Graph Convolutional Contrastive Learning for Few-Shot Open-Set Fault Diagnosis in Electro-Hydrostatic Actuators

Data-driven intelligent diagnosis methods for aerospace electro-hydraulic actuators (EHA) under few-shot conditions have achieved some progress. However, most approaches rely on a closed-set assumption, constructing diagnostic models solely for known fault categories while neglecting the occurrence of open-set scenarios involving unknown fault types during actual operation. This severely limits the algorithms' generalization capability and practical applicability. Aiming at this key issue, this paper proposes a small-sample fault diagnosis method for EHA that integrates graph convolutional contrast learning with unknown category perception mechanism (GCN-CN) of open-set data. The method introduces a cross-state difference-driven contrast learning strategy through graph convolutional neural networks, and constructs a Kullback-leibler scatter edge feature matrix fusing time-frequency features during the graph construction process, in order to enhance the network's ability of recognizing sample differences. On this basis, an unknown fault recognition mechanism based on the similarity distribution characteristics is designed to realize the effective discrimination of whether the input samples belong to an unknown fault category. Experimental results show that the proposed diagnostic network structure can still achieve excellent diagnostic performance while reducing the data volume requirement, and the fault diagnosis accuracy can reach over 99%. Meanwhile, the proposed fault category diagnosis mechanism can achieve more than 80% diagnosis accuracy for unforeseen fault types.