Applying Unsupervised Machine Learning Method on FRA Data to Classify Winding Types

Over the years utilities have accumulated a large number of measured FRA data whilst the transformers' design information such as winding types may or may not necessarily be known. Different winding types own different equivalent electrical parameters, i.e. capacitance and inductance. For instance, the interleaved winding has higher series capacitance whilst the plain disc winding has lower series capacitance. As a result, unalike features are caused at specific frequency ranges of FRA. Consequently it is possible to correlate FRA characteristics with known design features. Hierarchical clustering is an unsupervised machine learning algorithm that groups similar objects together. In this paper, using the National Grid FRA database as an example, winding types are identified by Hierarchical Clustering method through grouping similar FRA data. In addition, a pre-processing technique called Dynamic Time Warping (DTW) is used to scale frequencies with the same FRA features before applying Hierarchical Clustering, and this has been proved to be the most suitable unsupervised machine learning methods to classify winding types. National Grid has been retiring transformers, and each transformer retired would go through forensic examination and knowledge acquired can then be used for asset management. Same faults may occur to same winding types and result in similar distortions of FRA features. With the technique employed in this paper, in combination with expertise knowledge and forensic information accumulated, the utility will be able to develop a strategy to manage similar type of transformers and achieve effective asset management.