Abstract：Intelligent fault diagnosis for aero-engine rotor systems is of great importance for ensuring flight safety and improving maintenance efficiency. However, the high scarcity of authentic fault data and class imbalance in aero-engine rotors severely constrain the performance of existing intelligent diagnosis models in few-shot scenarios. To address these challenges, this paper proposes a multimodal general foundation model for intelligent fault diagnosis based on contrastive learning. The proposed method constructs a cross-modal alignment framework between vibration frequency-domain features and structured Chinese natural language descriptions A Transformer-based general foundation model is pre-trained on large-scale general rotating machinery datasets to learn universal "signal-semantic" fault representations. During the fine-tuning phase, condition-aware semantic s are introduced to inject the operating state information of the target equipment into the model, thereby guiding rapid adaptation under few-shot diagnostic scenarios. Experimental results on rotor system datasets demonstrate that the proposed method achieves a diagnostic accuracy of 73.11% in the one-shot setting, representing an improvement of nearly 40% over existing methods. When the number of samples per class increases to ten, the accuracy further improves to 99.23%. Ablation studies indicate that model pre-training and operating-condition s significantly accelerate convergence. These results suggest that the proposed method exhibits strong generalization capability and few-shot adaptability, providing an efficient and reliable intelligent diagnostic solution for high-value equipment such as aero-engines.
2026, Vol. 47 
