Acoustic Feature Representation Based on Timbre for Fault Detection of Rotary Machines

A new acoustic feature representation is proposed to detect faults of rotary machinery using acoustic signals. Acoustic features include the amplitude, frequency, and timbre, with the former two often used as diagnostics features. The timbre is also an indicator of the abnormal operation of machinery. The present study therefore focuses on the use of timbre-based features. Changes in stationary parts of observed acoustic signals (e.g., a change in period, which corresponds to the number of rotations) can be considered a physical quantity of the timbre. Because a rotary machine normally operates at a certain rotational speed, differences in the rotational period between adjacent frames are ideally zero in such normal operation. In contrast, the period differences should not be zero for a machine with a fault because observed signals contain characteristics of both the normal and faulty states and these characteristics vary over time. The present study therefore attempts to exploit the period difference of acoustic signals as a feature representation for fault detection. Experimental analysis conducted using acoustic signals recorded by microphones demonstrates that the proposed feature extraction contributes to the fault detection of rotary machines.