Open Access
Subscription or Fee Access
Learning Condition Indicators for Rolling Element Bearings via Deep Variational Autoencoding
Abstract
Feature extraction from time-series data is required for condition monitoring of dynamic mechanical systems. However, conventional techniques based on the knowledge of a human expert are not reliable for large and multidimensional datasets. The consumption of grease life in rotorcraft bearings has been studied and grease life models factor in to the determination of airworthiness and maintenance requirements. Features of acoustic emission (AE) signals have demonstrated potentially higher sensitivity to lubrication condition due to increased AE generation through asperity contact, than can be measured in structural vibration signals. The objective of this effort is, therefore, to utilize deep representation learning to automatically extract the most relevant features associated with lubrication condition from time-series data and assess their capacity as condition indicators of bearing grease life. AE and vibration data are collected from an accelerated grease consumption test on an otherwise healthy bearing with reduced grease. The time-series data is transformed into spectrograms and unsupervised learning is performed on these images using a convolutional variational autoencoder. The encodings of these spectrograms in the latent space of the autoencoder and their correlation with experiment time as an estimate of grease consumption were examined, where it was observed in both cases, via dimensionality reduction and by probing the generative model, that though the trends differ, the latent spaces of both AE and vibration data models are highly structured based on experiment time. However, the encodings of AE signals present a clearer and more descriptive trend when compared with vibration signal encodings.
DOI
10.12783/shm2019/32493
10.12783/shm2019/32493