Feature Selection for Audio-Based Fault Detection in Pumps
Abstract Whilst extensive work has been done on fault detection in bearings using sound, very little has been accomplished with other machine components and machinery partly due to the scarcity of datasets. The recent release of the Malfunctioning Industrial Machine Investigation and Inspection (MIMII) dataset opens the opportunity for research into malfunctioning machines like pumps, fans, slide rails, and valves. In this paper, we compare common features from audio recordings to investigate which best support the classification of malfunctioning pumps. We evaluate our results using the Area Under the Curve (AUC) as a performance metric and determine that the log mel spectrum is a very useful feature, at least for this dataset, but that other features can enhance detection performance when ambient noise is present (improving AUC from 0.88 to 0.94 in one case). Also, we find that mel Frequency Cepstral Coefficients (MFCC) perform substantially poorer as features than a sampled mel spectrogram.