Rolling bearing fault diagnosis is important for condition monitoring of rotating machinery. In practice, variations in operating conditions and heterogeneity among source domains of-ten cause distribution shifts in vibration signals, while labeled target-domain data are usually unavailable. To address this issue, this paper proposes TMDANN-A, a multi-source domain adaptation framework for vibration-based bearing fault diagnosis. The framework combines adversarial domain alignment, boundary-aware discriminative learning, and condition-guided source weighting to improve domain invariance and class separability. A hierarchical CNN-Transformer encoder is used to capture local impulsive features and long-range tem-poral dependencies. Experiments on the JNU and PU datasets achieve 98.51% average accu-racy over six transfer tasks. On an industrial test-rig dataset, the method further achieves 95.42% average accuracy, 95.44% Macro-F1, and 95.22% mean fault recall. The results verify its effectiveness under complex operating conditions.