FedPM-SGN: A federated graph network for aviation equipment fault diagnosis by multi-sensor fusion in decentralized and heterogeneous setting

Aviation equipment fault monitoring and diagnosis play a pivotal role in ensuring flight safety. The data-driven approaches have flourished and shown satisfying performance in academia and industry. However, its application to aviation equipment is hindered by two issues: heterogeneous data islands and multi-sensor information fusion. Aviation equipment dispersity and uneven load distribution lead to incomplete and inconsistent monitoring data for each piece of equipment, resulting in a heterogeneous data island. Furthermore, the complex structure of aviation equipment requires a multi-sensor monitoring system that makes high-dimensional and strongly-associated data. Motivated by this, a federated relation self-perception graph network with the prototype-based feature matching mechanism (FedPM-SGN) is built to address these two problems. In this federated learning framework, each client is equipped with a novel relation self-perception graph network (SGN), where a relation self-perception layer upon the attention mechanism is devised to perceive the complex interaction between multi-sensor measurements and establish various graph topologies for different fault states. Furthermore, to mitigate the feature space heterogeneity in each client, a supervision loss named contrastive guiding (CG) is added to the objective of each SGN, inducing the local feature to match the corresponding prototype and keep the non-corresponding prototype away. Experiment results on rotor system and centrifugal pump datasets illustrate that the FedPM-SGN can achieve a satisfying performance in the decentralized and heterogeneous scenario, presenting a promising prospect in collaborative fault diagnosis.