【作者】 陶加贵；

【导师】 唐炬；

【作者基本信息】 重庆大学 ， 电气工程， 2013， 博士

【摘要】 组合电器在运行电压、热、力等作用下的内绝缘时效老化和在生产、运输、调试装配、运行和维修过程产生或留下的各种潜伏性缺陷，会逐渐扩展致使内绝缘的电气强度下降而导致故障。由于前期潜伏性故障主要以局部放电的形式表现出来，因而通过对其准确辨识和危害性评估，能有效指导检修工作，制定更经济、更合理的检修策略，实现状态检修和设备全寿命周期经济管理。同时，鉴于多信息融合是监测技术的必然趋势和最终走向，论文改变了传统的单一传感器独立诊断的研究思路，基于DS证据理论建立了组合电器内部局部放电多信息融合辨识和危害性评估系统。论文以实验为基础，利用IEC60270法、荧光光纤检测法和超高频法建立了组合电器内部局部放电多传感器检测系统。通过大量试验测试，分析了组合电器内部高压导体金属突出物、绝缘子表面金属附着物、绝缘子内部气隙和自由金属微粒等四种典型绝缘缺陷下，三种传感器检测性能、特征信息的差异及随局部放电发展的变化规律，通过理论证明和实验验证探明了所使用的光检测法、超高频法的信号与放电量的相互关系。多信息源能从多个角度体现局部放电的属性。根据各源信息属性的差异，分别以时域、频域和db4为母小波的三层小波包分解提取了超高频时间解析信号的波形特征，作为辨识局部放电的细节信息；以偏斜度、陡峭度等统计参数提取了基于相位解析的放电指纹的分布特征；同时，结合超高频信号累积能量与放电量之间的相关性信息，基于DS证据理论首次建立了局部放电多信息融合辨识系统。通过对现场装置中的两个独立缺陷和样本数据的测试，验证了系统具有相对于独立传感器诊断更好的容错性、稳定性和更高的辨识率。针对辨识过程中，基于相位分析模式的特征量存有相关性而导致过高的计算代价和信息影响不均衡的问题，以支持向量机one-versus-one多分类算法为基础，结合“最佳脑损坏”理论，设计了全局最优排序准则，建立了多类支持向量机特征回归消去选择方法，通过排序选定其中9组作为放电指纹最优特征子集，并通过与主成分分析法的对比测试，验证了其具有相对较好特征降维效果。根据实验测试和现场运行经验，提出和定义了用于局部放电危害性评价的“正常、注意、警戒”三级状态。通过对三种传感器的检测信息及信息之间相关性的分析，结合组合电器相关检测规范和标准，选取IEC60270法所检测的放电量、光检测信号的最大幅值和放电指纹随着局部放电发展的变化规律特征作为危害性评估的主要源信息。针对三种特征信息与局部放电发展程度之间并非绝对的正相关关系的特性，结合试验中三种信息随着状态等级的变化规律，分别采用模糊隶属度分析法和神经网络建立了特征信息与局部放电危害程度之间的映射关系。最终，通过DS证据理论首次建立了局部放电危害性多信息融合评估系统。通过测试表明：①对于相似结构的缺陷模型，三种独立信息都能较好的反映局部放电的严重程度；②准确辨识局部放电类型有利于明细特征信息计算方法，提高信息挖掘的合理性和可靠性及基本概率分配设置的客观性和容错性；③多信息融合能有效综合三种方法的评估信息，相比独立信息评估具有更佳的评估性能、更高的准确率和可靠性。更多还原

【Abstract】 Gas insulated switchgear has inner insulation aging under the effect of voltage,heat and forces, in addition with certain hidden defects caused by producing, transportand operation, all of which will decrease the inner insulation voltage and then resultedin electric faults. Since the hidden fault appears mostly in the form of partial discharge,its accurate recognition and harmfulness assessment can give instructions for statedetection and life cycle economic management. In view of the advantage ofmulti-information fusion, the thesis constructs information fusion recognition andharmfulness assessment system of gas insulated switchgear based on DS evidencetheory, with the disregard of the traditional single sensor detection.The thesis proposes a multi sensor detection system of gas insulation switchgearwith the application of IEC60270method, fluoresense optical fiber detection and ultrahigh frequency. The experiments in a large number give three kinds of sensors detection,feature information differences and variations with partial discharge, under the defectsanalysis of metallic outshoot in gas insulation switchgear, insulator surface metalattachment, insulator inner gap and free metal particles, which demonstratecorresponding relations between three kinds of sensors and partial discharge both intheory and application.Multi information sources can give the representation of partial dischargecharacteristics in multi perspectives. According to the differences of the multiinformation characteristics, waveform feature of ultrahigh frequency time analyticsignals is extracted by wavelet decomposition in time, frequency and db4domains.Distribution feature of discharge fingerprint based on phase resolved analysis isextracted using skewness, measurable slop and other statistical parameters. Combinedwith relative information between ultrahigh frequency signals accumulation energy anddischarge amount, the partial discharge multi information fusion recognition system isconstructed based on DS evidence theory, with the demonstration of better faulttolerance, stability and recognition accuracy than independent sensor detection, underthe testing of independent defects and sample data.In the process of identification, taking the unbalanced effect from the correlation ofPRPD features into accounts, a global optimum sort criteria is designed to establish thesupport vector machine recursive feature elimination method, based on theone-versus-one Support Vector Machine classification algorithm and “Optimal Brain Damage” theory. Feature sets are optimized by principle component analysis andsupport vector machine recursive feature elimination, nine which are chosen as theoptimal feature subsets, with the demonstration of a relatively better effect for reductionof the feature dimension benefited from the SVM-RFE method.In accordance with the experiments testing and field operation experiences, threelevels of operation states are proposed and defined, namely normal, be aware of, and onalert, for the assessment of partial discharge harmfulness. With the combination of gasinsulation switchgear operation experience and relating detecting regulation, and underthe analysis of three kinds of sensors information and their correlation, the dischargeamount by IEC60270, maximal optical detection signal and variation of dischargefingerprint under partial discharge, are viewed as main principles of harmfulnessassessment. Based on the fact that the three kinds of feature information are notabsolutely positively correlated with the development degree of partial discharge, andcombined with the variation regulation of the three information in company with thestate degrees, the mapping relation between the feature information and the damagedegree of partial discharge is constructed, under the application of the fuzzymembership analysis and neural network respectively. Finally, the thesis constructs amulti information fusion assessment system of partial discharge harmfulness throughthe DS evidence theory originally.With the testing demonstration, it follows that:①All the three kinds ofindependent information can give a proper reflection on the order of severity of partialdischarge, with respect to defect models in similar structures.②The accuraterecognition of partial discharge types can give benefit to feature information calculationin detail, enhance the rationality and reliability of information mining, and improve theobjectivity and fault-tolerance of basic probability assignment settings.③Multiinformation fusion can provide a comprehensive treatment for the assessmentinformation in three kinds of methods, and can obtain better assessment performanceaccuracy and reliability, in comparison to the independent information assessments.更多还原