Feature extraction for rotary machine acoustic diagnostics focused on periodic period

A new feature extraction method is proposed to detect abnormality of rotary machines using sounds. It is important to detect the abnormality at an early stage to efficiently maintain industrial machines. The rotary machines generally yield abnormal sounds in operation with a technical failure. Acoustic sensors, i.e., microphones, have an advantage in diagnostic that they can avoid a direct contact to the machines. For employing acoustic information, however, the acoustic feature suitable for abnormality detection has to be investigated because it is difficult to extract a periodic period originated number of rotations, especially low frequency due to signal-to-noise ratio detection range is low, based on Fourier transform. In the present study, we attempt to estimate the rotational period based on peak selection method for the feature parameter. Experimental investigation carried out by simulating motor failures demonstrates that the period estimate is the feature suitable for abnormal sound detection: The rotational periods is estimated when the normal operation sound period is electromagnetic frequency and abnormal one is another period. It varies in the histogram by more than 20% and the outliers that were not detected in the normal operation mode are observed.