Partial Discharge Diagnosis in Gas-Insulated Switchgear Using a Fault-Mechanism-Enhanced Conditional Generative Adversarial Network
ID：50
Submission ID：54 View Protection：ATTENDEE
Updated Time：2025-11-04 13:34:10
Hits：111
Oral Presentation
Start Time：2025-11-09 10:35 （Asia/Shanghai）
Duration：15min
Session：[S3] 3. Power system and automation High voltage and insulation technology » [S3] 3.Power system and automation High voltage and insulation technology
No files
Abstract
Data-driven insulation defect diagnosis models aim to extract health information of gas-insulated switchgear (GIS) from large-scale monitoring data, and have achieved promising progress in recent years. However, in practical operating environments, the available fault monitoring data are often limited in scale and diversity. Moreover, most existing models lack the integration of physical fault mechanisms, making it challenging to establish reliable models for real-world applications. To address these issues, this paper proposes a fault-mechanism-enhanced conditional generative adversarial network (FME-CGAN) for partial discharge (PD) diagnosis in GIS. First, based on GIS insulation failure mechanisms, the operating fault problem is modeled as a search space, where potential equipment states and fault causes are defined as states, and operational or decision actions are represented as state transitions. A Monte Carlo Tree Search (MCTS) is then employed to simulate fault logic graphs under specific states, which are further encoded into a multilayer perceptron to construct a logical verification model. Subsequently, a conditional generative adversarial network is used to generate synthetic fault samples, while the logical verification model is embedded behind the discriminator to perform anomaly detection on the generated data. This mechanism not only validates the generated samples against fault logic but also provides a new backpropagation pathway for hyperparameter optimization. Finally, the discriminator is leveraged to achieve accurate GIS PD classification. Experimental results demonstrate that the proposed FME-CGAN achieves a diagnostic accuracy of 98.93%, outperforming baseline models by over 5%. These results verify that the proposed method significantly enhances the reliability of GIS PD modeling and analysis while incorporating fault-mechanism constraints into the diagnostic framework.
Keywords
gas-insulated switchgear, partial discharge, fault-mechanism-enhanced, conditional generative adversarial network, Monte Carlo Tree Search.
Speaker
Comment submit