<p>During ship operation, a large number of sensors are usually installed to obtain massive amounts of data to provide information for analyzing the ship’s condition. However, the large amount of data acquired often contains only a small amount of labeling information, making it difficult to build reliable fault diagnosis models using traditional data-driven methods. Federated learning can combine multiple SSL models to enhance the diagnostic performance of SSL models, thereby building a robust global model capable of handling serve complex local fault tasks. However, these semi-supervised federated learning (SSFL) methods ignore the harmful federal migration problem caused by inconsistent inter-client data reliability. Therefore, a trustworthy SSFL framework is proposed in this paper to mitigate the impact of unreliable data with a federated global model. In the federation process, the designed federated learning mechanism takes advantage of the trustworthy global teacher model to guide the local student model in mining the deeper features of the unreliable samples. On the other hand, when inter-client trustworthiness is inconsistent, the teacher-student model consistency (TSC) norms in the federated center are applied to drive the joint learning of the federated global model and the local clients to construct the optimal federated model. The experiment results show that TSC-SSFL improves the fault diagnosis accuracy by 14.55% compared to the baseline method when the data is affected by noise pollution and missing sample features. In addition, the fault diagnosis results in different experiment scenarios are all better than the comparison methods, which fully demonstrates the effectiveness and superiority of TSC-SSFL.</p>