Electro-Hydrostatic Actuator (EHA) is the core actuation unit in aerospace flight control systems, and its health status directly affects maneuverability and flight safety, making fault diagnosis of great engineering significance. However, EHA operates under varying conditions across flight phases, while fault samples are rare and costly to label. To tackle the dual challenges of domain shift and label scarcity, this paper proposes a dual-alignment unsupervised domain adaptation fault diagnosis method combining Multi-Kernel Maximum Mean Discrepancy (MK-MMD) with domain-adversarial training. A one-dimensional convolutional neural network serves as the feature extractor. The classifier is first trained on labeled source data using a cross-entropy classification loss to ensure fault discrimination. Then, the MK-MMD loss explicitly aligns global distributions in the reproducing kernel Hilbert space, while the domain-adversarial loss via a gradient reversal layer implicitly eliminates fine-grained discrepancies. The classification loss, MK-MMD loss, and domain-adversarial loss are jointly optimized to learn domain-invariant features that preserve fault-discriminative capability while being insensitive to operating condition variations, enabling cross-condition diagnosis without target labels. Experiments on EHA data demonstrate that the proposed method outperforms existing methods on both transfer tasks with strong noise robustness.

Electro-Hydrostatic Actuator, fault diagnosis, transfer learning, domain-adversarial network