With the development of Autonomous Marine Vehicles, research on real-time health monitoring based on data from remote monitoring and shore-based control center has attracted more and more attention. However, the feature extraction capability of each local client model is limited by the size and quality of data collected by remote monitoring. For clients with zero-shot, it is necessary to design an efficient federated learning framework which allows shore-based federation center with few-shot help to strengthen the feature extraction capability of clients. In this paper, a semi-supervised federated learning strategy based on adaptive learning mechanism of prototype points is proposed to solve the problem that clients with zero-shot have difficulty in establishing reliable fault diagnosis models when the quality of client data is inhomogeneous. By designing an adaptive prototypical network learning mechanism driven by sample quality and a dynamic in-and-out mechanism for local pseudo-labeled dataset, the few-shot learning capability of federation center as well as the reliability of local client’s prototype points are enhanced to alleviate the impact of unreliable pseudo-label on semi-supervised learning. The experiment validation on the rolling bearing dataset provided by Case Western Reserve University can come to an accuracy increasement of 13.12%, which indicates that the proposed Adpro-SSFL outperforms existed semi-supervised federated learning method when data suffered by high missing rate due to the interference of external factors on the signal acquisition sensor.

随着自主海洋车辆的发展，基于远程监控和岸基控制中心数据的实时健康监测研究越来越受到关注。然而，每个本地客户端模型的特征提取能力受到远程监控收集的数据的大小和质量的限制。对于零样本的客户端，有必要设计一个高效的联邦学习框架，让岸基联邦中心在少样本的帮助下，增强客户端的特征提取能力。针对零样本客户端在客户端数据质量不均匀的情况下难以建立可靠的故障诊断模型的问题，提出一种基于原型点自适应学习机制的半监督联邦学习策略。通过设计样本质量驱动的自适应原型网络学习机制和本地伪标签数据集的动态进出机制，增强联邦中心的小样本学习能力以及本地客户端原型点的可靠性减轻不可靠的伪标签对半监督学习的影响。在凯斯西储大学提供的滚动轴承数据集上进行的实验验证可以达到 13.12% 的准确率提升，这表明所提出的 Adpro-SSFL 在数据因高缺失率而遭受高缺失率的情况下优于现有的半监督联邦学习方法。外界因素对信号采集传感器的干扰。