To address the issues of insufficient multi-domain alignment and inadequate multi-source information collaboration in multi-source domain adaptive fault diagnosis, an end-to-end multi-source unsupervised domain adaptation fault diagnosis method is proposed, named the adaptation network based on multi-granularity alignment and evidential reasoning rule (MAERAN). Firstly, a multi-granularity alignment module is designed to achieve progressive multi-level distributed alignment from shallow to deep. Specifically, the shallow domain discriminator is used to align the shallow features of the shared domain to reduce the distribution bias of shallow features. Next, a multi-view sub-network is constructed to extract the deep features from each pair of source and target domains, and the domain adversarial strategy is implemented to align the feature distribution. By mining domain invariant features through two-stage domain alignment, the difference of cross-domain distribution is effectively reduced. At the same time, fine-grained class alignment is performed to enhance category distinguishability. Then, a multi-source information collaboration decision-making module based on evidential reasoning rule is proposed, which adaptively weights and fuses multi-source diagnostic information to enhance the generalization ability and diagnostic performance of the model. Finally, experimental results on two bearing datasets validate the effectiveness and superiority of the proposed method in multi-source transfer tasks.