Abstract: Objectives Addressing the challenges of data scarcity and imbalanced distribution of fault samples in the fault diagnosis of marine hydraulic steering gears, and considering that relying solely on real data struggles to support high-precision modeling while depending exclusively on virtual simulation data fails to adequately capture complex marine operating conditions, this paper proposes a diagnosis method based on virtual-real fusion to overcome the bottleneck of limited fault data in steering gear fault diagnosis. Methods First, a high-fidelity virtual model of the marine hydraulic steering gear is constructed, and its parameters are calibrated using real data to ensure consistency between the virtual model and actual operating conditions. On this basis, five typical states are simulated: slight oil pump leakage, severe oil pump leakage, slight valve sticking, severe valve sticking, and normal system operation, thereby generating a large amount of virtual data constrained by physical principles. Subsequently, the virtual generated data is effectively fused with the limited real data: the virtual data provides abundant fault samples, while the real data corrects operational deviations and introduces environmental noise characteristics. Addressing the characteristics of the fused data, a hybrid diagnostic model combining Temporal Convolutional Network (TCN, for extracting global features) and Gated Recurrent Unit (GRU, for modeling dynamic dependencies) is constructed. This model achieves multi-scale temporal feature extraction and dynamic dependency modeling, thereby deeply mining fault features and enhancing diagnostic accuracy. Finally, the effectiveness and advantages of this virtual-real fusion diagnostic method are validated through comparative experiments and ablation analysis. Results Experiments demonstrate that the virtual prototype not only compensates for the insufficiency of real data but also ensures data interpretability and controllability through physical modeling; after fusion with real data, the model achieves an accuracy of 99.6% in the diagnosis of five typical fault types and exhibits strong generalization capability.Conclusions The proposed virtual-real fusion diagnostic method fully leverages the complementary advantages of virtual prototypes and real data, effectively alleviating the issue of scarce fault samples in actual marine environments, and provides a reliable technical pathway with engineering application value for fault diagnosis of marine hydraulic steering gears.