Comparison of Algorithms for Clustering of Partial Discharge Signals under DC Voltage

Under HVDC voltage application, the analysis of partial discharges shows significant shortcomings compared to measurements under AC voltage application. As due to the missing phase-angle information of the test voltage at DC, a clear differentiation of noise and partial discharge signals, as well as the differentiation of several partial discharge sources, is still a challenging task. Therefore, an approach for signal detection and clustering based on intra- and interclass correlation combined with histogram-thresholding was developed and tested by means of measured partial discharge signals. This clustering algorithm differentiates acquired signals automatically into different signal groups in order to allow further and separate investigation.As an alternative method, a k-medoids clustering as a well-known unsupervised learning technique, was tested on the measured signals. This method provides a fast and reliable performance as it is deterministic. This contribution shows the feasibility of the k-medoids algorithm applied on the signals of a partial discharge test under DC voltage application. A comparison of the histogram-thresholding clustering and the k-medoids algorithm points out the pros and cons and demonstrates which kind of clustering algorithm is the more appropriate solution. The results of this investigation can be considered for the development of a fully automated and unsupervised measurement system for partial discharge analysis under DC voltage.