Fault Tracing on Transformer Substations via Graph Kernel on Event Graphs

Abstract This paper targets fault tracing of transformer substation devices. Traditional AI-inspired methods have the potential to save vast amount of manpower. But there are two challenges they face: inadequate representation of connection topology of substation devices and the lack of interpretability of the fault tracing process, both of which are highly valued in industrial practices. Thus, this paper constructs a framework to address the problem of fault tracing by preforming graph kernel computations on specially designed event graphs. In addition, we propose an Edge-Labeled Weisfeiler Lehman Graph Kernel (EL-WL), integrating edge information into tree distance calculation of the graph kernel. Besides, we use a variant of unique class association rule mining algorithm performed on subtree histogram vectors to mine interpretable fault tracing rules. Experiment results showed noticeable performance improvement on real-world data.