A Novel Algorithm for Separating Multiple PD Sources in a Substation Based on Spectrum Reconstruction of UHF Signals

Field noise interference suppression and effective extraction of signal characteristics are two keys to partial-discharge (PD) signal detection and analysis. In this paper, the autoregressive moving average (ARMA) process is utilized to model ultra-high-frequency (UHF) signals radiated by PDs. The estimation, which is based on high-order cumulants, of the ARMA orders and parameters is given theoretical analysis and implementation. The spectra of the detected signals are reconstructed with the model estimated. Then, characteristic frequencies are selected based on the reconstructed spectra and Fisher-like class separation measures. The radial basis function neural network is trained for the separation of the observed signals. Using the proposed method, UHF signals generated by electromagnetic simulation software are efficaciously modeled, reconstructed, and separated from mixing Gaussian white noises of varying signal-to-noise ratios and fixed-frequency signals. Finally, the step to obtain the number of PD sources in the assumed substation is proposed. UHF signals collected in a substation are processed by the proposed procedure, and PD sources are separated. This separation result is compared with PD sources localization results calculated by the time delay sequence, and the effectiveness of the method in the substation field interference circumstances is verified.