Multi-scale spatiotemporal feature-assisted physical information graph temporal convolutional network for aero-engine degradation trend prediction

Continuous degradation trend prediction (DTP) of aero-engine is crucial to ensure the safe and reliable operation of aircraft systems. Recently, data-driven deep learning methods have been extensively developed in the field of aero-engine DTP. However, the multi-sensor data obtained from actual measurements have complex temporal and spatial dependencies. Moreover, purely data-driven aero-engine DTP methods lack the guidance of aero-engine physical mechanisms. To address these challenges, a multi-scale spatiotemporal feature-assisted physical information graph temporal convolutional network (ST-PGTCN) is proposed for the DTP task of aero-engine. First, ST-PGTCN designs a method that can capture the spatiotemporal dependencies between multiple sensors to construct spatiotemporal feature graph. Then, graph temporal convolutional network is developed based on graph convolutional neural network and temporal convolutional neural network to model the degradation trends in spatiotemporal feature graphs. Finally, a multi-penalty loss function consisting of an improved delay prediction accuracy penalty and an aero-engine thermodynamic physics information penalty is proposed to enhance the practicality and physical consistency of the prediction results of ST-PGTCN. The proposed ST-PGTCN is evaluated on the N-CMAPSS dataset. Compared with other conventional loss functions and state-of-the-art DTP methods for aero-engine, ST-PGTCN shows superior prediction performance. The prediction results also reveal the feasibility and reliability of integrating aero-engine knowledge and deep learning models.