The distribution of vibration signals of rolling bearings under variable working conditions is significantly different. When using multi-source domain knowledge, the existing deep transfer diagnosis methods often face the problems of " negative transfer " interference caused by source domain differences and insufficient accuracy of cross-domain feature alignment, which seriously restricts the reliability of fault diagnosis. In order to solve the above problems, a deep transfer fault diagnosis method (DTRSR) based on seed replacement and residual reconstruction is proposed. Aiming at the distribution drift under multi-source conditions, a deep transfer learning framework integrating multi-source information extraction and residual feature alignment is constructed. The framework uses the shared weight convolutional neural network to mine the common features of multi-source domains, adopts the multi-source domain training mode, and uses the shared one-dimensional convolutional neural network (1D-CNN) joint model commonness to identify the characteristics of multiple known working conditions. It solves the problem that the knowledge coverage of the single source domain is limited and it is difficult to deal with the unknown variable working conditions of the target domain. In terms of transfer learning mechanism, a novel cross-domain alignment method based on bi-objective residual seed reconstruction is proposed. This method introduces a small sample learning strategy, uses a small number of labeled samples in the target domain ( such as only a single digit per class ) as " seeds ", introduces a clustering algorithm to extract the distribution center of the feature space of the target domain, and uses the Nearest neighbor matching strategy to match the class center of the source domain and the seed of the target domain one by one. This residual reconstruction mechanism can not only achieve fine-grained conditional distribution alignment, but also effectively overcome the risk of " negative transfer " caused by the large inter-domain gap in the variable working condition scenario, and ensure the high robustness of the model in multi-source cross-domain diagnostic tasks. At the same time, the balanced distribution adaptation (BDA) mechanism is introduced to dynamically adjust the weights of edge distribution and conditional distribution, which achieves fine-grained feature alignment and significantly improves the accuracy of feature alignment. The experimental results of multiple sets of multi-source transfer tasks (such as BC-A, AC-B, etc.) designed on the CWRU bearing dataset show that the proposed method is superior to single-source domain transfer and traditional joint distribution adaptive (BDA) method in classification accuracy and diagnostic stability. The t-SNE visualization and robustness analysis further confirm that the method can effectively enhance the separability between classes and the degree of polymerization in the domain, which provides an effective solution to the problem of negative migration and insufficient alignment accuracy under variable conditions.