Application-oriented small-sample transient gas-path fault diagnosis for aero-engines based on multi-source data fusion

During aero-engine transient operation, gas-path parameters fluctuate sharply and the system exhibits strong coupling, which makes fault features weak and difficult to distinguish. In practical applications, fault samples are also scarce and class boundaries are often blurred, which severely limits the performance of data-driven diagnostic methods. To address these challenges, a multi-source data fusion framework for aero-engine transient fault diagnosis is proposed by jointly integrating spectral distance-regularized generative adversarial network and graph convolutional network (GCN). First, this study develops a dynamic real-time aero-engine model that incorporates heat transfer effects, and uses hardware-in-the-loop simulation to generate transient gas-path data with high dynamic consistency and physical fidelity for fault sample augmentation. Next, spectral distance is introduced as a structural consistency constraint between real and generated multi-source gas-path samples, and a dynamic weighting mechanism is employed to emphasize transform-domain components with larger generation discrepancies, thereby improving the fidelity and separability of generated fault samples. Furthermore, the original and generated samples are organized into a topologically correlated graph, and cross-sample structural relationships are extracted by the GCN to achieve robust fault identification under complex transient operating conditions. Experimental results show that the proposed method achieves classification accuracies of 98.83% and 97.49% at the maximum takeoff and subsonic cruise operating points, respectively. Under small-sample and weak-discrepancy fault scenarios, the proposed method significantly outperforms other state-of-the-art methods in diagnostic accuracy, stability, and generalization capability.