Robust Feature Extraction and Classification of Acoustic Partial Discharge Signals Corrupted With Noise

Abstract

Partial discharge (PD) can be used as an indicator of impending failure in electrical plant insulation making the accurate classification of particular occurrence patterns useful for anticipating forthcoming outages. In this paper, we propose a feature extraction method that is robust to noise and can effectively select the most discriminant features of PD signals. Specifically, we follow three main steps. First, the spectrum of the PD signals is obtained using fast Fourier transform. Then, the low-frequency components are truncated and selected as PD representative features. Finally, these features are fed to the classifier and the detection accuracy is evaluated. In this paper, we consider the classification problem between three different types of acoustic PD signals, which are sharp, surface, and void PDs. Eight different classification models are adopted to test the PD detection accuracy along with the proposed scheme. Results on a benchmark data set illustrate the effectiveness of the proposed method on PD detection, while it yields a 100% classification accuracy (CA) for noise-free PD data. The robustness of the proposed method is also verified, where it achieves a CA up to 95.98% and 99.62% for noisy PD signals contaminated with high level of white and random noise, respectively. Furthermore, the proposed method is applied to actual PD signals corrupted with real noise; a CA between 98.16% and 98.64% is achieved.