Abstract
The measurement of the emitted electromagnetic energy in the UHF region of the spectrum allows the detection of partial discharges and, thus, the on-line monitoring of the condition of the insulation of electrical equipment. Unfortunately, determining the affected asset is difficult when there are several simultaneous insulation defects. This paper proposes the use of an independent component analysis (ICA) algorithm to separate the signals coming from different partial discharge (PD) sources. The performance of the algorithm has been tested using UHF signals generated by test objects. The results are validated by two automatic classification techniques: support vector machines and similarity with class mean. Both methods corroborate the suitability of the algorithm to separate the signals emitted by each PD source even when they are generated by the same type of insulation defect.
Highlights
One of the causes of power failures and blackouts is the breakdown of the insulation systems of electric assets, produced, in some cases, by their deterioration
The weights, Wk, for k = 1, . . . 50, are initialized to zero. Both the learning rate and the number of iterations are obtained empirically based on previous experimental measurements done in the radio-frequency localization of partial discharge (PD) sources [12]
In order to corroborate the suitability of the independent component analysis (ICA) algorithm reliably and systematically, we trained supervised classifiers to discriminate the signals coming from the two PD sources and to use them to separate the outputs of the ICA
Summary
One of the causes of power failures and blackouts is the breakdown of the insulation systems of electric assets, produced, in some cases, by their deterioration This aging can be premature and originates due to small and persistent discharges, called partial discharges (PDs). A validation technique based on a supervised classification effectively confirms that the signals recovered by ICA accurately match those originating from the sources. This validation was carried out using two automatic classification techniques: support vector machines (SVMs) [13,14] and a naive approach consisting of classifying each datum with the class of the closest mean.
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call