while toy-train recordings use eight microphones inside and outside the track perimeter.The toy-train setup assigns dynamic microphones to outer channels 1–4 and condenser microphones to inner channels 5–8.
- 3.1. Toy-car sub-dataset: The toy-car sub-dataset contains over 177 k sound samples across five microphone channels, including over 8 k anomalous samples.Samples are 12 seconds long, use five voltage-controlled speeds, and include three damage-depth levels for shafts, gears, and tires.
- 3. DETAILS OF SUB-DATASETS: Both sub-datasets include environmental-noise recordings reproduced through four loudspeakers with the same microphone settings used for machine sounds.Noise sources combine newly recorded files with files from ToyADMOS.
- 3. DETAILS OF SUB-DATASETS: Tables 1 and 2 document anomaly conditions and variation settings, while Table 3 provides example domain-shift task configurations.Further dataset details are available through the ToyADMOS2 repository.
4. SAMPLE DOMAIN-SHIFT TASK SETTINGS AND BENCHMARK
The benchmark illustrates ToyADMOS2 evaluation under domain shifts using limited target-domain normal data and varying noise conditions. Results show that anomaly-detection difficulty depends on task, domain, and damage level.
- The example evaluation used 1,500 source-domain normal samples, four normal target-domain samples per domain, and 200 unknown test samples.
- Target domains represented model-and-parts, operating-speed, microphone-and-noise, and combined shifts.
- Samples were evaluated at clean, 6 dB, 0 dB, and -6 dB SNR after mixing machine sounds with environmental noise and down-sampling to 16 kHz.
- A simple unsupervised autoencoder baseline was trained separately for each SNR condition using merged source- and target-domain normal data.
- Around 0.99 AUC was obtained for Toy car across all damage levels, while higher damage made source-domain/clean Toy-train detection easier.These are benchmark results for the DCASE 2020 baseline under the reported domain-shift configurations.
5. CONCLUSIONS
ToyADMOS2 is a large-scale dataset for evaluating anomalous-sound detection under domain shifts. It covers fixed-task and moving-task machine inspection with varied operating conditions and substantial normal and anomalous sound collections.
- ToyADMOS2 contains fixed-task and moving-task sub-datasets with domain shifts in machine models, part configurations, operating speeds, and microphone arrangements.
- Each sub-dataset contains over 27 k normal and over 8 k anomalous machine-operating sounds recorded at a 48-kHz sampling rate.
- The dataset was designed for evaluating systems under domain-shift conditions and was released for free download.