Comparative Study Of Machine Learning Techniques For Transfomer Fault Detection

Abstract

Many large structures, due to their large estimated load, are usually equipped with their own
transformers due to their higher voltage supplied from the utility. These transformers form the “heart” of the
entire structure as when the transformer is out of order, essentially the entire structure can no longer function.
Due to this, maintenance of transformer is given pre-eminence in any maintenance planning. However,
conventionally, these works are carried out based on a fixed time and mostly the detection system is done
manually. This is not an accurate measure of maintenance needs as these may vary with the usage of the
transformer and the environment it is operated in. This could lead to a lot of wastage of resources. To minimize
wastage, it would be ideal to carry out only specific servicing and only when it is required. In this project, the
possibility of transferring the task of deciding this ideal maintenance schedule to intelligent techniques is
studied. For this, three machine learning (ML) techniques are compared in detecting transformer faults which
could then be used for predictive maintenance scheduling. 340 DGA data samples are used for the training
and testing of the models and data pre-processing like normalization, logarithm and percentage methods are
performed on a raw data to improve performance. A software-based fault detection system is also developed
using three standard DGA fault interpretation methods like refined Roger’s ratio, IEC ratio and Doernenburg
ratio. The obtained result indicates that the accuracy and efficiency of the ML algorithms such as kNN, SVM
and ensemble improves with proper pre-processing techniques. The Quadratic SVM proved to be the best
performing algorithm under logarithm pre-processing technique with an accuracy of 85%.