Source-linked AI summary
MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection
Harsh Purohit, Ryo Tanabe, Kenji Ichige, Takashi Endo, Yuki Nikaido, Kaori Suefusa, Yohei Kawaguchi
TL;DR
Industrial anomaly detection lacks public datasets of machine sounds recorded under normal and anomalous conditions in real factories. The paper introduces MIMII, a multitype machine-sound dataset with real-life anomalies and an autoencoder benchmark for unsupervised detection. The evaluation identifies non-stationary sounds and noise as key issues for anomaly detection.
Problem
Public datasets do not cover different industrial machine sounds under normal and anomalous operating conditions in real factory environments, despite the need for automated inspection.
Method
The paper records machine sounds from four machine types with normal and anomalous conditions, simulates factory environments with background noise, and evaluates an autoencoder using normal-only training.
Results
The evaluation finds that non-stationary machine sounds and noise are key issues for developing unsupervised anomalous sound detection.
Takeaways & Limitations
MIMII provides a freely available benchmark for industrial machine-sound anomaly detection and related research.
Takeaways & Limitations
The benchmark assumes that a trained autoencoder produces high reconstruction error for anomalous machine sounds.
Abstract
from arXiv · showhide
Factory machinery is prone to failure or breakdown, resulting in significant expenses for companies. Hence, there is a rising interest in machine monitoring using different sensors including microphones. In the scientific community, the emergence of public datasets has led to advancements in acoustic detection and classification of scenes and events, but there are no public datasets that focus on the sound of industrial machines under normal and anomalous operating conditions in real factory environments. In this paper, we present a new dataset of industrial machine sounds that we call a sound dataset for malfunctioning industrial machine investigation and inspection (MIMII dataset). Normal sounds were recorded for different types of industrial machines (i.e., valves, pumps, fans, and slide rails), and to resemble a real-life scenario, various anomalous sounds were recorded (e.g., contamination, leakage, rotating unbalance, and rail damage). The purpose of releasing the MIMII dataset is to assist the machine-learning and signal-processing community with their development of automated facility maintenance. The MIMII dataset is freely available for download at: https://zenodo.org/record/3384388
1. INTRODUCTION
Factory inspection increasingly needs affordable automation because expert-based malfunction discovery is constrained by rising demand and a shortage of field engineers. The paper addresses the absence of public real-factory machine-sound datasets by introducing MIMII for unsupervised anomaly detection.
- Industrial inspection needs automation because malfunction discovery relies heavily on field-engineer experience while inspection requests increase and experts are scarce.
- Existing acoustic datasets advance scene and event recognition, but no public dataset covers varied machine sounds under normal and anomalous conditions in real factories.
- MIMII contains sounds from valves, pumps, fans, and slide rails, with seven product models for each machine type.
- The main task is unsupervised detection of anomalous conditions in 10-second segments, training only on normal sounds and testing normal-versus-abnormal discrimination.
- The dataset includes 26,092 normal sound files and real-life anomalous sounds, and is freely released for research on automated facility maintenance.
2. RECORDING ENVIRONMENT AND SETUP
MIMII recordings use an eight-microphone circular array positioned near each machine, capturing multichannel 10-second audio in reverberant factory conditions. Background noise from multiple real factories is also mixed with target sounds to simulate real environments.
- The TAMAGO-03 circular array contains eight microphones and supports evaluation of both single-channel and multichannel approaches.
- The array was placed 50 cm from machines, or 10 cm for valves, and recorded eight separate channels for each 10-second segment.
- Machine sounds were recorded as 16-bit audio sampled at 16 kHz in a reverberant environment, with each machine recorded in a separate session.
- Background noise from multiple real factories was continuously recorded and later mixed with target machine sounds to simulate real environments.
3. DATASET CONTENT
The dataset covers four industrial machine types and seven individual machines per type, combining extensive normal recordings with varied real-life anomalous conditions. Factory background noise is mixed into machine signals at controlled signal-to-noise ratios.
- MIMII includes solenoid valves, water pumps, industrial fans, and linear slide systems, spanning stationary and non-stationary machine sounds.
- Each machine type includes seven individual machines, and the dataset contains 26,092 normal sound segments across them.
- Recorded anomalous scenarios include contamination, leakage, rotating unbalance, and rail damage, with relatively few anomalous segments for each machine type.
- For a target SNR γ, background-noise power bj is tuned using γ = 10 log10 (a/bj) before adding the target machine and noise segments.
4. EXPERIMENT
The experiment benchmarks an autoencoder for unsupervised anomaly detection using log-Mel features and normal-only training. Performance varies by machine sound characteristics and factory-noise level, revealing non-stationarity and noise as key challenges.
- The benchmark evaluates an autoencoder-based unsupervised anomaly detector, motivated by prior successful use of autoencoders for unsupervised anomaly detection.
- The model uses first-channel audio, log-Mel spectrograms with 64 mel filters, and five combined frames forming a 320-dimensional input vector.
- The autoencoder is trained on normal segments, while anomalous segments and an equal number of normal segments form the test set; anomaly scores use ten-second averaged reconstruction error.
- AUCs are lower for valves than other machines because impulsive, sparse, non-stationary sounds yield small time-averaged reconstruction errors, whereas stationary fan sounds are easier to detect.
- For some machine models, AUC decreases rapidly as factory-noise levels increase, identifying non-stationarity and noise as challenges for unsupervised detection.
5. CONCLUSION AND FUTURE DIRECTIONS
The paper introduces MIMII as a real-world dataset for industrial-machine malfunction investigation, evaluates autoencoder-based unsupervised detection, and identifies non-stationarity and noise as key challenges. The dataset is intended as a benchmark for improving anomaly detection and supporting broader research, including domain adaptation.
- Conclusion: 26,092 normal and 6,065 anomalous sound segments were collected for investigating malfunctioning industrial machines.Background noise from multiple real factories was mixed with machine sounds to simulate real environments.
- Conclusion: Autoencoder-based unsupervised anomalous sound detection was evaluated using the MIMII dataset.The evaluation uses the dataset as a basis for assessing anomaly-detection performance.
- Conclusion: Non-stationary machine sounds and noise were identified as key issues for unsupervised anomaly detection.These issues are highlighted as challenges to overcome in developing anomaly detectors.
- Future directions: The dataset provides a benchmark for improving anomaly-detection accuracy in MIMII.The authors position the reported results as a basis for future method development.
- Future directions: MIMII can support evaluation of domain adaptation by training on selected machine models and testing on models from other manufacturers.The paper also describes the dataset as freely available and intended to accelerate research on machine-sound event detection.