Energy Informed Temporal—State Neural Network for Aircraft EDP Spring Failure Prediction

ABSTRACT Early spring failures in aircraft engine driven pump (EDP) systems cause subtle flow and pressure changes that may compromise flight safety if undetected. Based on the definitions of early spring failure from the Aircraft Maintenance Manual, we constructed a dataset for early spring failures. To address the issue of subtle early fault features, we collected the hidden features at each time step from a gated neural network. We propose an energy‐informed neural network to predict flow and pressure, while amplifying the subtle early failure characteristics of the spring from the energy perspective. Additionally, we introduce a temporal‐state neural network that extracts temporal features and state features from all hidden features. These two types of features are fused with attention mechanisms for weighted integration. Temporal features represent those with clear dependency relationships on current and prior features, while state features dynamically highlight the subtle changes caused by spring failure based on the data characteristics. Experimental results demonstrate that the proposed method can effectively identify abnormal spring failures in engine‐driven pump systems.