This paper proposes a multimodal feature fusion approach for aero-engine gas-path fault diagnosis. By designing a multimodal feature fusion model, temporal fault data are transformed into Gramian Angular Summation Field (GASF) images and Recurrence Plot (RP) images. GASF images capture global trends, while RP images highlight local nonlinear characteristics, thereby enhancing fault recognition and overall algorithm performance. Experimental results demonstrate that this approach improves diagnostic accuracy compared with single-modality networks. Furthermore, a lightweight student model is constructed to substantially reduce model parameters and memory usage, while knowledge distillation (KD) ensures the diagnostic accuracy of the student model. By further integrating logit and feature distillation and employing reinforcement learning for automatic weight optimization, the student model achieves accuracy comparable to or even surpassing that of the teacher model. Compared with the independently trained teacher model, the parameter count is reduced by 93.78%; compared with the independently trained student model, accuracy increased by 11.09%. These results validate the effectiveness of the proposed method and provide a practical reference for aero-engine gas-path fault diagnosis.
Yang et al. (Tue,) studied this question.