Research on Recognition of Multiple Partial Discharge Sources in Switchgear Based on the Combination of GST-TEV and ResNet-18

Abstract

Switchgear can develop insulation defects due to electrical, thermal, and chemical stresses during manufacturing or prolonged operation. Moreover, as voltage levels rise, multiple insulation defects can coexist within the switchgear. Traditional partial discharge (PD) recognition methods often suffer from poor generalisation and low accuracy, limiting their practical applications. This paper proposes a method to identify multiple PD sources by combining generalised S-transform (GST) with the ResNet-18 network. PD tests confirm that the designed monitoring device effectively detects transient earth voltage (TEV) signals from diverse single and mixed insulation defects. Given the non-stationary nature of TEV signals, this paper employs the generalised S-transform (GST) for time–frequency analysis. The findings demonstrate that the GST method offers high time–frequency resolution, significantly improving the feature extraction of various partial discharge sources. Additionally, deep learning algorithms are employed to classify the time–frequency image dataset derived from GST-TEV. The results demonstrate that, compared to traditional manual feature extraction methods, the ResNet-18 network efficiently extracts GST-TEV features from both single and mixed partial discharge sources, achieving a recognition accuracy of 99.41%. This study provides new methods and theoretical support for identifying multiple partial discharge sources in switchgear.