Comparison of Conditional and Non-Conditional Data Augmentation Approaches with Generative Adversarial Networks: A Case Study on Bearing Fault Diagnosis

In order to perform fault classification using Machine Learning algorithms, sufficient and balanced data is required. Nevertheless, in a lot of use cases data is not available in a sufficient manner to allow a stable and valid training of Machine Learning based algorithms for fault classification. There are various methods for augmenting real measured data, whereby a Generative Adversarial Network (GAN) is one of the most suitable approaches for synthetic data generation. Conditional and non-conditional data augmentation approaches are available for implementation of GAN algorithms for synthetic data generation. In the scope of the proposed paper, the performance of a conditional GAN (cGAN) and traditional GAN is evaluated and compared using vibration data of a rolling bearing measured on a bearing test rig. Both GAN approaches are used with optimized losses, considering the Wasserstein Distance and a gradient penalty in the loss function. The loss function of a cGAN contains label information, enabling the model to be trained as a holistic system for all labels, rather than being trained per label as in a traditional GAN. The training of the both GAN networks show a high performance. The synthetically generated data compared to the real data also shows sufficient similarity. Nevertheless, the distribution of the generated data is different, as the data generated with a non-conditional GAN has a better quality of results. Different methods are used to evaluate the data, for example statistical or cluster analysis, highlighting the differences in the generated data.