Data-model interaction-driven transferable graph learning method for weak-shot onsite FTU health condition assessment

Abstract

The large amount of monitoring data provided by the onsite hydropower unit has promoted the development of data-driven Francis turbine unit (FTU) health condition assessment (HCA) technology. However, these methods are usually trained in fully annotated source domains and applied to sparse onsite scenarios, leading to weak-shot learning problems. To fully explore the potential state representation from the source domain annotated by the mechanism simulation model, an innovative knowledge graph-based data-model-interaction framework is proposed for solving weak-shot onsite FTU health condition assessment. First, based on the selected critical onsite monitoring data, the pseudo-data obtained from the computational fluid dynamics calculation of the mechanism digital-twin (DT) model are used to fully annotate source domain. Secondly, the mixed pseudo-actual data are converted into graphs by node similarities to capture the correlations between signals. The explicit edge connection relationships in the graph structure allow state sharing across domain nodes and suppress loss of accuracy due to differences in domain distribution. Then, a transferable graph constructor with cross-domain parameter sharing is designed to learn knowledge-based construction strategies from the fully annotated source domain. Further, the transfer of state knowledge from theoretical domain to actual domain can further strengthen the sample’s representation in weak-shot domain. Finally, a graph-driven health benchmark model (HBM) is constructed to excavate the reconstruction-enhanced knowledge graphs, achieving FTU state presentation and degradation assessment. The proposed method has been applied in a dataset collected from onsite FTU, which not only achieves the best performance in multiple SOTA comparison tests, but also has an acceptable time consumption (5.24 s/100 graphs), and has the possibility of industrial field deployment.