Abstract
Power transformers are critical components in electrical power distribution systems, and their operational reliability is essential for maintaining a continuous power supply. Diagnosing the health of transformers presents significant challenges due to their complex physical and electrical structures. Frequency Response Analysis (FRA) has emerged as a widely adopted technique for assessing transformer conditions by analyzing impedance and transfer functions across a broad frequency range. However, applying FRA effectively remains challenging due to the lack of standardized methodologies for conducting tests and interpreting results. Variations in transformer design, fault conditions, and operating environments contribute to inconsistencies in diagnostic outcomes.
This study addresses the issue of non-standardized FRA testing and result interpretation, which limits effective classification of transformer health conditions. While FRA is extensively used, its reliance on subjective expert interpretation of graphs restricts its scalability in industrial applications. To overcome this limitation, the study develops a standardized FRA testing methodology that integrates machine learning and artificial intelligence algorithms for automated fault classification.
The primary objective of this research is to improve the reliability of FRA diagnostics by establishing a repeatable process applicable across various transformer types and fault scenarios. The study introduces intelligent classifiers based on machine learning algorithms, including Decision Trees, k-Nearest Neighbors, and Support Vector Machines, trained on FRA datasets to classify transformer conditions. Additionally, numerical indices such as the Correlation Coefficient, Cross-Correlation, and Cosine Distance are used to quantify changes in transformer characteristics, enabling objective health assessments.
The research employs a mixed analytical and experimental approach to develop and validate the proposed standardized FRA methodology. FRA datasets were collected from transformers under various fault conditions, including core and winding displacements, insulation deterioration, and bushing defects. The experimental setup
vi
involved FRA equipment to capture impedance and transfer function data across a wide frequency range, with statistical software and machine learning platforms, such as MATLAB, used to calculate numerical indices and implement classification algorithms.
Machine learning models, including Decision Trees, k-Nearest Neighbors, and Support Vector Machines, were applied to classify fault conditions using FRA data and numerical indices. These indices, such as the Correlation Coefficient and Cosine Distance, quantified deviations in resonance frequencies and impedance spectra, facilitating a consistent comparison of FRA results across different fault scenarios. This approach reduced subjectivity in diagnostics, enabling more consistent interpretation of FRA data.
The results demonstrated that machine learning models substantially enhanced FRA diagnostics. Specifically, Support Vector Machines and k-Nearest Neighbors classifiers achieved accuracy rates above 90% in identifying faults, such as core displacement, winding deformation, and insulation failures. Ensemble classifiers, such as Boosted and Bagged Trees, further improved classification accuracy to over 93%. These models effectively detected and classified transformer faults with minimal error, showing that the integration of numerical indices and machine learning enhances the reliability of FRA diagnostics.
A key finding was that certain indices, particularly the Correlation Coefficient and Cosine Distance, were highly sensitive to changes in transformer conditions, making them valuable for fault detection. These indices captured shifts in resonance frequencies and impedance spectra, offering clear indications of fault type and severity. Additionally, the use of machine learning reduced dependence on expert judgment, improving diagnostic accuracy and efficiency.
Compared to traditional diagnostic methods like Dissolved Gas Analysis and Partial Discharge testing, the machine learning-enhanced FRA methodology demonstrated superior accuracy and practicality. While these traditional methods are effective for detecting specific faults, they lack the comprehensive diagnostic capability of FRA,
vii
especially for mechanical and insulation issues. The machine learning-enhanced FRA approach provided faster results and required less manual interpretation, making it more efficient for large-scale applications.
This research shows that a standardized approach to FRA testing, combined with machine learning algorithms and numerical indices, significantly improves transformer fault detection and classification. In integrating Decision Trees, k-Nearest Neighbors, and Support Vector Machines with FRA data, this study offers more reliable, automated diagnostics applicable to various transformer types and fault conditions. These findings support more effective transformer maintenance strategies, reducing downtime and operational costs in power systems.