Although current data generation methods can enhance the classification performance of existing fault diagnosis models by augmenting the limited training set, their performance improvement is still restricted due to the low quality and insufficient diversity of generated data. To tackle this issue, we propose a cross-domain fault audio generation method based on the improved star generative adversarial networks, namely StarGAN-Aug. In the StarGAN-Aug, the model structure is first optimized using a pre-trained JDC network and a style encoder, leading to efficiently extract the specific frequency in fault signals and generate diverse audio samples across domains. Moreover, we add a comprehensive objective to optimize the model training and enhance its ability to ensure that the generated samples have both high fidelity and diversity. Experimental results show that using our generated audio samples, the classification performance of existing diagnostic models can be significantly improved.