Picture of DNA strand

Pioneering chemical biology & medicinal chemistry through Open Access research...

Strathprints makes available scholarly Open Access content by researchers in the Department of Pure & Applied Chemistry, based within the Faculty of Science.

Research here spans a wide range of topics from analytical chemistry to materials science, and from biological chemistry to theoretical chemistry. The specific work in chemical biology and medicinal chemistry, as an example, encompasses pioneering techniques in synthesis, bioinformatics, nucleic acid chemistry, amino acid chemistry, heterocyclic chemistry, biophysical chemistry and NMR spectroscopy.

Explore the Open Access research of the Department of Pure & Applied Chemistry. Or explore all of Strathclyde's Open Access research...

Selecting appropriate machine learning classifiers for DGA diagnosis

Aizpurua Unanue, Jose Ignacio and Catterson, Victoria and Stewart, Brian G and McArthur, Stephen and Lambert, Brandon and Ampofo, Bismark and Pereira, Gavin and Cross, James (2018) Selecting appropriate machine learning classifiers for DGA diagnosis. In: 2017 IEEE Conference on Electrical Insulation and Dielectric Phenomena. IEEE Dielectrics and and Electrical Insulation Society, Piscataway, N.J., pp. 153-156. ISBN 978-1-5386-1194-4

Text (Aizpurua-etal-IEEE-CEIDP-2017-Selecting-appropriate-machine-learning-clasifiers)
Accepted Author Manuscript

Download (508kB) | Preview


Dissolved gas analysis (DGA) is a common method of assessing transformer health. There are a number of machine learning classifiers reported to give a high accuracy on specific datasets, such as Artificial Neural Networks or Support Vector Machines. When these methods reach the same conclusion about the type of fault present, this can give an increased confidence in the veracity of the diagnosis. However, it is critical to analyze and quantify the strength of these classifiers in the presence of conflicting data to test their practicality for usage in the field. This paper investigates the adequacy of different machine learning based DGA diagnosis models in the presence of conflicting data. The proposed method will aid engineers with the selection of machine learning models so as to maximize the usability and accuracy in the presence of conflicting data.