Fault diagnosis of power transformer using feature selection techniques and KNN

Effective and correct fault diagnosis of power transformer is essential prerequisite for reliable power delivery. This paper presents a novel approach for power transformer incipient fault diagnosis based on feature selection techniques and K nearest neighbor algorithm. ReliefF, Laplacian Score and Fisher Score are employed to rank and select the most discriminative feature from 22 different features. 268 DGA samples are used to establish fault diagnosis model with the selected features. Fault diagnosis performance is evaluated with 5-fold cross validation methods. Fault diagnosis accuracy of the proposed methods achieves 72.5%. Results and comparisons against other conventional methods show relative superiority of feature selection based DGA interpretation approach in term of fault diagnosis accuracy.