To enhance the reliability of hydropower generation, extensive research has been conducted on the health condition assessment (HCA) of Francis turbine units (FTUs). Common HCA approaches involve establishing health benchmark models (HBMs) and constructing performance degradation indicator (PDI). However, previous studies have the following limitations: (1) In feature extraction, PCA is commonly used for dimensionality reduction. However, it only supports linear transformations and has limited ability to handle complex nonlinear features. (2) Euclidean distance or cosine similarity is often used to measure node distances but tends to overlook the probability distribution characteristics of feature samples. To address these issues, this paper proposes a sparse probability feature graph construction method for FTUs’ HCA. Initially, a simulation model of the FTUs is developed, utilizing computational fluid dynamics theories to produce simulated pulsation signals. Subsequently, dictionary learning technology is introduced to perform sparse feature extraction, generating a feature matrix. Based on this, considering the probability distribution of features, the Wasserstein distance is employed to compute the distances between nodes, resulting in the sparse probability feature graph. To assess the operating state of FTU, T-SNE is used to extract the comprehensive feature of health label and degraded data. Then, the PDI is calculated by deriving from the comprehensive health labels and comprehensive degraded data. Validation experiments are conducted to verify the effectiveness of proposed method.

Francis Turbine Units,Health Condition Assessment,Graph Representation Learning,Dictionary Learning