Abstract: Rolling bearings are critical load-bearing components in rotating machinery, and their reliable operation is paramount for industrial safety. In practical applications, bearing fault diagnosis often faces challenges such as scarce labeled data and variable working conditions, which lead to significant distribution discrepancies between the source and target data. While Unsupervised Domain Adaptation (UDA) methods based on deep learning have been introduced to address these domain shifts, existing approaches primarily focus on global marginal distribution alignment. They frequently neglect the fine-grained, class-conditional distribution discrepancies, which inevitably leads to the confusion of fault features at the decision boundaries and negative transfer, especially under severe domain shift scenarios. To overcome these critical limitations and ensure highly accurate cross-domain knowledge transfer, a novel unsupervised domain adaptation framework termed Prototype-Aligned Conditional Domain Adversarial Network (PACDAN) is proposed for cross-domain bearing fault diagnosis. The proposed method operates through three core innovations. First, the framework constructs a time-frequency dual-stream feature extraction network. It processes raw one-dimensional temporal vibration signals alongside their corresponding frequency-domain amplitude spectra in parallel branches. This dual-stream architecture deeply and comprehensively captures the multi-scale fault patterns embedded in non-stationary signals, providing a robust and semantically rich feature representation for subsequent adaptation. Second, a distribution-morphology-aware class prototype alignment mechanism is designed to explicitly model and align the intra-class structures. Instead of merely computing the distance between geometric centers, this strategy synchronously matches the first-order position (expectation) and the second-order morphological divergence (central moment) of the feature distributions across domains. By enforcing the statistical isomorphism of the cross-domain distributions, the method compels samples of the same category to tightly cluster around their respective prototypes in the shared feature space. Consequently, it effectively suppresses feature dispersion and structural collapse under severe domain shifts. Third, the framework integrates a conditional domain adversarial learning strategy to achieve precise, class-level domain alignment. The class prototypes, which are dynamically updated using an Exponential Moving Average (EMA) mechanism to guarantee stability during training, are explicitly embedded into the domain discriminator as prior conditional information. This strategic embedding forces the generator and discriminator to perform their adversarial minimax game within specific, fine-grained class subspaces. Therefore, the discriminator focuses on aligning the conditional distributions rather than the global distribution, sharply defining the class boundaries and avoiding cross-class misclassification. Extensive validation experiments are conducted on two widely recognized public platforms: the CWRU dataset and the JNU rotor test rig dataset. The method is rigorously evaluated across 18 distinct cross-condition transfer tasks involving various motor speeds and defect sizes. The experimental results demonstrate that the proposed PACDAN framework achieves remarkable average classification accuracies of 99.71% and 99.12% on the CWRU and JNU datasets, respectively. Furthermore, comprehensive ablation studies, confusion matrices, and t-SNE feature visualizations confirm the necessity of the core components and the superiority of the learned discriminative features. Compared to classical adversarial networks and recent state-of-the-art domain adaptation methods, the proposed approach demonstrates significant performance improvements and exceptional stability under severe domain-shift scenarios, exhibiting superior robustness and immense potential for generalization in unlabeled industrial applications.