Abstract: Rolling bearings are critical components of rotating machinery, and their health status directly affects the safety and reliability of industrial equipment. In practical fault diagnosis, labeled vibration data are typically available under specific working conditions, whereas target-domain data collected under different speeds, loads, or noise environments are often unlabeled. Because of distribution shifts across diverse operating conditions, diagnostic models trained only on source-domain data often suffer significant performance degradation when applied to target-domain datasets. Unsupervised domain adaptation can alleviate this issue by transferring diagnostic knowledge from labeled source data to unlabeled target data. However, most existing methods primarily focus on global distribution alignment while neglecting class-conditional distribution discrepancies, which may lead to feature confusion, ambiguous decision boundaries, and negative transfer under severe domain shifts. To address these limitations, this paper proposes a prototype-aligned conditional domain adversarial network (PACDAN) for cross-domain bearing fault diagnosis. The proposed architecture comprises three key modules: a time-frequency dual-stream feature extraction network, a distribution-morphology-aware prototype alignment unit, and a prototype-conditioned adversarial learning component. First, the raw vibration signal and its corresponding frequency-domain amplitude spectrum are used as two complementary input modalities. The time-domain branch extracts transient impact characteristics and local temporal features from vibration signals, whereas the frequency-domain branch captures dominant fault frequency components and spectral energy distribution information. The two groups of extracted features are fused to generate highly discriminative feature representations, thereby benefiting both fault classification and cross-domain adaptation. Second, stable class prototypes are constructed using labeled source-domain samples and high-confidence pseudo-labeled target-domain samples. Meanwhile, a dynamic prototype updating strategy is carefully designed to refine prototype representations throughout the training process. With the updated class prototypes, the proposed prototype alignment mechanism encourages cross-domain samples belonging to the same fault categories to cluster compactly in a unified feature space. This mechanism effectively reduces discrepancies in class-conditional distributions and suppresses feature dispersion under severe domain shifts. Third, the dynamically optimized class prototypes are incorporated into the domain discriminator as category-aware prior knowledge. Unlike conventional global adversarial learning schemes, the proposed method conducts adversarial optimization in class-specific feature subspaces rather than merely aligning global data distributions. This design enables the model to learn feature representations that are both domain-invariant and class-discriminative. Extensive experiments are conducted on two public bearing datasets, namely the Case Western Reserve University (CWRU) and Jiangnan University (JNU) datasets, covering 18 cross-condition transfer tasks. For each transfer task, all labeled source-domain samples and 20% of the unlabeled target-domain samples are used for model adaptation, whereas the remaining unlabeled target-domain samples are used for model evaluation and testing. The proposed PACDAN achieves average diagnostic accuracies of 99.71% and 99.12% on the CWRU and JNU datasets, respectively. Compared with several representative domain adaptation methods, PACDAN delivers higher and more stable diagnostic performance, especially under severe domain shifts. The effectiveness of prototype alignment, conditional adversarial learning, and time-frequency feature fusion in improving intra-class compactness, inter-class separability, and cross-domain feature alignment is verified through ablation experiments, confusion matrices, and t-SNE visualization results. Under Gaussian noise disturbances ranging from −6 to 0 dB, PACDAN achieves an average diagnostic accuracy of 95.97%, demonstrating its robustness in noisy environments. These results indicate that PACDAN can effectively reduce discrepancies in class-conditional distributions and improve the reliability of cross-condition unlabeled bearing-fault diagnosis.