Classification and separation of partial discharge signals by means of their auto-correlation function evaluation

This paper describes a K-Means Clustering classification algorithm for the separation of Partial Discharge (PD) signals and pulsating noise due to multiple sources occurring in practical objects. It is based on the comparison of the Auto-Correlation Function (ACF) of the recorded signals assuming that the same source can generate signals having similar ACF while ACF differ when signals with different shapes are compared. The ACF has been selected for its capability of well summarize both time- and frequency-dependent features of the signals. A correlation index that presents the best compromise between strong and weak discrimination among pulses, has been selected out of different distance measurements. The final result of the algorithm is a set of classes containing signals having similar shape which can be processed successively for signal source identification. Meaningful applications of the proposed algorithm are also reported. Improvements in separation effectiveness can enhance the clearness of the PD patterns and, consequently, the quality of the defect identification.