A novel dual-knowledge embedded graph convolutional network method combining knowledge quantification for gas turbine gas path analysis

Data-driven algorithms provide a powerful method for rapid gas path analysis (GPA) of gas turbines. The conventional GPA models based on only sensor data could be improved in the generalization capability. Incorporation of knowledge offers a potential solution to enhance the diagnosis performance of data-driven GPA models. This paper proposes a novel Quantified-Knowledge Guided Graph Convolutional Network (QK-GCN) model to enhance the physical interpretability and generalization. The QK-GCN model embeds dual forms of knowledge: firstly, the graph topology is constructed from the physical relations among 9 gas path sensors, identified through reachability analysis of the thermodynamic equations; secondly, the corresponding edge weights in the graph are assigned by quantifying the relation similarities between sensors via knowledge representation learning technique. The proposed QK-GCN model is developed and verified using the annual field data from a GE 9F-class gas turbine. The proposed model achieves a lower overall error value and a narrower error distribution in all comparative studies. Embedding relation similarity as the GCN edge weight improves the root mean square error by 15.3 % for compressor efficiency degradation (DEC) and 33.5 % for flow rate degradation (DGC). Furthermore, the proposed QK-GCN model maintains strong generalization capability even with a 30 % training data size. The results indicate that the proposed QK-GCN model provides a reliable, generalizable, and interpretable solution for gas turbine GPA. The dual-knowledge method is promising for application to the performance prediction and fault diagnosis of other energy and power systems.