<p>Bearing fault diagnosis under data imbalance and complex operating conditions remains a challenging problem for intelligent maintenance systems. The main difficulty lies not only in the scarcity of fault samples, but also in effectively capturing the complementary characteristics of time-domain transient signals and frequency-domain spectral patterns. To address these challenges, this paper proposes a Dual-Domain Generative Adversarial Network with Cross-Domain Structural Regularization (CSR-DGAN). The proposed framework is designed from a problem-driven perspective, dual domain-specific generators and discriminators are employed to explicitly model the data distributions in the time and frequency domains, enabling the learning of complementary representations. To further enhance the consistency between the two domains, a cross-domain structural regularization term is introduced to align the generated representations and improve their structural coherence. Moreover, residual connections are adopted to facilitate stable feature propagation and alleviate training instability, while the self-attention mechanism enables the model to capture long-range dependencies within each domain. These components are integrated to jointly improve representation quality and adversarial training stability. Extensive experiments on multiple bearing fault datasets demonstrate that CSR-DGAN outperforms existing generative augmentation methods in terms of distribution similarity, cross-domain consistency, and downstream diagnostic performance, highlighting the effectiveness of the proposed problem-driven dual-domain generative framework for robust fault diagnosis under imbalanced conditions.</p>