In recent years, the researches of fault diagnosis has increasingly focused on the application of transfer learning. The majority of current transfer learning approaches are based on the assumption that the label spaces of the source and target domains are identical. However, it is more common that the target domain label space is a subset of the source domain label space, which is called partial migration problem. To address this issue, a new double-stage partial adversarial network in cross-domain fault diagnostics (DS-PAN) has been proposed in this paper. This method uses local adversarial to realize feature alignment of source domain and target domain, and trains two feature classifiers to process the same samples, and then reverse trains feature extractor to ensure that the original data can be mapped to the ideal feature space. Experiments were carried out on the bearing data collected on the experimental platform, and several typical diagnostic methods were selected for comparison. The results prove the effectiveness and superiority of the proposed method.