An Efficient Method to Localize and Quantify Axial Displacement in Transformer Winding Using Support Vector Machines

Abstract

Power transformers are an inevitable and expensive equipment in an electrical power system. Condition monitoring uses predictive analysis to determine whether a problem is present or absent in order to prevent transformer failures and guarantee the transformer's safe operation. Among various condition monitoring techniques, Sweep Frequency Response Analysis (SFRA) is a powerful and reliable tool to detect winding deformations. However, the diagnosing potential of SFRA is still its infant state. Any mechanical damage in the transformer winding will change the equivalent circuit parameters and this change will be reflected in the FRA traces. By comparing the FRA traces of the testing transformer with normal winding the fault can be detected. To locate and quantify the axial displacement these FRA traces need to be acknowledged precisely. Support Vector Machine (SVM), a supervised machine learning technique helps to locate and quantify the axial displacement with the help of features extracted from the FRA traces of testing transformer and nominal winding. A series of axial displacements is simulated in FEMM Software and corresponding equivalent circuit parameters are used to generate FRA traces. Furthermore, features are extracted from these FRA traces to train the SVM model to enable it to predict the location and quantity of axial displacement accurately. Finally, the accuracy of this SVM model is tested through randomly created axial displacements data. The result indicates the ability of this technique to be used as an intelligent and accurate diagnostic tool.