Adaptive Physics-Data Fusion for Data-Efficient Prognostics of Aircraft Hydraulic Filters

Abstract

Due to the stringent cleanliness requirements of the aviation hydraulic system, predicting the health status of hydraulic filter is essential for ensuring flight safety. However, it is still a challenge to have an accuracy prediction due to complex physical degradation mechanisms, and insufficient individual degradation data in early stage. Therefore, an Adaptive Physics-Data Fusion (APDF) model is proposed to predict hydraulic filter health status by integrating the advantages of physical and data information. Specifically, a parameter-updated Ergun equation, describing the pressure drop across the filter with the contaminant deposition, is taken as physical knowledge to guide the design of neural network structure and loss function. In addition, an adaptive weighting strategy in the loss function is also developed to dynamically balance the physical and data contributions. The experimental results from the hydraulic filter degradation experiment show that the model achieves a pressure drop prediction accuracy with RMSE of 0.0076 MPa. Notably, under limited data in the early prediction stage (only with 40% training data), the APDF represents 87.5% improvement compared to traditional data-driven models (such as LSTM). This performance highlights significant advantages in scenarios with limited aviation sensor configurations.