An Unsupervised Domain Adversarial Partial Discharge Recognition Model Based on Dual Classifiers for SF6 and C4F7N/CO2/O2 GIS

Accurately identifying the types of partial discharge (PD) in gas-insulated switchgear (GIS) is very important for the stable operation of power grid. As environmentally friendly gas, C4F7N/CO2/O2 is expected to replace SF6 in GIS in the future. However, when the existing PD recognition models is used for C4F7N/CO2/O2 gas mixture, there is a significant drop in recognition accuracy. This paper compares the PD signals of SF6 gas and C4F7N/CO2/O2 gas mixtures in time and frequency domain, and the high-dimensional features extracted by the encoder are visualized. Furthermore, a dual-classifier unsupervised domain adversarial (DCUDA) PD recognition model is proposed. Specifically, DCUDA model is composed of three main components: an encoder, two classifiers, and a domain discriminator. DCUDA model employs adversarial training to maximize the extraction of share features between the source and target domain, requiring only a small amount of C4F7N/CO2/O2 gas mixtures PD data for training. Experimental results show that DCDUA model achieves recognition accuracies of 99.99% and 99.81% in SF6 gas and the C4F7N/CO2/O2 gas mixture respectively when ultra-high frequency (UHF) signal and ultrasonic (UL) signal are mixed input, providing significant theoretical and technical references for future PD recognition of C4F7N/CO2/O2 gas mixture GIS.