Dynamic Feature-Aware Graph Convolutional Network With Multisensor for Remaining Useful Life Prediction of Turbofan Engines

Multiple sensors have been employed to monitor the entire operational cycle of turbofan engines to enhance their reliability. The data collected from multiple sensors comprehensively reflect the operating status of a turbofan engine. However, multisensor systems may gather redundant information. In addition, the dynamic nature of turbofan engine operations results in changing optimal sensor groups for accurately reflecting the remaining useful life (RUL) and these dependencies among sensors. To tackle the challenges, this study proposes a dynamic feature-aware graph convolutional network (DFAGCN), which is designed to construct dynamic dependencies among sensors and dynamically select the optimal sensor group, thereby improving the accuracy of predicting the RUL. DFAGCN is divided into three parts. First, the dynamic feature-aware (DFA) module combines attention mechanisms and cosine similarity to explore the dynamic dependencies among sensors at different timestamps, and the optimal sensor group is selected based on the attention weights and degree centrality of the sensor nodes to form graph embeddings that accurately represent the degradation features of the equipment. Second, the gated recurrent unit (GRU) and graph convolutional network (GCN) are employed to learn the spatiotemporal feature representation of graph embedding. Finally, the regression layer is used to learn the mapping relationship between the feature representation and the RUL. The experimental results on two turbofan engine degradation datasets, C-MAPSS and N-CMAPSS, show that the proposed method can effectively select the optimal sensor group, extract dynamic dependencies among sensors, and achieve superior performance compared to other state-of-the-art (SOTA) methods.