Research on gas turbine health assessment method based on physical prior knowledge and spatial-temporal graph neural network

Health management plays a significant role in preserving the reliability and safety of gas turbines. An accurate assessment of the health status of gas turbines is critical for the realization of predictive health management. However, existing health assessment methods do not take into account the physical relationships and spatial states within the system. The focus of this study is to establish a methodology that can concurrently utilize physical relationships and data for the quantitative assessment of gas turbine health. This study proposes a novel Physical Spatial-Temporal Graph Convolutional Network (Phy-STGCN) approach that combines prior physical knowledge with data-driven techniques by incorporating the structural and operational mechanisms of gas turbine into a graph-based model. First, the concept of gas turbine health is defined, followed by the validation of the rationality behind health classification. Second, a temporal graph construction method based on K-nearest neighbor and prior physical knowledge is introduced. Third, a network architecture based on graph neural networks is proposed to incorporate the temporal and spatial dependencies present in the data. The proposed method is validated using gas turbine data, achieving a health assessment accuracy of 90.8%. At the initial stages of health degradation, the assessment accuracy can exceed 98.9%. Through a series of comparative and ablation experiments, the efficacy of the Phy-STGCN method in gas turbine health assessment is further substantiated. These results demonstrate that the proposed method can effectively leverage prior physical knowledge and the spatial coupling information between data, realizing a quantitative mapping from multi-source monitorable data to system states. This study provides insights for research on health assessment methods that integrate physical and data-driven approaches.