[Paper Review] MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection

Harsh Purohit, Ryo Tanabe+5 more|arXiv (Cornell University)|Sep 20, 2019

Music and Audio Processing24 references42 citations

TL;DR

Introduces the MIMII dataset for anomalous sound detection in four industrial machine types, and provides an autoencoder-based benchmarking baseline in unsupervised anomaly detection under real-world noisy conditions.

ABSTRACT

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

Motivation & Objective

• Create an open, real-world dataset of industrial machine sounds under normal and anomalous conditions to advance acoustic anomaly detection in maintenance contexts.
• Enable unsupervised learning benchmarks by training on normal sounds only and testing on mixed normal/anomalous sounds.
• Provide baseline autoencoder-based anomaly detection results across multiple machine types and noise levels to guide future research.

Proposed method

• Record multi-channel (eight-m microphone array) sounds from four machine types (valves, pumps, fans, slide rails) in real factory environments.
• Mix target machine sounds with background factory noise to simulate real environments at various SNR levels (6 dB, 0 dB, -6 dB).
• Extract log-Mel spectrogram features (frame size 1024, hop 512, 64 mel filters; 5 frames -> 320-d input) and train per-machine autoencoders to reconstruct normal sounds.
• Use an encoder–decoder network with fully connected layers (encoder: 64-64-8; decoder: 8-64-Output) and ReLU activations; optimize with Adam for 50 epochs.
• Evaluate anomaly detection via reconstruction error thresholding over 10-second segments; report AUC averaged over three runs.

Experimental results

Research questions

• RQ1 Can unsupervised autoencoders trained only on normal machine sounds detect anomalous operating conditions across different machine types?
• RQ2 How does background factory noise and non-stationarity of machine sounds affect unsupervised anomalous sound detection performance?
• RQ3 What are the baseline detection capabilities (AUC) for valves, pumps, fans, and slide rails under varying SNR conditions?
• RQ4 Is the dataset suitable for evaluating domain adaptation or cross-model generalization in anomaly detection?
• RQ5 How do multi-channel recordings compare to single-channel approaches for anomaly detection in industrial soundscapes?

Key findings

• The dataset contains 26,092 normal sound segments and 6,065 anomalous segments across four machine types and seven models each.
• Autoencoder-based unsupervised anomaly detection yields varying AUCs by machine type and model, with higher performance for fans and lower for valves.
• Average AUCs decrease with lower SNR, indicating noise degrades anomaly detection performance; valves show particularly lower AUC due to non-stationary, impulsive sounds.
• Across machine types, pumps and fans generally achieve higher AUCs than valves and slide rails, illustrating differences in acoustic characteristics for detection.
• Non-stationarity and background noise are key challenges for unsupervised anomalous sound detection in industrial environments.