Mechanical equipment operating under complex working conditions is highly prone to failure. If such failures are not diagnosed in a timely and accurate manner, they may lead not only to performance degradation and economic losses but also to serious safety accidents. Therefore, developing efficient and reliable intelligent fault diagnosis methods is of significant engineering importance. However, in real industrial scenarios, the number of fault samples in monitoring data is usually limited, resulting in data imbalance problems that severely constrain the accuracy and robustness of traditional diagnostic models. To effectively mitigate this issue, this paper introduces spectral graph convolution and a hybrid attention module, and proposes an improved cycle-consistent generative adversarial network for generating high-quality fault samples, thereby enhancing intelligent mechanical fault diagnosis under imbalanced data conditions. Specifically, spectral graph convolution models global pixel dependencies through sparse adjacency matrices, improving long-range feature interactions while reducing computational complexity. Meanwhile, the hybrid attention module dynamically assigns weights at both channel and spatial levels to highlight critical regions and strengthen feature representation. With the proposed improved cycle-consistent generative adversarial network, more realistic and diverse fault samples can be generated, effectively augmenting minority-class data and alleviating the limitations imposed by data imbalance on intelligent fault diagnosis performance. Experimental results on the Beijing Jiaotong University metro bogie dataset and the Soochow University bearing dataset show that the proposed method significantly outperforms comparison approaches in three image quality evaluation metrics and fault classification accuracy. These results validate its diagnostic effectiveness under imbalanced data conditions and demonstrate that it provides a practical and feasible solution for addressing data imbalance challenges in industrial applications.