Source-linked AI summary
ToyADMOS: A Dataset of Miniature-Machine Operating Sounds for Anomalous Sound Detection
Yuma Koizumi, Shoichiro Saito, Hisashi Uematsu, Noboru Harada, Keisuke Imoto
TL;DR
Large-scale ADMOS datasets are lacking because anomalous machine sounds are difficult to collect, limiting fair evaluation. ToyADMOS addresses this gap by deliberately damaging miniature machines to create a freely available dataset, with baseline results across three sub-datasets and a stated toy-to-real transfer limitation.
Problem
Large-scale datasets are lacking for anomaly detection in machine operating sounds, while anomalous sounds are difficult to collect and fair evaluation requires freely available data.
Method
ToyADMOS collects normal and anomalous operating sounds from deliberately damaged miniature machines under controlled laboratory recording conditions.
Results
The baseline achieved area under the receiver operating characteristic curves of 0.874, 0.981, and 0.843 for the toy-car, toy-conveyor, and toy-train sub-datasets, respectively.
Takeaways & Limitations
ToyADMOS and tutorial Python codes are freely available to support research into anomaly detection in machine operating sounds.
Takeaways & Limitations
Toy sounds do not necessarily match real machine sounds, so preprocessing parameters evaluated on ToyADMOS may need reconsideration for real-world ADMOS systems.
Abstract
from arXiv · showhide
This paper introduces a new dataset called "ToyADMOS" designed for anomaly detection in machine operating sounds (ADMOS). To the best our knowledge, no large-scale datasets are available for ADMOS, although large-scale datasets have contributed to recent advancements in acoustic signal processing. This is because anomalous sound data are difficult to collect. To build a large-scale dataset for ADMOS, we collected anomalous operating sounds of miniature machines (toys) by deliberately damaging them. The released dataset consists of three sub-datasets for machine-condition inspection, fault diagnosis of machines with geometrically fixed tasks, and fault diagnosis of machines with moving tasks. Each sub-dataset includes over 180 hours of normal machine-operating sounds and over 4,000 samples of anomalous sounds collected with four microphones at a 48-kHz sampling rate. The dataset is freely available for download at https://github.com/YumaKoizumi/ToyADMOS-dataset
1. INTRODUCTION
ToyADMOS addresses the lack of freely available large-scale ADMOS datasets, whose development is hindered by the difficulty of collecting anomalous sounds. It introduces a controlled, freely downloadable dataset built from deliberately damaged miniature machines.
- Motivation: Large-scale datasets support training and fair evaluation of deep neural networks in acoustic signal-processing tasks.Examples include WSJ0, VCTK, AudioSet, and Freesound.
- Research gap: No freely available large-scale datasets existed for anomaly detection in sounds because anomalous sounds are difficult to collect.Machine-condition inspection and fault-diagnosis tasks lacked even small datasets, encouraging synthetic-data evaluation.
- Contribution: ToyADMOS collects normal and anomalous operating sounds from deliberately damaged miniature machines under controllable laboratory recording conditions.The dataset targets both basic unsupervised ADMOS and advanced tasks including domain adaptation, noise reduction, data augmentation, and few-shot learning.
- Dataset design: The dataset covers product inspection, fixed-machine fault diagnosis, and moving-machine fault diagnosis, with individually recorded environmental noise and four microphones.Multiple machines with different detailed structures support variation across individuals, while repeated anomalous recordings support few-shot learning.
- Availability: ToyADMOS and related use materials are freely available for download.The dataset webpage provides the license and usage information.
2. DATASET OVERVIEW
ToyADMOS contains three task-specific sub-datasets with normal, anomalous, and environmental recordings collected under controlled conditions. Its design supports noise variation, multi-channel processing, individual-machine variation, and analysis of misdetections, while toy-to-real transfer remains limited.
- Sub-datasets: ToyADMOS comprises toy-car product inspection, toy-conveyor fixed-machine diagnosis, and toy-train moving-machine diagnosis sub-datasets.Each task uses a different toy and corresponding recording arrangement.
- Variation and controlled conditions: Multiple same-class toy models with different detailed structures provide operating-sound variation across individual machines.The dataset identifies each machine using the term “case,” supporting evaluation of domain adaptation to individual and noise-level changes.
- Sound categories: Each sub-dataset contains normal, anomalous, and environmental sound data, with anomalies produced by damaging components or adding extraneous objects.Environmental recordings simulate factory conditions using sounds collected at several actual factory locations and emitted from four loudspeakers.
- Recording format: Four omnidirectional microphones record the sounds, while IND files capture complete operations and CNT files capture continuous recordings cut every 10 minutes.Normal sounds include both IND and CNT files; anomalous sounds use IND files, and environmental noise uses CNT files.
- Variation and controlled conditions: Controlled recording conditions help analyze performance and misdetections because normal sounds share the same recording condition.This is especially relevant for unsupervised ADMOS, where rare normal sounds and training/test recording differences can cause misdetection.
- Limitation: Toy sounds do not necessarily match real machine sounds, so preprocessing parameters such as filterbanks must be reconsidered for real-world ADMOS systems.Toy and real sounds may share time-frequency structure while differing in detailed spectral shape because machine size influences sound.
3. DETAILS OF SUB-DATASETS
The three ToyADMOS sub-datasets instantiate product inspection and fixed- or moving-machine diagnosis with distinct miniature-machine configurations, recording layouts, sample formats, and deliberately induced faults.
- Toy-car sub-dataset: The toy-car sub-dataset models product inspection using a mini 4WD whose running sounds are recorded on an inspection device.Four cases combine two motor types and two bearing types; IND files are 11 seconds long.
- Toy-conveyor sub-dataset: The toy-conveyor sub-dataset models diagnosis of a fixed machine using three conveyor cases with microphones on the conveyor body and desk.Each case contains 60 hours of IND normal sounds and 120 hours of CNT normal sounds, with faults induced in pulleys, the belt, and voltage conditions.
- Toy-train sub-dataset: The toy-train sub-dataset models diagnosis of a moving machine using four cases combining train types and railway scales.The four microphones surround the track, and anomalous sounds arise from damaged carriages and straight or curved railway tracks.
- Toy-train sub-dataset: Toy-train recordings contain 66 hours of IND normal sounds and 197 hours of CNT normal sounds.Each case and channel includes 1,350 IND samples and 74 CNT samples.
- Environmental noise: The dataset includes environmental noise recordings for the toy-train configurations, totaling 12 hours for each case.The recording amount reflects differing microphone positions between HO-scale and N-scale cases.
4. EVALUATION AND BENCHMARK
A simple unsupervised autoencoder baseline was evaluated on each ToyADMOS sub-dataset under mixed-noise conditions. Performance varied across tasks, and false negatives exposed anomalous sounds with limited spectral or amplitude changes.
- Experimental setup: The benchmark evaluates a simple unsupervised anomaly-detection system using case 1, with 1,000 IND normal samples for training.Evaluation uses remaining normal and anomalous IND samples, mixed with randomly cropped environmental noise and downsampled to 16 kHz.
- Results: 0.874, 0.981, and 0.843 area under the receiver operating characteristic curves were obtained for toy-car, toy-conveyor, and toy-train sub-datasets, respectively.A file is classified as anomalous when the anomaly score exceeds the threshold for at least one time frame.
- Error analysis: False negatives frequently involved toy-car over-voltage sounds and toy-train curved-track sounds.Over-voltage sounds had little amplitude change, while the curved-track damage was far from all four microphones and produced small amplitude.
- Error analysis: The error analysis identifies detection of anomalies with little time-frequency change from normal sounds as a research direction.This direction follows the observed failures of the simple baseline on over-voltage and curved-track anomalies.
5. CONCLUSIONS
ToyADMOS is a freely available dataset for anomaly detection in machine operating sounds, accompanied by tutorial Python code. The paper reports its introduction and acknowledges technical assistance in data collection.
- ToyADMOS was introduced for anomaly detection in machine operating sounds (ADMOS).
- The dataset and tutorial Python codes are freely available on the Web.
- The authors hope ToyADMOS will advance research into anomaly detection in sounds.