Estimating Remaining Useful Life of Aircraft Engine System via a Novel Graph Tensor Fusion Network Based on Knowledge of Physical Structure and Thermodynamics

Accurate estimation of the remaining useful life (RUL) of aircraft engines is critical for aircraft health management and maintenance planning. To address such an issue, this article proposes a spatiotemporal graph attention tensor network (STGATN) based on knowledge of physical structure and thermodynamics. First, by utilizing engine sensor time-series data, we generate and construct an airflow state graph with thermodynamic knowledge and a structure state graph with structural layout knowledge. Then, by introducing a graph attention mechanism to extract spatial features of the two types of state graphs separately, and by proposing a tensor fusion module to embed and integrate the two groups of first-order feature vectors into a high-order tensor data. Furthermore, by designing the convLSTM layer to acquire temporal information of high-order tensor for accurate RUL prediction. Finally, experiments are conducted on the commercial modular aero-propulsion system simulation (CMPASS) dataset and the real engine test dataset. The comparative results show that our approach outperforms existing state-of-the-art methods.