Partial discharges and noise separation using spectral power ratios and genetic algorithms

Accurate measurements of partial discharge (PD) activity is essential for the application of this technique to condition-based monitoring. Noise and PD source characterization is necessary to fulfil that goal, since the interpretation of classical phase-resolved partial discharge (PRPD) patterns is usually complex for the measurements done in field. A successful pulse source separation prior to the identification seems to be the best option. The authors proposed in a previous work a method based on spectral power ratios (PR) to separate pulse sources with quite good experimental results. This technique calculates the spectral power in two frequency bands to obtain two parameters which, represented in a 2-dimensional map (PR map), characterize each pulse source by a cluster of points. The main difficulty of this technique is the choice of the appropriate frequency intervals that give a good separation of clusters, which sometimes can be cumbersome by manual means. Thus, this paper presents an unsupervised technique to select the two frequency intervals that gives the best separation among several clusters. This will give a great support for the system user to separate PD and noise sources in real measurements. The authors used genetic algorithms (GAs) to select these frequencies, with good results in several real experiments.