Transformer Faults Detection using Inrush Transients based on Multi-class SVM

Inrush currents with low damping can arise during transformer energisation due to abrupt load changes, asymmetrical flux, and saturation. Moreover, the development of such imbalanced high electromagnetic stress has been found to have a detrimental impact on winding insulation. Therefore, prediction of the inrush current during energisation can provide crucial information regarding the condition of the winding and core. This article proposes an intelligent method for identifying and labelling transformer faults using inrush transient signature data. The multi-SVM (support vector machine) model has been used to sort estimates of transformer failure states that overlap in their kernel density. The SVM model is trained to diagnose the transformer's inrush faults, winding defects, and healthy conditions using statistical feature parameters acquired from the inrush transient data. Furthermore, the model's validity is determined by comparing its performance to that of logistic regression, naive Bayes, and random forest models. Findings suggest that the SVM classifier performs reasonably well in predicting inrush arising due to core fault for low and medium values of gamma.