“Cough for me…” In a routine physical checkup, a doctor may press a stethoscope against a patient’s chest and use the sound of a cough to detect any abnormal respiratory conditions. Now, an R&D team from Japanese multinational conglomerate Hitachi has proposed a similar, sound-based technique for identifying malfunctioning industrial machines on a factory floor.
Hitachi researchers have released a first-of-its-kind dataset comprising sounds of functioning and malfunctioning industrial machines in real factory environments. The dataset is designed to aid in the development of acoustic detection techniques, and is introduced in the paper MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection.
The MIMII dataset classifies sounds into four major machine types: valves, pumps, fans, and slide rails, with samples from different machine models included in each category. There are a total of 26,092 ten-second sound files of machines in normal operating conditions, and 6,065 sound files of machines struggling under abnormal conditions. Researchers used only the normal machine operation sounds during the training phase, then introduced the anomalous sounds during the testing phase.
Researchers positioned a circular array of eight microphones at distances of 10 – 50cm to record the machine sounds in various real factories, processing the 10-second recordings as 16-bit audio signals sampled at 16 kHz. Because factories are not libraries, researchers had to deal with plenty of background noise, which they isolated and later mixed with the target machine sound to simulate real world conditions. The sound files are stored in the standard .WAV audio format.
Because every machine has its particular sound characteristics — even within a class or model type — researchers employed an autoencoder-based unsupervised anomaly detector. Researchers noted for example the sound signals of valves are “impulsive and sparse in time” which makes valve anomaly detection challenging compared to the relatively static and continuous sound signals of fans, which can more quickly and effectively expose anomalies. Researchers propose reducing detection degradation caused by non-static machines and factory background noise as areas for further study in unsupervised anomalous sound detection.
The MIMII dataset of machine sounds and factory background noise is the first to provide solutions for detecting anomalous conditions in industrial machinery via sounds, and has been open-sourced on Zenodo.
The paper MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection is on arXiv.
Journalist: Fangyu Cai | Editor: Michael Sarazen