An Improved One-step Onsite FTU Health Condition Assessment Method Using Semi-supervised Graph Discrepancy Learning

The data-driven health condition assessment (HCA) methods of Francis turbine units (FTUs) have become an important part of hydropower unit prognostics and health management (PHM). Among them, some multi-step interdependent supervised models are designed to perform the assessment task. However, in the industrial scenarios, the status labels of data are rare and difficult to obtain, and the multi-step interdependent models will greatly increase the unreliability and inefficiency of the method. To make full use of rare status labels, it is necessary to model and excavate the spatial-temporal relationships hidden in the data. In this paper, an improved one-step onsite FTU health condition assessment method using semi-supervised graph representation discrepancy learning is proposed. First, through KNN algorithm, a small amount of labeled data are used to light up a large number of unlabeled data to construct graphs. This self-supervised method converts the health state data and the degradation state data into the graphs respectively, which improves the sample's state representation ability. Then, by using NT -Xent Loss to maximize graph discrepancy, the graph convolutional networks are trained as a graph representator, which reads the degradation graph and output the quantitative assessed degradation values. Therefrom, an end-to-end assessment model based on graph representation discrepancy learning is proposed to output quantified PDI directly from the input monitoring data. Comparative experimental results show that the proposed method achieves a competitive result.