報告書番号

R10030

タイトル（和文）

油中ガス分析データによる電力用変圧器の様相判定手法

タイトル（英文）

Transformer Condition Classification based on Gas-In-Oil

概要 （図表や脚注は「報告書全文」に掲載しております）

電力会社では電力用変圧器の内部状態監視のため，油中ガス分析により内部不具合の診断を行っており、判定精度の向上に向けて継続的に検討が進められている。本研究では，2009年度に電気協同研究会が刊行した報告書「電力用変圧器改修ガイドライン」に付記されている油中ガス分析データに対して，変圧器の様相（「正常」，「過熱」，「放電」，「過熱＋微小放電」，「絶縁油混入」の５つの様相）を総て正しく判定できる判定手法を開発した。本手法は，9種類のガスの対数の2次多項式の正負に基づき対象様相か否かを判定する。未知のデータに対する判定精度の推定方法であるLeave-One-Outによる判定手法の推定判定精度は99%以上である。

概要 （英文）

For efficient maintenance of power transformers, Japanese electric power companies use analysis of dissolved gas-in-oil. Electric technology research association (ETRA) in Japan updates the criteria of analysis of gas-in-oil approximately every 10 years. The ETRA report on the most recent update in 2010 describes that the linear support vector machine can generate a linear discriminant to judge the existence of inner faults based on the data of gas-in-oil.. However, it cannot judge the types of faults (overheat, discharge, discharge-with-overheat, oil-incorporation-from-LTC). In this report, we propose a new method which can judge correctly the condition (non-existence of inner faults and types of faults) of all the data attached in the appendix of the ETRA report. It generates one quadratic discriminant for each condition based on the log of 9 gases in oil, because of the fact such that the log of composition of rate of gases are linear with heating temperature and that the data in the same condition form a cluster as vectors. To determine the coefficient of quadratic discriminant, we use linear support kernel machine (which is an extension of SVM and a kind of multiple kernel learning.) to make the discriminant more robust and to reduce their non-zero terms. By the leave-one-out cross validation, the largest judgment error rate is 1.1%. Therefore we expect that the quadratic discriminants by our new methods can judge new data highly precisely.

報告者

キーワード