Prototype-assisted two-stage domain adaptation network for electro-hydraulic actuator fault diagnosis under variable operating conditions

Electro-hydraulic actuators (EAs) are safety-critical components of large transport aircraft and their condition directly affects the reliability of primary flight-control surfaces. In practice, however, labeled fault data collected from physical actuators are scarce, whereas simulation data are relatively inexpensive to generate. The distribution discrepancy between simulated and hardware-in-the-loop (HIL) signals prevents a diagnostic model trained only on simulation data from being directly transferred to variable physical operating conditions. To address this simulation-to-HIL transfer problem, we formulates EA fault diagnosis as a closed-set unsupervised domain adaptation task and proposes a prototype-assisted two-stage domain adaptation network (PTDAN). In the first training stage, adversarial domain adaptation and local maximum mean discrepancy are combined to reduce global and class-conditional distribution discrepancies, respectively. In the second training stage, the feature extractor and domain classifier are frozen, and the label predictor is refined using source labels, confidence-filtered target pseudo-labels, and prototype-based consistency regularization. Experiments are conducted on labeled Simulink/AMESim co-simulation data and unlabeled HIL data under multiple relief-valve pressures and load torques. Across six transfer tasks, the second training stage improves the average target-domain accuracy by 14.82 percentage points over the first training stage. In the HIL transfer experiments, PTDAN achieves an average target-domain accuracy of 81.65%, compared with 60.84% for MSFN. These results demonstrate the potential of PTDAN to reuse low-cost simulation knowledge for diagnosing EAs under unlabeled and variable operating conditions, thereby reducing dependence on costly physical fault acquisition and manual annotation.