Heterogeneous Dynamic-Aware GNN for RUL Prediction of Aeroengine

Accurate remaining useful life (RUL) prediction is crucial for achieving predictive maintenance. The performance degradation process of equipment is, however, not only manifested as anomalies in individual variables but also as changes in the constraint relationships among multiple variables. Human operations and changes in operating conditions further complicate the state representation of the performance degradation process, making RUL prediction for complex systems like aeroengines a challenging task. This article proposes a novel graph flow learning model for RUL prediction in such complex systems with dynamic interactive operations, termed the heterogeneous dynamic-aware graph neural network (HDA-GNN). First, various correlation evaluation metrics are used to convert sensor signals into heterogeneous graphs to represent the complex relationships among sensor signals. Then, gated recurrent unit (GRU) and attention mechanisms are employed to dynamically calculate the edge weights of the heterogeneous graphs, and a graph convolutional network (GCN) is used to integrate the features of these dynamic heterogeneous graphs. Finally, the transformer is applied to learn the temporal features within the graph flows. This model is capable of learning dynamic features under human external operation disturbances from multiple perspectives, thereby reducing the impact of the randomness of human operations on dynamic features. Experimental results on the NASA commercial modular aero-propulsion system simulation (C-MAPSS) dataset demonstrate the effectiveness and superiority of this model for RUL prediction tasks in complex systems like aeroengines.