Improved DGA method based on rules extracted from high-dimension input space

Abstract

The diagnosis of incipient faults in power system elements such as transformers is usually based on the concentrations of dissolved gases existent in the insulation oil. There are consolidated DGA-based (dissolved gas analysis) methods in the literature, such as the Duval triangle. However, they present some limitations such as the existence of non-decision areas and erroneous results. Proposed is a simple methodology to improve the analysis of incipient faults based on rules extracted from a high-dimension space (21 attributes), formed by the gases concentrations and some of their interrelations. From such input space, the C4.5 method (decision tree) is used to extract a set of interpretable rules. Databases known in the DGA technical literature such as IEC TC 10 are adopted to analyse the proposed approach. When compared with a standard method, considering all data test folders in the performed 10-folder cross-validation statistical analysis, the extracted rules show greater accuracy with an error in the diagnosis of incipient faults of 6.25%, against 18.75% for the Triangle method in the worst case.