Performance of Edited Nearest Neighbour for Multi-Class Classification of Transformer Failure in Dissolved Gas Analysis

An imbalanced dataset happens when one class has significantly more samples than the others, which might bias classification algorithm results. Some transformer states are far more prevalent than others in dissolved gas analysis data, making it more difficult to categorise the less frequent states reliably. This research analyses the Edited Nearest Neighbours approach for identifying transformer faults based on dissolved gas measurement to overcome this issue. This study employed data from the UK's National Grid and DGALab, which included 790 transformer failure samples with six gas attributes and fault labelling. Then, several classifiers were evaluated, including XGBoost, Support Vector Machines, Convolutional Neural Networks, and Decision Trees. The Edited Nearest Neighbour and Decision Tree model outperforms other models in terms of accuracy (89.36%), recall (92.07%), and F1-score (89.25%). This technique performs well in managing the imbalance of multi-class data. Future research is to create adaptive undersampling approaches that modify based on the dataset's properties to improve classification performance. The study shows that the Edited Nearest Neighbors technique improves transformer fault classification accuracy on imbalanced datasets and highlighting its potential for robust fault detection.