Augmented mel-spectrogram VGG-16 model for axial and radial load classification at wire-race bearings

Abstract

Abstract Wire-race bearings are designed to be space-saving and are used in settings, where high precision under continuous high load and torque is required. Thus, load detection is a vital part of their condition monitoring. Measuring load using load cells defeats the space-saving purpose of these bearings. Using virtual sensors, by measuring axial and radial loading indirectly, via an accelerometer attached to the housing of the bearing, can provide an alternative. We extend an earlier approach of a virtual sensor using vibration spectrograms with convolutional neural networks. We propose to use transfer learning to reduce overfitting. Moreover we use Mel-spectrograms instead of standard spectrograms and apply feature augmentation techniques. Thereby the virtual sensor improves validation accuracy from 86.3 % to 94.2 %. Using the logarithmic Mel-scale enables to over-sample the characteristic rollover frequencies which appear up to 500 Hz. Frequency masking shows a higher degree of effect compared to normal spectrograms. Further, we find that by using Mel-spectrograms, the combined latency of the spectrogram generation and forward propagation is reduced by 61.6 %.