电工技术学报2025，Vol.40Issue(21)：6856-6870,15.DOI:10.19595/j.cnki.1000-6753.tces.241857
基于可解释人工智能的SF6/N2气体直流火花放电光谱解析与临界闪络缺陷识别
Spectral Analysis and Critical Flashover Defect Identification of DC Spark Discharge in SF6/N2 Gas Based on Explainable Artificial Intelligence
摘要
Abstract
Currently,gas-insulated switchgear(GIS)employing an SF6/N2 gas mixture with a volume ratio of 3∶7 is considered a highly viable technical solution.This is attributed to its optimal balance among dielectric performance,cost-effectiveness,and environmental impact.To further reveal the microscopic physicochemical reaction mechanisms during discharge and improve both insulation design and fault detection strategies,this study investigates the spectral characteristics of spark discharges in SF6/N2(3∶7)mixed gas environment under direct current voltage. Initially,experimental platforms simulating partial discharges were established within a sealed chamber containing SF6/N2(3∶7)mixed gas at 0.6 MPa.The test specimens consisted of scaled electrode models representing insulator surface contamination,conductor protrusion,and floating metallic particles,and were constructed using materials analogous to those found in actual GIS components.The study employed the QE65 Pro high-resolution spectrometer from Ocean Optics with synchronized triggering to capture instantaneous discharge spectra during spark events,while a Lumina series UV-visible imaging system simultaneously recorded the visible discharge process. Subsequently,this research developed an integrated analytical framework combining quantum optical analysis with artificial intelligence-based data processing.An explainable AI model was designed that identifies key spectral features characterizing different spark discharge types.These features were then visualized through attention heatmaps,assisting experts in interpreting discharge mechanisms based on quantum optical principles.This approach addresses limitations in deep learning models where decision basis lacks interpretability,enabling researchers to identify subtle spectral features that may not be directly perceptible to the human eye. Moreover,building upon the identified spectral features,an investigation of reaction mechanisms was conducted.The energy-level transitions of key elements were mapped and correlated with discharge conditions.This analysis revealed several critical physicochemical processes:the decomposition of SF6 into reactive fragments,nitrogen excitation pathways,electrode materials'metal surface sputtering effects,and the decomposition processes of hydroxides and sodium compounds on insulation surfaces.And the influence of these reactions on discharge development was analyzed.Studies of maximum energy transfer thresholds provided insights into discharge-induced damage mechanisms and material tolerance limits.These findings support the development of optimized insulation solutions with enhanced discharge withstand capabilities. Based on spectral lines identified through theoretical and database analyses,a multi-objective optimization algorithm was employed to select optimal diagnostic bands for defect type identification.Using these optimized spectral bands,we designed and constructed a novel multi-spectral sensor prototype incorporating micro-structured side-emitting optical fiber sensing elements.We validated the sensor's performance using high-power sources that simulated critical flashover conditions,achieving over 98%accuracy in distinguishing between three typical GIS defect types. This research has established a workflow that organically combines artificial intelligence with human analysis.The approach can guide GIS insulation design optimization,fault prevention,and operational maintenance.Based on these findings,future work can simultaneously advance the optimization of key insulation components,such as insulator materials and protective measures for metallic parts,and develop high-sensitivity multi-spectral online monitoring systems suitable for practical and complex GIS field environments.关键词
局部放电/光谱解析/能级跃迁/多目标优化/可解释人工智能/微结构侧发光光纤Key words
Partial discharge/spectral analysis/energy level transition/multi-objective optimization/explainable artificial intelligence(XAI)/micro-structured side-emitting optical fiber分类
动力与电气工程引用本文复制引用
冯宇宁,苑舜,蔡志远,黄翀阳,高磊..基于可解释人工智能的SF6/N2气体直流火花放电光谱解析与临界闪络缺陷识别[J].电工技术学报,2025,40(21):6856-6870,15.基金项目
辽宁省教育厅基本科研项目资助(LJKMZ20220472). （LJKMZ20220472）
