Application of the TF-ResNet-MSA Model in the Complex Fault Diagnosis of EHA System Plunger Pumps

To address the challenge of accurately diagnosing complex faults in Electro-Hydrostatic Actuation (EHA) system plunger pumps, which are characterised by multimodal signals, weak fault features, and class imbalance under high-pressure and high-dynamic working conditions, this paper proposes a novel TF-ResNet-MSA model enhanced with transfer learning. The model utilises a ResNet-50 backbone, incorporates a Multi-Scale Attention (MSA) mechanism for adaptive weighting of local and global time-frequency features, and integrates a Transformer encoder to capture long-range dependencies among vibration, pressure, and acoustic modalities. Furthermore, batch-wise dynamic weighting is introduced during training to mitigate label imbalance, thereby enabling end-to-end fine-tuning with limited labelled data. The experimental findings, derived from a simulated dataset, demonstrate that the proposed model attains an accuracy of 0.982 and an F1-score of 0.977 across five distinct fault types. This outcome signifies a 6.9 percentage point enhancement over the ResNet-50 baseline. In real-world noise and varying working conditions, the model demonstrated an accuracy of 0.953 and an F1-score of 0.947, exhibiting only a 2.9-point degradation from the simulation. This performance significantly surpasses that of ablated models employing MSA (0.903) or Transformer alone (0.896). The confusion matrix analysis indicates a misclassification rate of less than 2.2%, and the model demonstrates an accuracy of over 0.918 under SNR = 10 dB. The findings demonstrate the model's superior diagnostic precision and robustness in practical EHA scenarios, offering an intelligent solution for fault detection and predictive maintenance in aerospace hydraulic actuators that are highly efficient and interpretable.