In actual production, bearings are usually in a normal working state, which results in a lack of data for fault diagnosis (FD). Yet, the majority of existing studies on FD of rolling bearings focus on scenarios with ample fault data, while research on diagnosing small-sample bearings remains scarce. Therefore, this study presents an FD method for small-sample bearings, employing variational-mode decomposition and Symmetric Dot Pattern, combined with a pre-trained and fine-tuned Residual Network18 (VSDP-TLResNet18). The approach utilizes variational-mode decomposition (VMD) to break down the signal, determining the k value and the best Intrinsic-Mode Function (IMF) component based on center frequency and kurtosis criteria. Following this, the chosen IMF component is converted into a two-dimensional image using the Symmetric Dot Pattern (SDP) transform. In order to maximize the discrimination between two-dimensional fault images, Pearson correlation analysis is carried out on the parameters of SDP to select the optimal parameters. Finally, we use the pre-trained and fine-tuned method combined with ResNet18 for small-sample FD to improve the diagnosis accuracy of the model. Relative to alternative approaches, the suggested method demonstrates strong performance when dealing with small-sample FD.

In rolling bearing intelligent fault diagnosis (FD), lightweight models are constrained by issues such as noise interference and the scarcity of fault data, making it challenging to achieve real-time, high-accuracy diagnosis on resource-limited devices. To address these challenges, this study proposes a lightweight model that combines the hierarchical fine-grained decision fusion (HFDF) strategy with an improved EfficientNetV2 architecture (HFDF-EffNetV2). The model optimizes depth and width multiplicity factors to enhance parameter utilization efficiency. It uses pyramidal convolution (PyConv) combined with Fused-MBConv (Fused-MBPyConv) to obtain multi-scale time-frequency information. Additionally, an enhanced MBConv, termed BSMB-Conv-MLCA, integrates subspace blueprint separable convolution (BSConv-S) with mixed local channel attention (MLCA) extract deep-grained fault features. The HFDF strategy outputs confidence in stages and updates weights to prevent the model from falling into local overfitting when handling confusable samples. Experimental results on Case 1 and Case 2 show that HFDF-EffNetV2 achieved 100% accuracy with diagnostic times of 18.67 millisecond (ms) and 17.56 ms, respectively, and 1.85 million (M) parameters. Under noisy conditions, average accuracies reached 98.19% and 85.68%, respectively. Additionally, the model performed well with small samples, yielding accuracies of 98.69% and 97.51%. These results highlight its superior robustness to noise and lightweight performance compared with other advanced models.