The imbalanced data will seriously restrict the effect of fault diagnosis, we propose a method that can stably generate minority samples with multiple labels. We introduce the Wasserstein loss and gradient penalty into the c-GAN (conditional Generative Adversarial Nets) framework with the ability of multi-label sample generation to ensure the stability of sample generation while ensuring the flexibility and diversity of sample generation. The constructed CWGAN-GP (conditional - Wasserstein loss - Generative Adversarial Networks with gradient penalty) model is used to supplement the imbalanced dataset by generating few-shot samples, so that, the imbalanced scenes are transformed into balanced ones. The bearing case study results show that the proposed method can improve the diagnosis accuracy by 2.27% in the 8:2 (normal: fault) imbalance scenario, which proves that the proposed method can effectively solve the fault diagnosis problem under the condition of imbalanced samples.