【作者】 李中；

【导师】 苑津莎；

【作者基本信息】 华北电力大学（河北） ， 电工理论与新技术， 2010， 博士

【摘要】 大型电力变压器是电力系统的枢纽设备之一,其运行状况直接影响着电力系统的安全、可靠运行,其故障诊断技术研究具有十分重要的理论和实际意义。故障诊断可视为一个模式识别问题,也是近年新兴的人工免疫系统的一个重要应用领域。据此,论文从模式识别技术入手,重点对相似度及其测量、聚类、人工免疫系统模型和算法进行了深入研究,国际标准数据测试集上的测试结果验证了论文所提方法的可行性和有效性,应用所提方法完成了基于油中溶解气体分析的电力变压器故障诊断。论文在相似度评估、人工免疫系统模型和算法等方面取得了一定突破,取得的创新性成果主要有:(1)基于向量间差值的特性,提出了新的相似度测量方法。论文根据向量间差值与对象形状相似间的近似关系,设计了向量形态参数,结合经典的欧几里得距离,提出了两种考虑形状相似的形态相似距离,用于相似度的测量。形态相似距离计算简便,大量随机数据集聚类结果的统计分析表明,形态相似距离能够从大小和形状两个方面进行相似度的评估。多个UCI标准数据集和一个真实数据集的识别以及聚类实验表明,形态相似距离对于包含大小与形状信息向量的相似度测量,具有很高的准确度。(2)提出了一种抗体生成算法。在生物免疫系统中,抗体的超高速变异能力同遗传有关机理差异显著。据此,论文中提出的新算法没有采用通常人工免疫算法中大量应用的类似遗传算法的随机搜索和优化策略,而是依据抗体的浓度,区分不同情况,设计了抗体进化、抗体合并以及抗体新生三种不同的策略,快速提取和记忆抗原特征,有效地提高了算法的效率。(3)提出了一种具有自组织、自学习和自记忆能力的自组织抗体网络模型。模型基于生物免疫系统中抗体对抗原的快速学习与记忆能力,计算简便,只需根据样本数据的规模设置初始抗体个数,无需人工设置任何其他参数与阈值;模型中抗体类型与浓度的设计,有效提高了抗体的对抗原的学习和记忆能力,使该模型具有了数据分析的结构。多个UCI标准数据集上不同方法的仿真结果对比表明,相比其它智能方法,自组织抗体网络具有很高的识别准确率和数据浓缩率。(4)应用本文提出的自组织抗体网络模型和抗体生成算法,基于油中溶解气体分析数据,对电力变压器故障进行了分析。仿真结果表明该方法所获得的诊断准确率较高,可作为解决电力变压器故障诊断问题的一种新方法。更多还原

【Abstract】 A large electric power transformer is one of the key apparatus in the electric power system. Faults of a power transformer may have a great effect on stability of the power system. Therefore, there is great academic and engineering significance to do an earlier research on fault diagnosis technology of power transformers. To investigate this problem, this paper thoroughly studies the similarity measures, cluster analysis and artificial immune system (AIS), and achieves some breakthroughs on the theory of the similarity estimation and AIS. The proposed methods in this paper have been tested on some benchmark data sets from the UCI repository, and robust results are obtained. The innovative achievements are concluded as follows:1. Two kinds of distance measures to similarity estimation are proposed. The relation between the characteristic of differences and shape similarity is discussed, the Vector Shape Parameter (VSP ) is defined, merging the classical Euclidean distance, two new measures based on the analysis of the differences between vectors are presented, named as the Shape Similarity Distance (SSD) and the Morphology Similarity Distance (MSD) respectively. The FCM clustering results on many rand datasets show the new method can estimates similarity on both the size and shape of objects. A lot of classification and clustering results on some benchmarked datasets from the UCI repository and a real dataset conclude that, the presented method is one kind of similarity estimation measure which can achieve higher accuracy than the classical methods.2. One kind of antibody generation algorithm is proposed. In the biological immune system, antibody (Ab) has ultra-high-speed variation capacity with significant different mechanism of genetic system. From this point of view, the antibody generation algorithm does not use random search and optimization strategies like most of artificial immune systems used, but learn and memory the characters of antigen with three different strategies according to different situation: antibody evolution, antibody combination and antibody production. This method greatly enhances the efficiency of this algorithm.3. A self-organization, self-learning and self-memory named self-organization Antibody net (soAbNet) is suggested. The soAbNet is inspired by reinforcement learning and immune memory ability of antibody to antigen. It is easy to calculate, only need to define the number of initial antibodies, without any other parameters and thresholds. The concentration of antibody is designed, and it effectively improve memory capacity of antibody, and the data analysis ability of this model. The proposed approaches have been tested on a variety of benchmark dataset from the UCI repository. In all the experiments, this method demonstrates effective performance compared with other methods.4. Faults diagnosis of power transformers is discussed based on the antibody generation algorithm and soAbNet. The Dissolved gases analysis experimental results show that this method has higher accuracy, this provides a new approach to solve faults diagnosis of power transformers problem.更多还原