6.22.21
Industrial Music

An automatic aleatoric composition created from the sounds of the MIMII dataset

The Malfunctioning Industrial Machine Investigation and Inspection (MIMII) dataset is composed of over 30,000 recordings of factory sounds.  Created by the Research and Development Group of Hitachi, Ltd., this dataset is the first of its kind.  The recordings include “normal” and “abnormal” sounds of industrial machines recorded on factory floors and are intended to provide a benchmark dataset for the development of automated facility maintenance through machine learning.  By training an artificial intelligence in sound-based machine fault diagnosis and anomalous sound detection it could be possible to diagnose mechanical faults and perhaps even apply predictive maintenance models to industrial machinery using computer monitored microphones.

As interesting as that is, the MIMII dataset provides a different opportunity, namely to explore the largest repository of factory field recordings ever created.   In keeping with the purpose of its creation- to teach machines how to monitor machines- this edition of Special Collections is completely automated.  A 6 channel computer jukebox will crawl through the sound files, selecting files to be played at random and for varying durations.

The original paper, MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection, can be read at https://arxiv.org/pdf/1909.09347.pdf 

The MIMII dataset can be downloaded at https://zenodo.org/record/3384388#.YNJF6m5lBUM