Intelligent mechanical fault diagnosis techniques have been extensively developed in recent years. Owing to the advantage of data privacy protection, federated learning has recently received increasing attention; this approach can utilize monitoring data from multiple local clients to train an optimal global diagnosis model. However, low-quality data are often present for some clients, including mislabeled data and incomplete data that lack some health states. Furthermore, the data distributions are usually different across different clients owing to variations in machine operating conditions. Therefore, the performance of the federated diagnostic model may be limited by low-quality data and data distribution discrepancies. To address this issue, a federated transfer learning method with low-quality knowledge filtering and dynamic model aggregation is proposed. First, a dynamic filter is designed by filtering out low label probabilities predicted by low-quality source domain models to construct high-confidence pseudo-labels for the target domain data. Then, the batch normalized maximum mean discrepancy (BN-MMD) distance metric is introduced into the training loss function to reduce the data distribution discrepancy between the source clients and the target client without private data leakage. While building the global model in each training round, a dynamic model aggregation algorithm is proposed to mitigate the influence of low-quality clients. This algorithm evaluates the weight of each client according to its contribution to the total consensus diagnostic knowledge and then aggregates the local models with adaptive weights. Consequently, it can overcome the drawback of the classical federated averaging (FedAvg) algorithm, where all local clients are assigned the same weight when constructing the global model. Experiments are conducted on three bearing datasets under various loads and speeds. Compared with some existing diagnostic methods, the proposed federated transfer learning method can reduce the impact of low-quality data and achieve higher diagnostic accuracy while preserving data privacy.