An intelligent fault diagnosis method of rolling bearing based on expansion and transfer of single working condition data under small samples

To address the challenge of limited labeled data and complex fault characteristics in small-sample rolling bearing diagnosis, an intelligent diagnosis method of rolling bearing based on single working condition data expansion and transfer under small samples is proposed. Eighteen state simulation schemes of rolling bearing including loose sensor connection under six working conditions were designed, and the variational mode decomposition-gray texture image (VMD-GTI) samples were obtained by using the vibration signal conversion method. To enhance diagnostic reliability under small-sample conditions, an improved deep convolutional generative adversarial network (DCGAN) is employed for small-sample image expansion, thereby enriching data diversity. To enhance feature representation and improve fault diagnosis accuracy, an enhanced ResNet-18-EMA-Conv is developed by incorporating exponential moving average (EMA) convolution into residual neural network (ResNet). To enable knowledge transfer under limited data, the improved DCGAN is adopted to augment single-condition samples, and the pretrained ResNet-18-EMA-Conv model is validated through both self-transfer and cross-condition transfer. The results show that when the number of original samples is 100, the optimal expansion ratio is 1:1. After expansion, the accuracy of fault diagnosis using ResNet is improved from 0.8968 to 0.9467, which is increased by 0.0499, and the accuracy of fault diagnosis model improved by EMA convolution is further improved from 0.9467 to 0.9732, which is increased by 0.0265. Using the self-transfer of single-working condition data, the accuracy of fault diagnosis is improved from 0.9732 to 0.9810 again, which is improved by 0.0078. The average testing accuracy increased from 0.8523 to 0.9378 and increased by 0.0855 under six working conditions. These results confirm that the proposed framework effectively enhances diagnostic accuracy and generalization for complex bearing fault modes under small samples and multi-condition scenarios.