Digital Filtering Methods for Interferences on Partial Discharges Under DC Voltage

In this contribution, a new separation method for partial discharge signals and disturbances is introduced. In contrast to frequency rejection filter techniques like low or high-pass filters as de-noising methods, a correlation filter is used to separate the acquired signals into different signal classes. For that, a test circuit for the measurement of partial discharge signals under dc voltage was set up. Two test samples for the generation of corona and surface discharges as typical partial discharge sources were used to obtain test signals for the presented classification algorithm. Besides partial discharge signals, additional disturbances from the power supply, which was realized by the application of a half-wave rectifier, were also measured. Spectral analyses of the measured signals show the difficulties when using common frequency-rejection filter methods since the relevant spectral contents of discharge signals and disturbances are located in the same frequency range. An application of these filter methods for de-noising would influence the waveform of the partial discharge signals and impede further analyses due to a limitation respectively a loss of their relevant spectral contents. The new approach provides a separation method for different signals without an influence on the pulse shape by using a correlation filter method. Histogram analyses of the correlation coefficients show a clear differentiation of the signal classes and present thresholds for the distinction between correlated and uncorrelated events. The consistence of each single signal class is proven to ensure a high similarity of the events obtained from the same test sample. The correlation between two signal classes lead to coefficients below the determined threshold value, which indicates a comparison of events from different classes. In a last experiment, the results from the histogram analysis were implemented in a classification algorithm to separate the raw data into specific classes and presents a new approach for the grouping of unknown signals in comparison to multi-channel processing tools like the 3 Center Frequency Relation Diagram (3CFRD).