Remaining Useful Life Prediction Method for Turbofan Engines Based on Graph Neural Networks

Accurate remaining useful life (RUL) prediction is crucial for assessing the health status of turbofan engines and preventing potential failures. In turbofan engine RUL prediction tasks, multiple sensors are required to monitor operating conditions. However, the current mainstream RUL prediction models for turbofan engines have limitations in capturing the spatial correlations among sensors. To address these limitations, this article proposes a novel graph neural network (GNN)-based RUL prediction method for turbofan engines, named adaptive dynamic fusion GNN (ADFG). This method combines the structural characteristics and monitoring data of turbofan engines to effectively model the spatial correlations among sensors. First, a multilevel hidden feature extraction method based on the multiscale convolutional neural network (MSCNN) structure is proposed, and the convolved data are used as node features to construct a more comprehensive sensor dynamic interaction relationship. Second, a dynamic graph construction method was designed to capture the potential dynamic relationships during sensor degradation in real time. Based on this, an adaptive fusion mechanism is proposed to fuse the correlations between fixed prior and dynamic similarity graphs. Finally, after updating the spatial features of the sensors through graph attention network (GAT) aggregation, an attention mechanism (AM) is utilized to fuse the local and global features of the sequences, and the temporal dependencies among the sequences are captured using bidirectional gated recurrent unit (BiGRU) to improve the prediction performance of the model. The effectiveness of the method proposed in this article is validated by the commercial modular aero-propulsion system simulation (C-MAPSS) dataset.