Combining and comparing various machine‐learning algorithms to improve dissolved gas analysis interpretation. Issue 15 (19th June 2018)

• Record Type: Journal Article
• Authors: Senoussaoui, Mohammed El Amine; Brahami, Mostefa; Fofana, Issouf
• Is Part Of: IET generation, transmission & distribution. Volume 12:Issue 15(2018)
• Page Start: 3673
• Page End: 3679
• Publication Date: 2018-06-19
• Subjects: transformer oil -- power engineering computing -- Bayes methods -- decision trees -- nearest neighbour methods -- pattern classification -- fault diagnosis -- multilayer perceptrons -- chemical engineering computing; DGA fault classification -- classification models -- bootstrap aggregation -- Adaboost algorithm -- J48 decision tree -- k‐nearest neighbour -- multilayer perceptron -- Bayes network -- artificial intelligence techniques -- Duval triangle -- key gases -- ratio method -- DGA interpretation -- dissolved gas analysis interpretation -- machine‐learning algorithms; Electric power production -- Periodicals; Electric power transmission -- Periodicals; Electric power distribution -- Periodicals; 621.3105
• DOI: 10.1049/iet-gtd.2018.0059
• Languages: English
• ISSNs: 1751-8687

Abstract

Abstract : Since the discovery of dissolved gas analysis (DGA), it is considered as a leading technique for the diagnosis of liquid insulated power equipment. However, accurate analysis results can only be achieved if the measured gases closely reflect the actual equipment condition to enable an appropriate interpretation of these gases. In general, conventional techniques such as the ratio method, key gases, and Duval triangle combined or not with artificial intelligence techniques such as machine‐learning algorithms are used for DGA interpretation. Here, four well‐known machine‐learning algorithms are compared in terms of DGA fault classification – Bayes network, multilayer perceptron, k ‐nearest neighbour, and J48 decision tree. Moreover, the effect of applying ensemble methods such as boosting through the Adaboost algorithm and bootstrap aggregation (bagging) is analysed, and the performances of these algorithms are evaluated. The data for developing classification models was transformed into three forms, other than the raw data. The obtained results clearly presented the efficiency and stability of some algorithms such as the J48 tree and Bayes networks for DGA fault classification, in particular, when the data is appropriately pre‐processed. Moreover, the performance of these algorithms was found to consistently improve by integrating the concepts of multiple models or ensemble methods.