Separation of Partial Discharge Sources by Application of Feature Extraction and Clustering Methods

This study presents an automated methodology for the separation of PD events arising from different sources, as well as from disturbances. This approach enables the construction of distinct PRPD patterns for each identified source type, thereby reducing signal overlap and facilitating a more straightforward and accurate interpretation.

The method leverages spectral characteristics to distinguish PD pulses associated with defect mechanisms both from each other and from those caused by disturbances. In a reduced feature space based on spectral content, a mixture modeling algorithm is employed to cluster the pulses into separable groups, each corresponding to a potential PD source or disturbance contribution.

Experimental validation is conducted under controlled laboratory conditions. The methodology is applied to measurements obtained from various specimens, including a high-voltage rotating machine, as well as configurations consisting of a capacitor and a wire to generate void and corona discharges. The clustering results demonstrate the separation of PD signals associated with individual defect types and disturbances.