Abstract
Power transformers are vital for maintaining the reliability and stability of electrical
systems. However, their vulnerability to faults, such as partial discharges and winding
deformation, poses significant operational risks. Advanced diagnostic techniques are
essential for timely fault detection and predictive maintenance. This study investigates
the application of machine learning (ML) techniques in transformer fault detection using
Frequency Response Analysis (FRA) data. The study aims to evaluate the effectiveness of
various ML models, the impact of frequency variations, and the contribution of numerical
indices to fault classification accuracy. FRA data, comprising 50 to 70 measurements per
transformer, were segmented into eight frequency bands (20 kHz to 12 MHz). A systematic
approach utilizing a confusion matrix was applied to classify faults such as partial
discharges and winding deformation. The performance of ML models, including Decision
Trees and Subspace KNN, was assessed in terms of classification accuracy. Machine
learning models achieved fault classification accuracies ranging from 80% to 100% across
eight frequency bands (20 kHz to 12 MHz). Decision Tree models excelled in detecting
insulation faults, achieving 100% accuracy for faults such as thermal aging (Class A), electrical
stress (Class B), and moisture ingress (Class C). Subspace KNN models demonstrated
strong performance for core-related faults, with classification accuracies of 100% for core
displacement (Class B) and core buckling (Class C), but they faced challenges with lamination
deformation, achieving 75% accuracy. Contamination-related faults exhibited a 100%
False Negative Rate (FNR), indicating a need for model refinement. Fault detection was
consistent across frequency bands, with key diagnostic markers at 7.6 MHz, 8.25 MHz, and
8.7 MHz providing high diagnostic value. Machine learning integration into FRA-based
diagnostics enhances the accuracy and reliability of transformer fault detection. While current
results are promising, future research should focus on deep learning approaches and
enhanced feature extraction to address challenges such as data scarcity and fault diversity