PGMTKD: A Physics-Guided Multi-Teacher Knowledge Distillation Network for External Gear Pump Onboard Fault Diagnosis with Missing Modality

Flight safety and engine performance are directly impacted by the status of the external gear pump, which is an important part of the engine?s fuel system. However, diagnosing faults in external gear pumps still faces many chanllenges in practical applications. First, compared to the controlled conditions of a ground laboratory, the onboard flight environment has limited number of sensors, rendering it a situation with missing modalities. Second, due to constraints in aircraft computing resources, it is difficult to achieve the same level of fault diagnosis capabilities as in laboratory settings. To overcome the identified problems, this study introduces a Physics-Guided Multi-Teacher Knowledge Distillation Network (PGMTKD). To deal with missing modalities during flight, three different teacher networks were developed in this research. Among these, we implemented one physics-informed network that incorporates a pressure reconstruction model based on a spectral method in fluid dynamics, which serves as a physical constraint. The second teacher network using a GPT-based structure is designed to learn temporal relationships in the data. The third teacher network adopts an ResNet structure to extract spatial fault features. Through knowledge distillation, the fault-related knowledge is efficiently learned by a smaller 1D-CNN student model. The student network is a more practical choice for an onboard environment where computing power is limited. Lastly, the PGMTKD model outperformed several baseline methods using our laboratory established experimental test rig, showcasing improvements in diagnosis performance. Experimental results demonstrate that PGMTKD achieves higher diagnostic accuracy under conditions of limited modality data. The code is available at https://github.com/tsed563/PGMTKD.