Generally, high performance of deep learning (DL)-based machinery fault diagnosis methods relies on abundant labeled fault samples under various working conditions, while they are usually stored by different users and not communicated with each other due to data privacy protection. Federated learning (FL) is a possible solution, but the traditional federated averaging (FedAvg) algorithm in FL ignores the potential domain shift of different FL participants, which limits its further application. Therefore, a federated transfer learning (FTL) framework with discrepancy-based weighted federated averaging (D-WFA) is proposed to train the good global diagnosis model collaboratively as well as protect data privacy. First, local labeled source domain data and unlabeled target domain data are utilized to update multiple local models with generalization ability. Then, a maximum mean discrepancy (MMD)-based dynamic weighted averaging algorithm is designed to aggregate the updated local models with automatically learned weight. The proposed D-WFA overcomes the disadvantage of the traditional FedAvg algorithm, which assumes that all clients have the same contribution in constructing the global model during FL training. Experiment results on a bearing dataset show that the proposed D-WFA outperforms the traditional FedAvg and relative FTL method, which offers a promising solution in privacy-preserving machine learning for fault diagnosis.