I. Introduction
Detecting and identifying partial discharges (PDs) in gas- or oil-insulated switchgear and transformers with metallic enclosures are well-established procedures [1]. In traditional methods, the PD sensors are mounted inside the metallic enclosure which promises a high signal-to-noise ratio (SNR). But for operational switchgears and transformers without such PD sensors, arranging a shutdown specifically to fit internal couplers can be rarely justified [1]. So external nonintrusive sensors, which are easy to install and do not have operation interruptions, are becoming more popular for the devices which are not suitable to install internal sensors. Usually, the radiating electromagnetic wave from the PD source induces and forms a small pulse-like voltage on the metal tank surface. This is the so-called transient earth voltage (TEV) [2], [3]. The nonintrusive PD sensors detect those pulse-like TEV signals to determine PD existence. When detecting PDs on the external surface of the enclosure, one of the major problems that needs to be addressed is interference from the surroundings. Researchers have proposed many approaches [4]–[6] for the noise reduction. But difficulties still exist due to the following challenges:
Very low SNR. Reference [3] points out that the amplitude of TEV is proportional to the original PD and decreases with the distance from the PD source. When measured on the surface of metal cladding in field tests, the TEV signal is small that noises cannot be ignored. The noises from electronic equipment and the radio-frequency disturbances are believed to be major kinds of interference [7]. Currently, some thresholding methods, such as wavelet analysis, have performed well in PD extraction. But wavelet analysis could only divide the frequency range into several bands. This segmentation is not accurate enough to display the slight energy variations when PD's energy is smaller than noise in most frequency bands.
Lack of datasets for artificial-intelligence (AI)-based methods. In many PD denoising methods, AI has been used to recognize PDs and other pulse-like noises [8], [9]. A basic premise for these AI-based methods is the proper collection of a database with enough size. However, such a large database is almost impossible to collect in most field applications.