Predicting Remaining Useful Life During the Healthy Stage in Rolling Bearings

Quantitative predictions of the time before the spall initiation phase (origination of the first spall) in pristine ball bearings running under an applied load is of great industrial relevance, especially for systems that require high running accuracy and/or high-speed performance. Currently there are no available methodologies to predict the remaining life until the first spalling event exclusively from vibration signals. We present an end-to-end approach, based on deep learning (one dimensional convolutional layers combined with long short-term memory units), that is able to quantify the time before the origination of the first spall in ball bearings, having as sole input vibration measurements. The method has been validated on a set of bearings -- run to failure on independent but identical test-rigs -- which had not been considered during training.