A high-speed train gearbox imbalance fault diagnosis method based on virtual reality data generation fusion with enhanced randomness

Abstract

The collection of historical fault data for high-speed train gearbox gears requires long observation periods, and the occurrence probability of each fault type varies significantly, leading to severely class-imbalanced datasets that greatly limit the practical applicability of fault diagnosis models. Although traditional dynamic simulation methods can be used for data augmentation, they fail to adequately capture the complex disturbance factors present in real operating conditions, thereby limiting their effectiveness in improving diagnostic accuracy. To address these issues, this article proposes a novel fault diagnosis framework for high-speed train gearbox imbalance faults (VRGF-DM), which integrates randomness enhancement factors with a virtual–physical generative fusion network. First, a structure-coupled dynamic model of a high-speed train transmission gearbox is established based on a multibody dynamics modeling approach, enabling accurate representation of typical fault conditions, faithful reflection of vibration mechanisms in the physical domain, and generation of large amounts of vibration data under different gear fault modes. Second, an interactive generative adversarial network is designed for virtual–physical data fusion and combined with a random-loss variational autoencoder to improve the quality of simulated data while fully exploiting latent fault-related information. Finally, the fault diagnosis model is trained and tested using the augmented and completed dataset. Experimental results demonstrate that the Proposed method improves diagnostic accuracy by an average of 1.6% under relatively low data imbalance conditions and by 4.45% under highly imbalanced conditions.