In response to the challenges posed by complex and variable operating conditions of rolling bearings and the limited availability of labeled data, both of which hinder the effective extraction of key fault features and reduce diagnostic accuracy, this study introduces a model that combines a spatial attention (SA) mechanism with a multi-scale depthwise separable convolution module. The proposed approach first employs the Gramian angular difference field (GADF) to convert raw signals. This conversion maps the temporal characteristics of the signal into an image format that intrinsically preserves both temporal dynamics and phase relationships. Subsequently, the model architecture incorporates a spatial attention mechanism and a multi-scale depthwise separable convolutional module. Guided by the attention mechanism to concentrate on discriminative feature regions and to suppress noise, the convolutional component efficiently extracts hierarchical features in parallel through the multi-scale receptive fields. Furthermore, the trained model serves as a pre-trained network and is transferred to novel variable-condition environments to enhance diagnostic accuracy in few-shot scenarios. The effectiveness of the proposed model was evaluated using bearing datasets and field-collected industrial data. Experimental results confirm that the proposed model offers outstanding fault recognition performance and generalization capability across diverse working conditions, small-sample scenarios, and real industrial environments.

Bearings in mining scraper conveyors, essential transmission components, are frequently subjected to harsh environments, including high temperatures, heavy impact loads, dust contamination, and prolonged operational durations. Vibration signals from these bearings face acquisition challenges, compounded by a significant imbalance between normal and fault data, which complicates fault diagnosis. Conventional fault diagnosis methods often fail to adequately resolve these issues. To address these challenges, this study introduces a novel bearing fault diagnosis approach that combines an Improved Auxiliary Classifier Generative Adversarial Network (IACGAN) with a Swin Transformer (ST). Initially, continuous wavelet transform (CWT) is employed to convert vibration signals into two-dimensional time-frequency representations. Subsequently, a Convolution and Attention Fusion Module (CAFM) is incorporated into the ACGAN generator to enhance the quality and diversity of generated samples. Moreover, a gradient penalty term is added to the discriminator’s loss function to improve training stability and mitigate mode collapse. Finally, the Swin Transformer is trained on a hybrid dataset comprising high-quality time-frequency images produced by IACGAN and original samples. Experimental results show that the Swin Transformer achieves superior diagnostic accuracy under various imbalance conditions, reaching 99.1% when the data is balanced at a 1:1 ratio. Additionally, during a field application at the E2312 mining face of Gaohe Energy Coal Mine, the IACGAN-ST model attained a diagnostic accuracy of 96.06%, significantly outperforming benchmark models such as SVM, CNN, and ResNet. This approach effectively mitigates data imbalance and significantly enhances the fault diagnosis accuracy for scraper conveyor bearings.