Multirotational Speed Data Augmentation and Data Repair of High-Speed Train Wheelset Bearings Using Graph Speed Classifier GAN

Wheelset bearings fault samples of high-speed trains often suffer from insufficient numbers and missing points. However, working conditions are one of the most important factors affecting data augmentation and repair, which is not sufficiently emphasized by existing methods. Also, generative adversarial networks (GANs) as effective sample generation methods still have many problems, such as model collapse and training difficulties. To solve these problems, a data augmentation and data repair method for high-speed train wheelset bearing fault diagnosis with multispeed generation capability is proposed. By adding an independent speed classifier, the speed conditions of samples are classified. The relation among samples is also introduced to improve the accuracy of the classifier. A new loss function is designed to guide to generate samples of specified faults and speeds. Finally, a sample repair method based on the pretraining generator is studied to improve the utilization rate of real samples. The wheelset bearing dataset is utilized to validate that the proposed method can generate and repair samples for different speeds and fault types.