Rolling bearing fault diagnosis under limited data using conditional mask-enhanced VQGAN and convolutional-embedded vision Mamba

Abstract The rapid advancement of rolling bearing fault diagnosis technology has provided crucial support for the operation and maintenance of high-end equipment. However, the high cost of data acquisition and stringent safety requirements in industrial environments have led to scarce samples and uneven category distribution, thereby limiting the performance improvement of diagnostic models. In view of this, this study proposes a novel fault diagnosis method integrating a conditional mask-enhanced vector quantized generative adversarial network (CME-VQGAN) with a convolutional-embedded vision Mamba (ViM) network. To improve the similarity between generated pseudo-samples and real data, a conditional autoregressive representation modeling (CARM) mechanism is proposed. It extracts latent indices from original time–frequency images via an encoder and applies masking to emphasize common fault features as initial conditions. Based on the dependency relationships of key features, CARM drives the transformer to model the global dependencies of latent representations, thereby generating high-quality pseudo-samples. Furthermore, to address characteristic frequency band aliasing caused by compound faults, a global context modeling module is introduced in both the encoder and decoder of the CME-VQGAN, enhancing the representation of multi-source coupled fault features. Additionally, a multi-scale convolutional token embedding module in the ViM embedding layer strengthens attention to key local structural features, improving the discernibility of aliased fault characteristics arising from local non-stationarity, modulation-impulse mixtures, and phase perturbations. To verify the effectiveness and superiority of the proposed method, a series of comparative experiments were conducted on the Case Western Reserve University dataset and a self-constructed dataset. The experimental results demonstrate that the proposed method can effectively address the challenges of few-shot and imbalanced data, highlighting its practical value in real-world applications.