Abstract
Mechanical fault diagnosis has become critical with the growing complexity and automation of industrial systems. In practical scenarios, fault data are often distributed across decentralized clients with heterogeneous data distributions and privacy constraints, while different clients may adopt different model architectures due to deployment-specific requirements. Federated learning (FL) facilitates collaborative model training across decentralized clients without sharing raw data; however, standard parameter-aggregation-based FL generally relies on architectural consistency across clients, which is difficult to satisfy in model-heterogeneous mechanical fault diagnosis. To address this problem, this article proposes federated distilled data aggregation with proxy-based tridirectional training (FedDDA-P3T) for model-heterogeneous federated fault diagnosis. More specifically, each client independently distills its local dataset into a compact distilled dataset and uploads it to the server. The distilled data are not used merely as compact substitutes for local datasets; instead, they serve as model-agnostic knowledge carriers and are aggregated on the server to support knowledge transfer across heterogeneous client models. Moreover, the server-aggregated distilled data are redistributed to clients as shared training anchors. A proxy model is further introduced to bridge heterogeneous private models through these shared anchors, forming a proxy-based tridirectional training strategy. Experiments on mechanical fault diagnosis datasets demonstrate that FedDDA-P3T achieves superior diagnostic performance over representative state-of-the-art FL methods under model-heterogeneous settings.