Development of Data Augmentation Strategies for Rolling Element Bearings

Machine learning algorithms are used in condition monitoring of rolling bearings to classify damages based on structure-borne sound signals.
However, since it is not efficient to collect damaged bearings or to damage them artificially, real data sets almost exclusively contain data of undamaged bearings, while the damage classes are rarely represented.
This imbalance usually leads to overfitting of the model to the class of undamaged bearings.
To address this problem, data augmentation methods can be used to generate artificial data based on existing time series and, thus, to augment underrepresented classes.
In this work, promising basic data augmentation techniques for time series of rolling bearings have been selected and applied to the dataset of Paderborn University.
For each method, the effects on the envelope spectrum, which represents the model input, were checked to ensure that the cyclostationary properties of the time series and the characteristics of the envelope spectra are preserved.
The obtained data sets were used to train a convolutional neural network.
Based on the classification results, promising augmentation approaches were identified.
In future, they could be incorporated into generative models such as GANs and thereby reduce their computational effort when generating artificial data.