Dissolved gas analysis method based on novel feature prioritisation and support vector machine. Issue 8 (1st September 2014)

• Record Type: Journal Article
• Authors: Wei, Chenghao; Tang, Wenhu; Wu, Qinghua
• Is Part Of: IET electric power applications. Volume 8:Issue 8(2014)
• Page Start: 320
• Page End: 328
• Publication Date: 2014-09-01
• Subjects: feature extraction -- support vector machines -- power engineering computing -- particle swarm optimisation -- power transformers -- fault diagnosis; dissolved gas analysis method -- feature prioritisation -- DGA -- incipient faults detection -- oil‐filled transformers -- feature classification -- international gas interpretation standard analysis -- Kolmogorov‐Smirnov test -- maximum statistic distance -- multilayer support vector machine -- tuning parameters -- particle swarm optimisation -- bootstrap technique -- 10‐fold cross‐validation technique -- artificial intelligence classifiers -- gas feature extraction -- genetic programming; Electric power -- Periodicals; Electric power systems -- Periodicals; 621.305
• DOI: 10.1049/iet-epa.2014.0085
• Languages: English
• ISSNs: 1751-8660

Abstract

Abstract : Dissolved gas analysis (DGA) has been widely used for the detection of incipient faults in oil‐filled transformers. This research presents a novel approach to DGA feature prioritisation and classification, which considers not only the relations between a fault type and specific gas ratios but also their statistical characteristics based on data derived from onsite inspections. Firstly, new gas features are acquired based on the analysis of current international gas interpretation standards. Combined with conventional gas ratios, all features are then prioritised by using the Kolmogorov–Smirnov test. The rankings are obtained by using their values of maximum statistic distance. The first three features in ranking are employed as input vectors to a multi‐layer support vector machine, whose tuning parameters are acquired by particle swarm optimisation. In the experiment, a bootstrap technique is implemented to approximately equalise sample numbers of different fault cases. A common 10‐fold cross‐validation technique is employed for performance assessment. Typical artificial intelligence classifiers with gas features extracted from genetic programming are evaluated for comparison purposes.