MC-ABDS: A system for low SNR fault diagnosis in industrial production with intense overlapping and interference

Abstract

Planetary gearboxes are vital in industrial production due to their large transmission ratios. Therefore, accurate fault diagnosis of planetary gearboxes is crucial. However, in industrial applications, the acoustic fault signals from two different adjacent planetary gearboxes may overlap and interfere with each other, resulting in a low Signal-to-Noise Ratio (SNR) for each acoustic fault source, which in turn prevents accurate fault diagnosis. In this context, the Multi-Task Learning-Temporal Convolutional Network (MTL-TCN) is proposed to simultaneously output the orientation of the acoustic sources as well as the fault type to solve the problem of interference between adjacent acoustic sources. A Spatial information based Multi-Task Channel Attention (SMTCA) mechanism is also proposed to solve the problem of acoustic signal overlapping by using the orientation information to calculate the weight of the acoustic signal channel and assigning it to the fault diagnostic task, which combines the sound field information into the separation of fault sources. Finally, a Multi-Channel Acoustic based diagnose System (MC-ABDS) is proposed, which contains a customized microphone array as well as a sound field information and fault feature information extraction method called Multi-Task Attention TCN (MTA-TCN). The system is validated by the data collected in the anechoic chamber, and it is effective for the acoustic overlapping and interference that occurs when two adjacent planetary gearboxes are operating. The orientation accuracy of the acoustic source reached 99.98 %and the diagnostic accuracy of the fault reached 92.08 %.