Multifault Feature Wasserstein Generative Adversarial Networks for Fault Diagnosis in Unbalanced Data

Abstract

Due to the limitation of industrial conditions in production, raw sensor data are always shown as an unbalanced dataset, characterized by abundant normal operational data and scarce fault instances. This unbalance can degrade the performance of conventional fault diagnosis methods, leading to reduced accuracy and unstable model training. To address this challenge in bearing fault diagnosis, this article proposes a multifault feature Wasserstein generative adversarial network (MFF-WGAN) to enhance diagnostic precision. First, the framework employs a multiencoder denoising autoencoder (DAE) architecture to mitigate noise interference in raw sensor data. Subsequently, the proposed MFF-WGAN integrates label information into its adversarial loss function to enable simultaneous generation of diverse fault categories, while incorporating interclass feature discrepancies to refine sample quality. Finally, the developed multifault feature Wasserstein generation adversarial network is tested on the Case Western Reserve University bearing dataset and the laboratory bearing dataset. Computational results show that the proposed method can generate high-quality bearing samples with multiple faults effectively, which can obtain a higher diagnosis accuracy of 99.01% and 97.71% compared with the existing methods.