The robust feature extraction capability of deep learning (DL) has rendered it a popular choice for fault diagnosis in various industries. However, the accuracy of fault diagnosis models is often compromised due to an imbalance between normal and faulty data during the production process. To address this issue, this study proposes an unbalanced bearing fault sample diagnostic method based on a cycle generative adversarial network, using vibration fault diagnosis of rotating machinery rolling bearings as an illustrative example for validation purposes. The method utilizes continuous wavelet transform (CWT) to convert the original vibration signal into time-frequency maps, which are then utilized as input for the cycle generative adversarial network. To address issues related to unstable training and timely convergence of the model, spectral normalization and weight decay techniques are introduced, thereby enhancing the performance of the improved cycle generative adversarial network in generating additional fault samples. Finally, the Swin Transformer model is employed for fault diagnosis and compared with other methods such as Random Forest (RF), Sparse Autoencoder (SAE), Support Vector Machine (SVM), and Convolutional Neural Network (CNN). The diagnostic experiment encompasses sample expansion, imbalanced proportion between normal and faulty samples, as well as scenarios involving limited sample sizes. Experimental results demonstrate that the proposed method can generate higher quality synthetic samples when training data is scarce; moreover, utilizing the Swin Transformer for diagnosis yields superior accuracy compared to alternative approaches such as RF, SAE, SVM, or CNN models – thus showcasing significant potential in addressing imbalanced data within fault diagnosis.