This research addresses the challenges of zero-shot and limited-data fault diagnosis in industrial systems, where sufficient labeled data for all fault types are unavailable. To overcome this issue, several generative learning frameworks were developed to synthesize discriminative and attribute-consistent fault features.First, GAN-based models incorporating semantic distance and attribute-consistent constraints were proposed to transfer knowledge from seen to unseen fault classes, improving diagnostic performance under zero-shot conditions. These methods were validated and published in IEEE journals.To further enhance stability and robustness, a diffusion-based model (DP-CDDPM-AC) was developed, leveraging dual-path generation to improve attribute consistency and mitigate domain shift. In addition, a meta-learning-based generative framework (CycleML-AC) was proposed to handle limited-data scenarios.These studies have resulted in two published journal papers (IEEE TII, IEEE TIM) and two papers currently under review.