Winding Fault Detection in Power Transformers Based on Support Vector Machine and Discrete Wavelet Transform Approach

Abstract

Transformer winding faults (TWFs) can lead to insulation breakdown, internal short circuits, and catastrophic transformer failure. Due to their low current magnitude—particularly at early stages such as inter-turn short circuits, axial or radial displacement, or winding looseness—TWFs often induce minimal impedance changes and generate fault currents that remain within normal operating thresholds. As a result, conventional protection schemes like overcurrent relays, which are tuned for high-magnitude faults, fail to detect such internal anomalies. Moreover, frequency response deviations caused by TWFs often resemble those introduced by routine phenomena such as tap changer operations, load variation, or core saturation, making accurate diagnosis difficult using traditional FRA interpretation techniques. This paper presents a novel diagnostic framework combining Discrete Wavelet Transform (DWT) and Support Vector Machine (SVM) classification to improve the detection of TWFs. The proposed system employs region-based statistical deviation labeling to enhance interpretability across five well-defined frequency bands. It is validated on five real FRA datasets obtained from operating transformers in Gauteng Province, South Africa, covering a range of MVA ratings and configurations, thereby confirming model transferability. The system supports post-processing but is lightweight enough for near real-time diagnostic use, with average execution time under 12 s per case on standard hardware. A custom graphical user interface (GUI), developed in MATLAB R2022a, automates the diagnostic workflow—including region identification, wavelet-based decomposition visualization, and PDF report generation. The complete framework is released as an open-access toolbox for transformer condition monitoring and predictive maintenance.

1. Introduction

Transformer winding fault (TWF) emerges when the insulating material degrades, mechanical shock takes place, or thermal stress accumulates, generating nonlinear current distortion and impedance imbalance. These faults generally come with partial discharges, localized overheating, and mechanical vibrations in the transformer. At the site of a TWF, the current becomes highly non-uniform often showing abrupt spikes or distortions when disrupts the voltage distribution across the winding, leading to insulation stress and increased risk of dielectric breakdown. Therefore, conventional FRA interpretational methods commonly fail to detect these faults especially when deviations are subtle or masked by operational variations [1,2]. Identifying TWFs is challenging due to the subtle deviations in frequency response and their overlap with normal operational variations, which often go undetected by conventional FRA interpretation techniques [3,4].

According to diagnostic guidelines IEC 60076-18 [5], traditional FRA interpretation sensitivity significantly declines for impedance variations below 2–5%, making early-stage TWF detection unreliable without advanced signal processing.

Furthermore, TWFs exhibit nonlinear frequency response characteristics that resemble other phenomena such as tap changer operations, load tap variations, and natural aging effects—making them difficult to distinguish using conventional FRA methods [5,6,7]. Therefore, identifying and preventing TWFs through accurate diagnostic techniques is essential. Misdiagnosis or undetected faults can lead to insulation breakdown, equipment damage, costly outages, and even catastrophic transformer failures [8]. Therefore, it is crucial to detect TWFs quickly and accurately and to distinguish them from other non-critical variations in frequency response. Typical winding configurations and fault types are illustrated in Figure 1 [8].

In this work, the FRA spectrum is divided into five diagnostic frequency regions based on IEC 60076-18 and CIGRÉ WG A2.26: (i) Core Issues (10–1000 Hz), (ii) Bulk Winding Displacement (1–10 kHz), (iii) Winding Deformation (10–100 kHz), (iv) Inter-Turn Shorts & Tap Changer Influence (100 kHz–1 MHz), and (v) Partial Discharge/Floating Metal Effects (1–2 MHz).

So far, various methods have been proposed to detect transformer winding faults. Most recent approaches focus on machine learning techniques applied to frequency response data [9].

Wang et al. [10] proposed a novel classification method for radial transformer winding faults using Frequency Response Analysis (FRA) combined with a Particle Swarm Optimized Relevance Vector Machine (PSO-RVM). Fault types such as free buckling (FB) and forced buckling (FDB) were simulated and analyzed. Numerical indices were extracted from FRA signatures to train and validate the PSO-RVM model. Results showed that PSO-RVM achieved high classification accuracy while using fewer vectors than PSO-SVM, improving model sparsity and efficiency. Zhao et al. [11] introduced an enhanced method to classify transformer winding deformation faults using Frequency Response Analysis (FRA) polar plots and multiple SVM classifiers. Polar plots were constructed from FRA magnitude and phase data, and image texture features were extracted using digital image processing. Three independent SVMs were trained on the extracted features and combined into a robust ensemble classifier. Experimental validation showed superior accuracy over traditional FRA and single-SVM methods. Ezziane et al. [12] proposed a method using support vector machines (SVM) to diagnose transformer winding faults, including fault type, location, and severity. FRA signals were analyzed using statistical indicators, with datasets collected from both healthy and faulted winding conditions. The SVM was trained on this data to classify faults with high accuracy. Results confirmed the method’s potential for precise and intelligent transformer diagnostics. Wang et al. [13] proposed a hybrid diagnostic method combining Frequency Response Analysis (FRA) with a kernel-based extreme learning machine (KELM) optimized by a seagull optimization algorithm (SOA). FRA tests were conducted on lab-scale winding models under various fault conditions. Numerical features from the FRA traces were used to train the SOA-KELM for fault type and severity classification. Comparative results showed that SOA-KELM outperformed other classifiers in both accuracy and computational efficiency.

Bigdeli et al. [14] investigated the use of intelligent classifiers—PNN, decision tree, SVM, and k-NN—for identifying various transformer faults using Frequency Response Analysis (FRA). Faults such as axial and radial deformation, disc spacing variation, and core damage were tested on model transformers. Numerical and statistical features were extracted from segmented frequency ranges of the FRA data. The trained classifiers were then validated using real transformer measurements, with performance comparisons highlighting the diagnostic strengths of each model. Behkam et al. [15] applied advanced machine learning methods to diagnose disk space variation (DSV) faults in transformer windings using FRA data. Tests were performed on a 1.6 MVA transformer with varying DSV severities and positions, capturing all FRA components—magnitude, phase, real, and imaginary. Classifiers including SVM, Decision Tree, Random Forest, and XGBoost were trained to identify fault location and severity. The models achieved up to 100% accuracy in test datasets, confirming the potential of ML-driven FRA interpretation for precise DSV diagnosis. Saji et al. [16] proposed an SVM-based method for localizing and quantifying axial displacement in transformer windings using Sweep Frequency Response Analysis (SFRA). FEMM simulations were used to model varying axial displacement levels, from which FRA traces were generated. Key features were extracted and used to train an SVM model for accurate fault localization and severity estimation. The approach demonstrated high diagnostic precision, confirming SVM’s effectiveness for intelligent transformer condition monitoring. Muhammed et al. [17] developed an SVM-based diagnostic method to locate and quantify axial displacement in transformer windings using features derived from Frequency Response Analysis (FRA). Axial faults were modeled via FEM simulations, and equivalent circuit parameters were used to generate FRA traces. These were processed to extract statistical features for SVM training. Validation on a 3.3 MVA transformer confirmed the model’s ability to precisely detect and localize displacement faults.

Kang et al. [18] proposed a transformer winding fault classification method by integrating Frequency Response Analysis (FRA) with Support Vector Machines (SVM) optimized using the Bald Eagle Search (BES) algorithm. The BES was used to tune the kernel parameter and penalty factor of SVM for improved convergence and accuracy. Comparative results demonstrated that BES-SVM outperformed both traditional SVM and PSO-SVM models. The approach achieved higher diagnostic precision for identifying winding faults. Zhou et al. [19] introduced a novel method for diagnosing autotransformer (AT) winding faults using image-based features extracted from Frequency Response Analysis (FRA) polar plots. The approach applies gray-level gradient and difference statistics as inputs to a binary tree Support Vector Machine (SVM) classifier. Simulated faults, including short circuits, axial deformation, and capacitance variation, were accurately identified. The method demonstrated high accuracy in fault type and faulty winding localization for ATs. Wang et al. [20] proposed a diagnostic method for autotransformer (AT) winding defects based on transfer function pole distribution using Frequency Response Analysis (FRA). The method employs gray correlation analysis, Tamura texture features, and GGCM to extract diagnostic parameters. These features are fed into a Cuckoo Search-optimized SVM (CS-SVM) to classify defect types and affected windings. Results confirm the method’s high accuracy in identifying both single and mixed winding faults in AT systems.

Although several techniques have been proposed for detecting TWFs, most lack sufficient accuracy and rigorous validation for reliable fault identification. Using conventional mathematical approaches for TWF detection introduces significant complexity, and most existing methods have not been applicable to real-world substations. The proposed DWT-SVM classifier achieved a cross-validated accuracy ranging from 40% to 80% across five real transformer case studies, depending on the fault type and dataset complexity. Given these limitations, there is a need to develop an accurate, industry-aligned method with strong evaluation metrics, implemented using widely adopted engineering software platforms.

This paper proposes a highly accurate fault detection framework combining DWT and SVM classification to address the challenges of detecting transformer winding faults. While previous studies used DWT for feature extraction with varying mother wavelet preferences, this work increases the decomposition level to enhance classification accuracy and achieve more distinctive feature representation compared to earlier approaches. Considering the limitations observed in prior studies, this work employs a five-level DWT using the Daubechies 4 (db4) mother wavelet in combination with a SVM classifier for accurate detection of TWFs. The proposed approach is implemented entirely in MATLAB R2022a with an integrated graphical user interface (GUI), allowing users to load real FRA datasets, perform multi-level wavelet decomposition, visualize the diagnostic plots, and automatically generate structured PDF reports. Unlike earlier works that relied on synthetic or simulated networks, this study uses actual FRA data collected from five operating power transformers rated between 5 MVA and 40 MVA across substations. The deviation between baseline and current FRA signatures is processed region-wise, and DWT features are extracted from each region. Each region is statistically evaluated using indicators such as coefficient of determination(R²), MAPE (Mean Absolute Percentage Error), MAE (Mean Absolute Error), RMSE (Root Mean Squared Error), SD (Standard Deviation), and CC (Correlation Coefficient) to determine the degree of deviation. These features are then statistically labeled based on deviation severity and used to train the SVM classifier. Each region’s feature vector is passed as input to the SVM model, which classifies the data into normal or faulted classes (binary: 0 or 1). The practical and engineering focus of the study—supported by a GUI-driven implementation and real diagnostic datasets—demonstrates the framework’s industrial applicability and potential for standardized transformer health assessment.

The remainder of the study is organized as follows: Section 2 introduces the materials and methods of the proposed WTF detection of transformers based on SVM, and Wavelet transform approach. Section 3 introduces the results of studied cases studies. Section 4 introduces a discussion of the results and finally Section 5 presents the conclusion.

2. Materials and Methods

Figure 2 demonstrates the two basic stages of the proposed procedure. The first part engages with data processing and the second part involves output generation. In the first stage the procedure begins with FRA data acquisition where the user loads transformer FRA data. This is followed by Preprocessing where the system parses the raw data, computes the deviation between baseline and current signals, and organizes it by diagnostic frequency regions. Wavelet-Based Feature Extraction is then applied using multi-level discrete wavelet transform (DWT) with wavelets such as Daubechies 4 (db4), Symlet 5 (sym5), and Coiflet 5 (coif5) to extract statistical features like mean, standard deviation, and entropy. The extracted features undergo Statistical Deviation Labeling (SDL), where regions are flagged as “faulted” or “normal” based on metrics such as coefficient of determination (R²) and the mean absolute percentage error (MAPE).

Finally, in this first stage, SVM Classification is performed using these statistically labeled features to build region-aware predictive models. The second stage engages with user interaction and reporting. Initially, graphical user interface (GUI) Visualization presents the FRA fingerprint and D5 decomposition plots, along with automated fault summaries. Consequently, the PDF Report Generation module creates a detailed diagnostic report containing all visualizations, result tables, and maintenance recommendations. Together, these two stages form a seamless FRA diagnostic pipeline from data input to actionable output.

Figure 3 also illustrates a schematic of the individual elements of the proposed procedure. The FRA signal data is first loaded from operational transformers ranging from 5 MVA to 40 MVA. The signal is preprocessed to compute the deviation between baseline and current measurements across predefined frequency regions. Moreover, DWT is applied using the db4 at level five to extract diagnostic features from each region. Subsequently, statistical thresholds based on the R² and MAPE are used to label each region as either faulted or normal. Based on these statistical labels and extracted features, classification and decision-making are performed using a SVM algorithm. Finally, the results are visualized within the GUI and exported in a diagnostic PDF report. Since FRA signatures can exhibit subtle variations due to loading, tap positions, or environmental conditions, distinguishing between normal operational variation and true fault-related deviations remains a key challenge. For this reason, the system integrates both signal-domain and statistical-domain intelligence to improve fault detection reliability across different transformer units and operating contexts.

2.1. Data Collection and Preparation

In this study, FRA magnitude traces are used for the TWF detection. The proposed diagnostic architecture is coupled with the five real-world FRA datasets collected from operating substations located in Johannesburg Gauteng Province with transformer ratings ranging from 5 MVA to 40 MVA. Then, the TWF response signatures will be collected as data for feature extraction in the following stage. To improve robustness under practical field conditions, the system applies preprocessing measures such as linear interpolation to handle missing data points, Hampel filtering for outlier removal, and low-pass smoothing to reduce high-frequency noise artifacts before conducting DWT-based feature extraction. These steps ensure diagnostic reliability even with incomplete or corrupted FRA datasets.

2.2. Wavelet Transform

In this study, discrete wavelet transform (DWT) is employed to decompose the transformer FRA deviation signal into low- and high-frequency components using the Daubechies 4 (db4) wavelet. This decomposition results in two primary outputs: approximation coefficients (A5), representing the low-frequency, high-scale signal content, and detail coefficients (D5), which capture the high-frequency, low-scale information. As shown in Figure 4, the FRA signal is first passed through both a low-pass and high-pass db4 filter. The low-pass filter output (A5) is used for extracting statistical features for SVM classification, while the high-pass filter output (D5) is used to generate visual decomposition plots across diagnostic frequency bands [21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39].

Although the filtering process theoretically doubles the number of output samples—e.g., from an original signal of 1000 points to 2000 (1000 from each filter)—only one set is retained through downsampling to preserve the core signal characteristics. This enables efficient feature extraction without redundancy. The separation and labeling of A5 and D5 coefficients form the foundation for the multi-region analysis and classifier training conducted in this research.

To minimize redundancy and preserve the most relevant diagnostic information, downsampling is applied after the initial wavelet decomposition stage. This process retains only the essential approximation (cA) and detail (cD) coefficient sequences derived from the low-pass and high-pass outputs, respectively. As illustrated in Figure 5, the full-length filtering path retains all coefficients, while the downsampled wavelet transform (WT) path reduces the data volume by half—yielding compressed yet information-rich features. In practical implementation, this approach forms the basis of the discrete wavelet transform (DWT), effectively isolating the critical frequency-domain patterns used for fault classification in transformer winding diagnostics.

To progressively isolate fault-relevant frequency bands, the approximation component from each decomposition level undergoes further wavelet filtering. This recursive process forms a hierarchical structure known as the wavelet decomposition tree. Although theoretically infinite, the decomposition is typically limited to a fixed number of levels where signal distinctions become diagnostically meaningful. As illustrated in Figure 6, the FRA deviation signal is decomposed into five levels using the Daubechies (db4) wavelet, with D5 coefficients—spanning the 1000–10,000 Hz range—being used for fault classification due to their sensitivity to winding deformations.

In light of the above decomposition, feature extraction plays a critical role in TWF detection. Subtle winding defects often induce only slight deviations in FRA responses, making their patterns difficult to distinguish from normal variations caused by load changes, tap operations, or core saturation. To uncover these hidden patterns, this study applies the DWT using the Daubechies (db4) mother wavelet across five decomposition levels. The FRA signal deviations are used as input, and then systematically decomposed to isolate frequency bands relevant to physical fault phenomena. From the resulting wavelet coefficients—particularly the detail coefficients (D5)—statistical and pattern-based features are extracted for machine learning classification, marking the conclusion of the wavelet-based feature extraction process.

To ensure robustness against minor operational variations such as tap changer movements or load fluctuations, region-based statistical labeling applies strict quantitative thresholds (e.g., MAPE > 15%, R² < 0.80), filtering out insignificant deviations and focusing the classifier only on statistically meaningful fault patterns.

2.3. Support Vector Machine Algorithm

The theoretical robustness and simplicity of implementation have made SVM a widely adopted method for fault classification in transformer diagnostics. Compared to artificial neural networks (ANNs), SVMs tend to generalize better with lower risk of overfitting, producing more consistent classification boundaries—particularly useful in detecting subtle transformer winding faults from FRA data. SVM is also input-space agnostic, making it highly suitable for high-dimensional signal features.

The SVM classifier operates by training on labeled samples, where each data point comprises extracted features and a class label indicating either a fault or normal condition. During testing, the trained SVM model assigns new data samples to one of two categories based on learned decision boundaries. In this study, a binary classification strategy is adopted, with SVM constructing an optimal hyperplane that separates fault and normal FRA regions using support vectors. Figure 7 illustrates this process, showing how SVM distinguishes between healthy and faulty winding signatures.

In the proposed transformer fault detection framework, the SVM identifies a hyperplane that optimally separates the extracted FRA-based features of faulty and healthy winding conditions. The support vectors are the boundary points that lie closest to the decision surface. The optimal separating margin, defined as ${\displaystyle \frac{2}{\parallel W\parallel }}$, ensures maximum class separability, where $W$ is the normal vector to the hyperplane, and $b$ is a bias term. The separating hyperplane is mathematically expressed as follows in Equation (1).

$$W^T{x}_i+b=0$$

(1)

To ensure optimal classification of transformer winding health states, the support vector machine maximizes the geometric margin between the fault and normal classes. The margin $m$—the distance between the closest support vectors from each class—is inversely proportional to the norm of the weight vector $W$: as follows in Equation (2).

$$m={\displaystyle \frac{2}{\parallel W\parallel }}$$

(2)

By minimizing $\parallel W\parallel$, the classifier increases the margin, reducing the risk of overfitting and improving generalization. In this study, maximizing mmm ensures that the boundary formed between healthy and faulty FRA signal patterns is as robust and separable as possible, enhancing the reliability of transformer condition monitoring.

To achieve optimal separation, the objective is to minimize $\parallel W\parallel$, transforming the classification task into a convex quadratic optimization problem as follows in Equation (3).

$$\min {\displaystyle \frac{1}{2}}{\parallel W\parallel }^2$$

(3)

Subject to the constraint as follows in Equation (4)

$$y_i\left(W^T{x}_i+b\right)\ge 1$$

(4)

Lagrange multipliers $a_i$ are then introduced to solve the dual form of the problem, as follows in Equation (5).

$$\max L\left(a\right)={\displaystyle \sum }_{i=1}^N{a}_i+{\displaystyle \frac{1}{2}}{\displaystyle \sum }_{i=1}^N{\displaystyle \sum }_{j=1}^N{a}_i{a}_j{y}_i{i}_j\left(x_{i,}{x}_j\right)$$

(5)

Subject to the constraint expressed in Equation (6).

$${\displaystyle \sum }_{i=1}^N{a}_i{a}_j=0$$

(6)

where,

$a_i$—are the Lagrange multipliers

$y_i\in \left\{-1,1\right\}$—are the class labels (e.g., fault vs. normal)

$x_i$ and $x_j$ represent FRA feature vectors.

Solving this quadratic programming problem identifies the optimal support vectors and maximizes the separation margin, improving classification performance for incipient transformer faults. The solution yields the weight vector as follows in Equation (7).

$$W^{\ast }={\displaystyle \sum }_{i=1}^N{x^{\ast }}_i{x}_i{y}_j$$

(7)

Bias term is expressed as follows in Equation (8).

$$b^{\ast }=y_{sv}+{\displaystyle \sum }_{i=1}^N{a^{\ast }}_i{y}_j\left(x_{i,}{x}_{sv}\right)$$

(8)

The final decision function for classifying FRA data points is expressed as follows in Equation (9).

$$f\left(x\right)=sign\left({\displaystyle \sum }_{i\in v}^N{x^{\ast }}_i{y}_j\left(x_{i,}{x}_{sv}\right)+b^{\ast }\right)$$

(9)

Here, $K(x_i{y}_j)$ is the kernel function, and support vectors are those with ${a^{\ast }}_i>0$. By maximizing the margin and accurately classifying both training and unseen FRA samples, this model ensures robust winding fault detection and interpretation as shown in Figure 8.

The final stage in the proposed framework involves classification and decision-making to identify TWFs based on frequency response deviation signals. In this study, a binary classification system is implemented using a SVM trained on features extracted from the D5 wavelet decomposition of FRA signatures. Each sample is labeled as either fault (1) or normal (0). The classification workflow is structured into: (1) feature extraction, where the DWT-based statistical characteristics are computed from the FRA deviation signal; (2) model training, using real transformer FRA measurements collected from Gauteng Province; and (3) diagnosis and evaluation, where a trained SVM model classifies new input data using these features. To evaluate classification performance, a confusion matrix is generated within the custom MATLAB GUI, which automates the full pipeline from signal input to PDF report export.

2.4. Validation Method

To assess classifier performance and prevent overfitting, a 5-fold cross-validation procedure was applied in each case study. In this approach, the dataset was randomly partitioned into five subsets, where four subsets were used for training and the remaining subset for validation. The process was repeated five times, and the final classification accuracy was reported as the average across all folds. The results demonstrated variable classification accuracies across the five transformers—typically around 80% in most cases and lower (around 40%) in more challenging scenarios—highlighting both the strengths and practical limitations of the proposed framework when deployed on real-world FRA data. Furthermore, to enhance robustness against operational variations such as tap changer movements and load fluctuations, the framework employs statistical deviation thresholds (e.g., MAPE > 15%, R² < 0.8) within predefined diagnostic regions. Only deviations exceeding these thresholds are classified as faults, ensuring that normal operational shifts do not trigger false alarms.

2.5. GUI Practical Application and Efficiency Evaluation

To validate its usability, the developed MATLAB-based GUI was employed across all five case studies. On a standard workstation (Intel i7 CPU, 16 GB RAM), each full diagnostic process—comprising data loading, DWT feature extraction, SVM classification, and PDF report generation—was completed within approximately 8–12 s per transformer. The GUI required minimal user intervention beyond file selection, and no computational bottlenecks were encountered during execution. This confirms the tool’s efficiency for real-world transformer condition monitoring and supports its potential for deployment in industrial maintenance environments.

3. Results

This section presents the diagnostic results of FRA applied to five real-world power transformers located at substations in Gauteng Province, South Africa. Each transformer underwent testing using the Megger FRAX 101 2019 analyzer, with data processed via MATLAB R2022a for wavelet decomposition and SVM classification. A GUI was developed to visualize deviation signatures and generate automated diagnostic reports.

3.1. Transformer and Data Collection

All transformers analyzed in this study were subjected to on-site FRA testing under oil-filled conditions, using standard sweep frequency ranging from 20 Hz to 2 MHz. The test configuration followed a typical impedance-matched setup with 50 Ω termination, as illustrated in Figure 9, ensuring accurate signal integrity across the test bandwidth.

The Megger FRAX 101 setup, shown in Figure 10, includes a signal generator, source and load matching resistors (50 Ω), and dual voltage measurement ports for computing the frequency-dependent transfer function. The system measures the voltage across the transformer under test (V₂) and compares it with the input voltage (V₁) to generate a magnitude trace in decibels (dB), revealing resonance shifts and impedance anomalies caused by winding faults.

For region-based fault localization, the full frequency scan was segmented into five diagnostic bands, aligned with IEC 60076-18 [40] and CIGRÉ WG A2.26 [41] standards.

Table 1. Parameters used in the FRA-based TWF detection model.

FRA Diagnostic Parameter	Typical Range/Threshold
Frequency Range (Full Scan)	10 Hz–2 MHz
Core Issue Detection Band	10–1000 Hz
Bulk Winding Displacement Band	1000–10,000 Hz
Winding Deformation Band	10,000–100,000 Hz
Inter-Turn Short/Tap Changer Band	100,000–1,000,000 Hz
Partial Discharge/Floating Metal	1,000,000–2,000,000 Hz
DWT Decomposition Levels	5 levels (db4, coif5, sym5)
MAPE Fault Threshold	>15%
R2 Fault Threshold	<0.80
SVM Output Classes	0 (Normal), 1 (Fault)

defines these regions and lists diagnostic thresholds used during statistical deviation labeling.

3.2. Case Study 1

This case study focuses on a 6 MVA, three-phase power transformer equipped with an offload tap changer, tested at tap position 1 under oil-filled conditions. The objective was to evaluate the transformer’s winding and core integrity using FRA in conjunction with advanced signal processing and machine learning techniques. The FRA test was conducted across a wide frequency range (20 Hz to 1.99 MHz) using the Megger FRAX 101 system, and key transformer parameters are summarized in

Table 2. Case Study 1—Summary of Key Transformer and Test Parameters.

Attribute	Value
Rated Power (MVA)	6
HV Rating (kV)	88
LV1/LV2 Rating (kV)	1.22/1.22
Tertiary Rating (kV)	2.38
Impedance HV–LV1 (%)	6.07
Impedance HV–LV2 (%)	6.14
Impedance HV–Tertiary (%)	10.07
Tap Changer Type	Offload Tap Changer (DETC)
Tap Position During Test	1
Sweep Frequency Range (Hz)	20–1,988,018
Input Resistance (Ω)	50

. The data obtained serves as the basis for fault region decomposition, wavelet-based feature extraction, and classification using support vector machines (SVMs).

Figure 11 presents the baseline and fault-mode FRA signatures acquired from Case Study 1 using the Megger FRAX 101 analyzer, illustrating the frequency-dependent deviations observed across the measurement range.

Figure 12 illustrates the level-5 detail (D5) decomposition of the FRA deviation signal using the Daubechies 4 (db4) wavelet, segmented across five diagnostic frequency regions to support TWF localization in Case Study 1. This data has been attained using the graphical user interface (GUI) presented in Appendix A.

Table 3. Case Study 1—SVM Trained with coif5 Wavelet.

Region	Tested Patterns	Correct	Incorrect
Core Issues	354	283	71
Winding Bulk Displacement	208	166	42
Winding Deformation	208	166	42
Inter-Turn Shorts & Tap Changer	208	166	42
Partial Discharge & Floating Metal	63	50	13

presents the classification results of the SVM trained using features extracted via the coif5 mother wavelet. After applying 5-fold cross-validation, all five FRA fault regions were classified with an approximately 80% correct prediction rate, confirming strong generalization ability across the dataset. A total of 1041 patterns were tested across the defined regions.

The performance of the sym5-based SVM classifier is shown in

Table 4. Case Study 1—SVM Trained with sym5 Wavelet (Cross-Validated).

Region	Tested Patterns	Correct	Incorrect
Core Issues	354	283	71
Winding Bulk Displacement	208	166	42
Winding Deformation	208	166	42
Inter-Turn Shorts & Tap Changer	208	166	42
Partial Discharge & Floating Metal	63	50	13

, indicating comparable classification results across all regions with no major misclassification deviation from coif5-based performance. This further supports the reliability and consistency of wavelet-based SVM schemes for transformer winding fault discrimination, even under cross-validation constraints.

Results for the db4-based classifier are shown in

Table 5. Case Study 1—SVM Trained with db4 Wavelet (Cross-Validated).

Region	Tested Patterns	Correct	Incorrect
Core Issues	354	283	71
Winding Bulk Displacement	208	166	42
Winding Deformation	208	166	42
Inter-Turn Shorts & Tap Changer	208	166	42
Partial Discharge & Floating Metal	63	50	13

, confirming its robust diagnostic ability across all five frequency regions. While minor errors were observed after cross-validation, the db4-based model maintained consistent performance comparable to the coif5 and sym5 wavelet models. Based on its favorable compactness, interpretability, and empirical stability, db4 was ultimately selected for the final diagnostic implementation.

To benchmark different diagnostic approaches, four fault detection schemes were compared based on their achieved classification accuracies. As shown in

Table 6. Case Study 1—Scheme Comparison Accuracy (Cross-Validated).

Method	Accuracy (%)
DWT-SVM	80
WT-SVM	50
DWT-FFA	16.67
WPT-SVM	80

, the proposed DWT-SVM and WPT-SVM models each achieved 80% accuracy following 5-fold cross-validation, significantly outperforming conventional WT-SVM (50%) and DWT-FFA (16.67%) techniques. These results demonstrate the superiority of wavelet-transform-based feature extraction combined with statistical deviation labeling for FRA signature interpretation.

Table 7. Case Study 1—Statistical Summary by FRA Region.

Region	CC	SD	MSE	RMSE	MAE	MAPE	R2	Label
Core Issues	1.0000	0.0000	100.00	10.000	10.000	44.558	−0.1064	Fault
Winding Bulk Displacement	0.9859	1.3494	12.500	3.5355	3.2692	6.539	0.7730	Fault
Winding Deformation	1.0000	0.0000	9.000	3.0000	3.0000	5.281	0.7790	Fault
Inter-Turn Shorts & Tap Changer	1.0000	0.0000	9.000	3.0000	3.0000	6.806	0.8114	Normal
Partial Discharge & Floating Metal	1.0000	0.0000	9.000	3.0000	3.0000	10.640	0.6791	Fault

summarizes the statistical error metrics extracted across the defined diagnostic frequency regions. Metrics such as the correlation coefficient (CC), standard deviation (SD), mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and the coefficient of determination (R²) were used both for region labeling and to support the engineering interpretation of winding fault presence.

3.2.1. Engineering Interpretation & Recommendation

The transformer tested in Case Study 1 was analyzed using discrete wavelet transform (DWT) features extracted from five well-defined FRA diagnostic frequency regions. All three trained classifiers—using coif5, sym5, and db4 mother wavelets—achieved approximately 80% classification accuracy following cross-validation, correctly identifying all statistically fault-labeled regions with minimal generalization error. Among the tested methods, DWT-SVM using the db4 wavelet demonstrated the most consistent performance and superior interpretability, supporting its selection as the preferred diagnostic model.

The statistical summary highlighted significant deviations in the Core Issues, Winding Bulk Displacement, Winding Deformation, and Partial Discharge & Floating Metal frequency bands. These deviations were evidenced by elevated MAPE values (e.g., 44.56% in the core region), negative or low R² coefficients, and distinctly higher error statistics, indicating early mechanical and dielectric anomalies. Only the Inter-Turn Shorts & Tap Changer region remained within statistically acceptable thresholds, suggesting stable insulation and turn-to-turn integrity.

Based on the diagnostic findings, offline maintenance actions are strongly recommended for the affected regions. Recommended tests include winding resistance measurement, insulation resistance evaluation, dissolved gas analysis (DGA), and internal visual inspection if possible. The transformer appears to exhibit early-stage core shifting and localized winding deformation, which require proactive investigation to prevent further deterioration. It is recommended to standardize DWT-SVM (db4) as the primary diagnostic approach for this transformer configuration to ensure consistent and reliable health assessments during future condition monitoring.

3.2.2. Root Cause Analysis Based on FRA Diagnostic Standards

Based on the observed deviations in the Frequency Response Analysis (FRA) results for Case Study 1—specifically the exceptionally high MAPE value (44.56%), the negative R² (−0.1064), and the elevated MAE (10) in the low-frequency range (10 Hz to 1000 Hz)—the most probable root cause of the anomaly is core clamping looseness or mechanical movement of the magnetic core.

According to international standards such as IEC 60076-18 and technical guidelines from CIGRÉ WG A2.26, deviations in this frequency band are typically associated with disturbances in the magnetic circuit, including core shifting, relaxation of yoke clamping pressure, or mechanical displacement caused by thermal stress, transport vibration, or operational shocks.

The absence of abnormalities in the mid- and high-frequency bands supports the conclusion that the winding structures remain mechanically intact, thereby ruling out inter-turn short circuits or major insulation failures. Visual inspection following shutdown is expected to reveal symptoms such as loosened yoke bolts, displaced core laminations, or compression marks on the clamping structures.

Therefore, the root cause of the detected anomalies is most plausibly mechanical in nature, specifically related to compromised core clamping integrity. Mechanical re-tensioning of the core bolts, followed by re-testing via FRA after maintenance, is recommended to restore the transformer’s structural stability and to confirm the corrective actions were effective.

3.3. Case Study 2

This case study investigates a 6 MVA, three-phase transformer equipped with an offload tap changer, tested at tap position 1 under oil-filled conditions. The purpose of the test was to assess the mechanical and dielectric health of the windings and core using FRA. The FRA scan was conducted using the Megger FRAX 101 system, spanning 20 Hz to 1.99 MHz.

Table 8. Case Study 2—Summary of Key Transformer and Test Parameters.

Attribute	Value
Rated Power (MVA)	6
HV Rating (kV)	92.5
LV1/LV2 Rating (kV)	1.22/1.22
Tertiary Rating (kV)	2.36
Impedance HV–LV1 (%)	6.12
Impedance HV–LV2 (%)	6.18
Impedance HV–Tertiary (%)	10.14
Tap Changer Type	Offload Tap Changer
Tap Position During Test	1
Sweep Frequency Range (Hz)	20–1,988,018
Input Resistance (Ω)	50

summarizes the transformer’s electrical and test configuration details. The results from this case serve as the second benchmark for validating the wavelet-SVM classification framework.

Figure 13 illustrates the measured FRA signatures for Case Study 2. A visible deviation is observed when comparing the baseline trace with the case trace, particularly in the mid-to-high frequency regions, indicating the presence of mechanical anomalies.

Figure 14 presents the level-5 detail (D5) decomposition of the deviation signal using the db4 wavelet. Faulted behavior is evident in several regions, such as winding displacement and deformation zones.

Table 9. Case Study 2—SVM Trained with coif5 Wavelet (Cross-Validated).

Region	Tested Patterns	Correct	Incorrect
Core Issues	354	283	71
Winding Bulk Displacement	208	166	42
Winding Deformation	208	166	42
Inter-Turn Shorts & Tap Changer	208	166	42
Partial Discharge & Floating Metal	63	50	13

presents the classification performance of the SVM model trained using features extracted from the coif5 wavelet. After applying 5-fold cross-validation, the classifier achieved approximately 80% correct classification across all five FRA fault regions, successfully identifying 831 out of 1041 patterns, thereby demonstrating the strong discriminatory capability of the coif5-derived features.

Table 10. Case Study 2—SVM Trained with sym5 Wavelet (Cross-Validated).

Region	Tested Patterns	Correct	Incorrect
Core Issues	354	283	71
Winding Bulk Displacement	208	166	42
Winding Deformation	208	166	42
Inter-Turn Shorts & Tap Changer	208	166	42
Partial Discharge & Floating Metal	63	50	13

summarizes the SVM classification output based on sym5-derived features. The classifier yielded similar cross-validated performance across all fault regions, confirming the robustness and generalization capability of sym5 wavelet features in representing winding anomalies.

Table 11. Case Study 2—SVM Trained with db4 Wavelet (Cross-Validated).

Region	Tested Patterns	Correct	Incorrect
Core Issues	354	283	71
Winding Bulk Displacement	208	166	42
Winding Deformation	208	166	42
Inter-Turn Shorts & Tap Changer	208	166	42
Partial Discharge & Floating Metal	63	50	13

reflects the classification results for the db4-based model. As with the coif5 and sym5-based classifiers, all fault zones were correctly identified with approximately 80% cross-validation accuracy. db4 again emerged as the preferred wavelet for real-world transformer diagnostics due to its compact support, excellent localization properties, and superior interpretability.

Table 12. Case Study 2—Scheme Comparison Accuracy (Cross-Validated).

Method	Accuracy (%)
DWT-SVM	80
WT-SVM	50
DWT-FFA	16.67
WPT-SVM	80

benchmarks the accuracy of four diagnostic schemes under cross-validation. The DWT-SVM and WPT-SVM models each achieved 80% accuracy, validating their effectiveness for FRA-based winding fault classification. Conventional approaches such as WT-SVM and DWT-FFA significantly underperformed, highlighting the importance of both feature extraction and statistical deviation labeling in intelligent diagnostics.

Table 13. Case Study 2—Statistical Summary by FRA Region.

Region	CC	SD	MSE	RMSE	MAE	MAPE	R2	Label
Core Issues	1.0000	0.0000	9.000	3.0000	3.0000	4.3067	0.96245	Normal
Winding Bulk Displacement	0.9905	2.1205	188.81	13.741	13.577	18.737	0.03224	Fault
Winding Deformation	0.9776	1.0760	10.798	3.286	3.1058	5.5632	0.50136	Fault
Inter-Turn Shorts & Tap Changer	1.0000	0.0000	9.000	3.0000	3.0000	6.6682	0.75290	Fault
Partial Discharge & Floating Metal	1.0000	0.0000	9.000	3.0000	3.0000	9.9957	0.45604	Fault

provides the region-wise statistical error metrics used to label fault presence during training. Significant deviations were detected, particularly in the Bulk Displacement, Winding Deformation, Inter-Turn Shorts & Tap Changer, and Partial Discharge & Floating Metal regions, indicating mechanical and dielectric anomalies.

3.3.1. Engineering Interpretation & Recommendation

Deviations were detected in four of the five diagnostic regions: Winding Bulk Displacement, Winding Deformation, Inter-Turn Shorts & Tap Changer, and Partial Discharge & Floating Metal. These regions exhibited elevated MAPE values, increased RMSE, and relatively low R² coefficients, confirming the presence of mechanical and dielectric disturbances. Only the Core Issues region remained statistically normal, suggesting that the magnetic core structure has not been significantly affected.

Among all schemes tested, the DWT-SVM approach using the db4 wavelet again emerged as the most reliable method for fault classification under cross-validation. Therefore, it is recommended to adopt DWT-SVM (db4 wavelet) as the standard approach for ongoing transformer diagnostics, ensuring both robustness and scalability for real-world monitoring programs.

3.3.2. Root Cause Analysis Based on FRA Diagnostic Standards

Based on the statistical deviations observed in the FRA for Case Study 2—particularly the high MAPE (18.74%) and RMSE (13.74) in the frequency range of 1 kHz to 10 kHz—the most probable root cause is bulk winding displacement. According to IEC 60076-18 and CIGRÉ WG A2.26 diagnostic guidelines, abnormalities in this frequency band are typically attributed to mechanical movement or loss of clamping pressure within winding assemblies.

The additional deformation noted in the 10 kHz to 100 kHz range suggests axial or radial movement within the winding structure. Furthermore, elevated deviations in the high-frequency region (100 kHz to 1 MHz) imply looseness around tap changers or localized inter-turn effects.

The absence of significant deviation in the core frequency region (10 Hz to 1 kHz) confirms that the magnetic core remains mechanically stable, focusing the diagnosis on winding-related anomalies. Given that all faulted regions correspond to movement-sensitive frequency bands, it is plausible that mechanical stress during factory acceptance tests, transport, or thermal cycling contributed to localized winding misalignment.

Therefore, the root cause of the detected anomalies is attributed to winding bulk movement and axial deformation, compounded by possible mechanical loosening near the offload tap changer. A full mechanical inspection, re-tightening of the winding clamps, and subsequent FRA revalidation following maintenance are strongly recommended to ensure full mechanical recovery and operational reliability.

3.4. Case Study 3

This case focuses on a 0.31 MVA single-phase distribution transformer equipped with an offload tap changer, tested at tap position 1 under factory-filled conditions. The aim was to verify winding integrity using FRA in conjunction with DWT and machine learning classification. The test was conducted over a wide frequency sweep using the Megger FRAX 101, and transformer parameters are shown in

Table 14. Case Study 3—Summary of Key Transformer and Test Parameters.

Attribute	Value
Rated Power (MVA)	0.31
HV Rating (kV)	22
LV1 Rating (kV)	0.4
Impedance HV–LV1 (%)	4.68
Tap Changer Type	Offload Tap Changer
Tap Position During Test	1
Sweep Frequency Range (Hz)	20–1,988,018
Input Resistance (Ω)	50

.

Figure 15 shows the baseline vs test-mode FRA signatures captured from Case 3 using the Megger FRAX 101. Notable divergence is seen in mid and high-frequency bands.

Figure 16 presents the D5-level wavelet decomposition (db4) of the deviation signal, segmented by diagnostic FRA regions.

Table 15. Case Study 3—SVM Trained with coif5 Wavelet (Cross-Validated).

Region	Tested Patterns	Correct	Incorrect
Core Issues	354	212	142
Winding Bulk Displacement	208	125	83
Winding Deformation	208	125	83
Inter-Turn Shorts & Tap Changer	208	125	83
Partial Discharge & Floating Metal	63	38	25

introduces the classification performance of the SVM model trained using the coif5 wavelet for Case Study 3. Following 5-fold cross-validation, the model achieved approximately 40% classification accuracy across the diagnostic regions. Although reduced from initial in-sample estimates, the coif5-based features still demonstrated meaningful capacity to capture fault signatures in the FRA data.

Table 16. Case Study 3—SVM Trained with sym5 Wavelet (Cross-Validated).

Region	Tested Patterns	Correct	Incorrect
Core Issues	354	71	283
Winding Bulk Displacement	208	42	166
Winding Deformation	208	42	166
Inter-Turn Shorts & Tap Changer	208	42	166
Partial Discharge & Floating Metal	63	13	50

presents the results for the SVM model trained with the sym5 wavelet. In this case, the cross-validated classification performance was notably lower, highlighting the sensitivity of SVM training outcomes to the choice of mother wavelet. The sym5 model yielded 20% accuracy across the tested regions, emphasizing the need for careful wavelet selection.

Table 17. Case Study 3—SVM Trained with db4 Wavelet (Cross-Validated).

Region	Tested Patterns	Correct	Incorrect
Core Issues	354	142	212
Winding Bulk Displacement	208	83	125
Winding Deformation	208	83	125
Inter-Turn Shorts & Tap Changer	208	83	125
Partial Discharge & Floating Metal	63	25	38

summarizes the classification results for the db4 wavelet-based SVM. The db4 model achieved an overall cross-validated accuracy of 40%, comparable to that of coif5, reinforcing its diagnostic suitability despite the increased complexity of the fault patterns in this case.

Table 18. Case Study 3—Scheme Comparison Accuracy (Cross-Validated).

Method	Accuracy (%)
DWT-SVM	40
WT-SVM	50
DWT-FFA	16.67
WPT-SVM	40

benchmarks the classification accuracies of four diagnostic schemes under cross-validation. Both DWT-SVM and WPT-SVM achieved 40% classification accuracy, outperforming DWT-FFA, but falling short of WT-SVM performance under this case study, indicating the difficulty posed by subtle transformer anomalies in this specific scenario.

Table 19. Case Study 3—Statistical Summary by FRA Region.

Region	CC	SD	MSE	RMSE	MAE	MAPE	R2	Label
Core Issues	1.0000	0.0000	9.000	3.0000	3.0000	4.5113	0.91653	Normal
Winding Bulk Displacement	1.0000	0.0000	9.000	3.0000	3.0000	4.6573	0.80664	Normal
Winding Deformation	0.9643	1.6630	396.24	19.906	19.837	45.826	−11.396	Fault
Inter-Turn Shorts & Tap Changer	0.9732	1.1787	10.880	3.2985	3.0817	12.213	0.57752	Fault
Partial Discharge & Floating Metal	1.0000	0.0000	9.000	3.0000	3.0000	12.127	0.22778	Fault

compiles the statistical error metrics used for SVM region labeling. Faults were clearly detected in the Winding Deformation, Inter-Turn Shorts & Tap Changer, and Partial Discharge & Floating Metal regions based on elevated MAPE, significant RMSE, and diminished or negative R² values.

3.4.1. Engineering Interpretation & Recommendation

The diagnostic analysis for Case Study 3 reveals clear deviations in the Winding Deformation, Inter-Turn Shorts & Tap Changer, and Partial Discharge & Floating Metal regions, while the Core and Bulk Displacement zones remained statistically within normal parameters. These conclusions are substantiated by elevated MAPE values (>12%), sharp MAE spikes, and poor R² values, notably the extremely negative R² (−11.396) in the Winding Deformation region.

Among the schemes tested, the DWT-SVM classifier utilizing the db4 wavelet remained the most robust diagnostic model under cross-validation conditions, despite the reduction in classification accuracy compared to earlier non-validated estimations. As such, it is recommended to continue using this methodology for assessing this transformer configuration, complemented by offline diagnostic tests such as Winding Resistance Measurements, Dissolved Gas Analysis (DGA), Insulation Resistance Testing, and Visual Inspection.

3.4.2. Root Cause Analysis Based on FRA Standards

The statistical evidence in Case Study 3 points to a significant anomaly in the Winding Deformation region, with a high MAPE value of 45.83%, an elevated RMSE of 19.91, and a negative R² (−11.396). These indicators are consistent with abnormal mechanical behavior, likely arising from axial or radial displacement of winding structures.

According to international diagnostic standards such as IEC 60076-18 and CIGRÉ WG A2.26, deviations in the 10 kHz–100 kHz frequency band typically correspond to winding distortions. Complementary deviations observed in the Inter-Turn Shorts & Tap Changer and Partial Discharge & Floating Metal regions further support the likelihood of compromised dielectric integrity.

The most probable root cause is mechanical shifting or looseness within the winding structure, potentially initiated during handling, factory acceptance testing, or vibration during service operation. Immediate physical inspection and mechanical revalidation are therefore warranted to address the identified risks and to restore full operational reliability.

3.5. Case Study 4

This case study investigates a 10 MVA, three-phase transformer equipped with an on-load tap changer (OLTC), tested at the extreme raise tap position. The aim is to evaluate insulation and winding integrity through FRA supported by statistical signal features and supervised machine learning classification. The test was executed using the Megger FRAX 101 over the standard sweep range of 20 Hz to ~2 MHz. Key transformer and test setup details are summarized in

Table 20. Case Study 4—Summary of Key Transformer and Test Parameters.

Attribute	Value
Rated Power (MVA)	10
HV Rating (kV)	44
LV1 Rating (kV)	11
Impedance HV–LV1 (%)	11.3
Tap Changer Type	On-load Tap Changer
Tap Position During Test	Extreme Raise (Onload)
Sweep Frequency Range (Hz)	20–1,988,018
Input Resistance (Ω)	50

.

Figure 17 shows the frequency domain response comparison between baseline and test traces, highlighting significant variations at both mid and high frequency regions.

Figure 18 illustrates the level-5 decomposition of the FRA deviation signal using the db4 wavelet. The absolute D5 coefficients are plotted across five fault-specific frequency regions to aid in fault isolation.

Table 21. Case Study 4—SVM Trained with coif5 Wavelet (Cross-Validated).

Region	Tested Patterns	Correct Patterns	Incorrect Patterns
Core Issues	354	142	212
Winding Bulk Displacement	208	83	125
Winding Deformation	208	83	125
Inter-Turn Shorts & Tap Changer	208	83	125
Partial Discharge & Floating Metal	63	25	38

summarizes the classification performance of the SVM model using features extracted from the coif5 wavelet. After cross-validation, the classifier achieved moderate performance, correctly identifying 40% of the patterns across all fault categories, suggesting the need for robustness improvements under real-world deployment conditions.

The sym5-based SVM model exhibited comparable behavior, maintaining 40% classification accuracy across all diagnostic regions after cross-validation, highlighting the consistent yet moderate discrimination ability of sym5 features under statistical variation. The results of the SVM trained with sym5 Wavelet are shown in

Table 22. Case Study 4—SVM Trained with sym5 Wavelet (Cross-Validated).

Region	Tested Patterns	Correct Patterns	Incorrect Patterns
Core Issues	354	142	212
Winding Bulk Displacement	208	83	125
Winding Deformation	208	83	125
Inter-Turn Shorts & Tap Changer	208	83	125
Partial Discharge & Floating Metal	63	25	38

.

Similarly, the db4-based classifier achieved a 40% cross-validation accuracy across the tested regions. Despite its lower performance under validation compared to previous results, db4 remains favored for its better feature compactness and computational efficiency in fault feature extraction. The cross-validated results of the SVM trained with db4 wavelet are shown in

Table 23. Case Study 4—SVM Trained with db4 Wavelet (Cross-Validated).

Region	Tested Patterns	Correct Patterns	Incorrect Patterns
Core Issues	354	142	212
Winding Bulk Displacement	208	83	125
Winding Deformation	208	83	125
Inter-Turn Shorts & Tap Changer	208	83	125
Partial Discharge & Floating Metal	63	25	38

.

A comparison among different diagnostic schemes is presented in

Table 24. Case Study 4—Scheme Comparison Accuracy (Cross-Validated).

Method	Accuracy (%)
DWT-SVM	40
WT-SVM	50
DWT-FFA	16.67
WPT-SVM	40

. The DWT-SVM and WPT-SVM methods each achieved a cross-validated accuracy of 40%, while traditional WT-SVM and DWT-FFA approaches performed less competitively.

The region-based statistical deviation analysis is summarized in

Table 25. Case Study 4—Statistical Summary by FRA Region.

Region	CC	SD	MSE	RMSE	MAE	MAPE	R2	Label
Core Issues	1.0000	0	9.000	3.000	3.000	6.6968	0.91099	0
Winding Bulk Displacement	1.0000	0	9.000	3.000	3.000	5.8349	0.91553	0
Winding Deformation	1.0000	0	9.000	3.000	3.000	7.9835	0.86153	0
Inter-Turn Shorts & Tap Changer	0.98334	1.1787	398.12	19.953	19.918	49.907	–10.283	1
Partial Discharge & Floating Metal	0.27609	2.1418	15.206	3.8995	3.2698	9.7595	–60.933	1

. Significant deviations were observed particularly in the Inter-Turn Shorts & Tap Changer and Partial Discharge & Floating Metal regions, based on elevated MAPE values and strongly negative R² metrics.

3.5.1. Engineering Interpretation & Recommendation

The transformer evaluated in Case Study 4 demonstrated localized deviations predominantly in the Inter-Turn Shorts & Tap Changer and the Partial Discharge & Floating Metal regions. Despite moderate overall classification rates under cross-validation, the statistical deviation metrics strongly flagged these regions as abnormal. In particular, the elevated MAPE (49.91% and 9.76%) and the markedly negative R² values (−10.283 and −60.933) confirm significant disturbances.

Thus, it is strongly recommended to initiate offline diagnostics, specifically Winding Resistance Testing, Dissolved Gas Analysis (DGA), Insulation Resistance Testing, and isual Inspection of bushings and OLTC (On-Load Tap Changer) structures. Given its consistent diagnostic separation and statistical robustness, the DWT-SVM (db4 wavelet, cross-validated) approach remains the preferred method for this transformer configuration and should be adopted for future health assessments.

3.5.2. Root Cause Analysis Based on FRA Standards

According to IEC 60076-18 [40] and CIGRÉ WG A2.26 [41] standards, deviations in the high-frequency bands (100 kHz–2 MHz) are closely associated with inter-turn short circuits and floating metallic particles. The strong attenuation, increased phase distortion, and poor signal correlation detected in this band suggest insulation degradation or metallic contamination near the OLTC or bushing terminals. Since the mid-frequency zones remained stable, core displacement and major winding deformation are unlikely causes. Thus, the most probable root cause is localized insulation degradation combined with ungrounded metallic particles or early tap changer wear, requiring corrective maintenance.

3.6. Case Study 5

This case study analyzes a 20 MVA, three-phase transformer equipped with an on-load tap changer, tested under oil-filled conditions at the extreme raise position. FRA measurements were captured across a broadband frequency sweep (20 Hz to ~2 MHz) using the Megger FRAX 101 device.

Table 26. Case Study 5—Summary of Key Transformer and Test Parameters.

Attribute	Value
Rated Power (MVA)	20
HV Rating (kV)	88
LV1 Rating (kV)	11
Impedance HV–LV1 (%)	11.6
Tap Changer Type	On-load Tap Changer
Tap Position During Test	Extreme Raise (Onload)
Sweep Frequency Range (Hz)	20–1,988,018
Input Resistance (Ω)	50

presents a summary of the key transformer nameplate and test parameters for this unit.

Figure 19 presents the FRA signatures captured for Case Study 5 using the Megger FRAX 101 analyzer. The baseline and fault-mode curves reveal notable deviations, particularly in the high-frequency domain, indicating abnormalities in specific transformer regions.

Figure 20 illustrates the wavelet decomposition (db4, D5 level) of the FRA deviation signal across five diagnostic frequency regions. Elevated coefficients are observed in the Partial Discharge & Floating Metal region, aligning with statistical findings, and supporting fault localization.

The SVM was first trained using features extracted with the Coif5 wavelet. As shown in

Table 27. Case Study 5—SVM Trained with Coif5 Wavelet (Cross-Validated).

Region	Tested Patterns	Correct	Incorrect
Core Issues	354	283	71
Winding Bulk Displacement	208	166	42
Winding Deformation	208	166	42
Inter-Turn Shorts & Tap Changer	208	166	42
Partial Discharge & Floating Metal	63	50	13

, a cross-validated accuracy of approximately 80% was achieved across the five diagnostic regions. While not perfect, the results demonstrate high model generalization with Coif5 features under more realistic validation conditions.

The Sym5 wavelet was also evaluated as a feature extractor. As illustrated in

Table 28. Case Study 5—SVM Trained with Sym5 Wavelet (Cross-Validated).

Region	Tested Patterns	Correct	Incorrect
Core Issues	354	283	71
Winding Bulk Displacement	208	166	42
Winding Deformation	208	166	42
Inter-Turn Shorts & Tap Changer	208	166	42
Partial Discharge & Floating Metal	63	50	13

, the cross-validated model achieved identical performance to the Coif5 case, confirming the stability of the multi-resolution wavelet-based training approach.

Table 29. Case Study 5—SVM Trained with db4 Wavelet (Cross-Validated).

Region	Tested Patterns	Correct	Incorrect
Core Issues	354	283	71
Winding Bulk Displacement	208	166	42
Winding Deformation	208	166	42
Inter-Turn Shorts & Tap Changer	208	166	42
Partial Discharge & Floating Metal	63	50	13

presents the final and selected model based on db4 wavelet decomposition. Consistent with previous wavelets, approximately 80% cross-validation accuracy was achieved, confirming the robustness of db4-derived features under real-world noise and variation.

Table 30. Case Study 5—Scheme Comparison Accuracy (Cross-Validated).

Method	Accuracy (%)
DWT-SVM	80
WT-SVM	50
DWT-FFA	16.67
WPT-SVM	80

compares the performance of four diagnostic schemes. DWT-SVM and WPT-SVM achieved superior accuracies (80%), notably outperforming WT-SVM and DWT-FFA.

Finally,

Table 31. Case Study 5—Statistical Summary by FRA Region.

Region	CC	SD	MSE	RMSE	MAE	MAPE	R2	Label
Core Issues	1.0000	0	9.000	3.000	3.000	48.695	0.83183	Fault
Winding Bulk Displacement	1.0000	0	9.000	3.000	3.000	8.6295	0.75824	Fault
Winding Deformation	1.0000	0	9.000	3.000	3.000	5.8521	0.75396	Fault
Inter-Turn Shorts & Tap Changer	1.0000	0	9.000	3.000	3.000	9.7571	0.87226	Normal
Partial Discharge & Floating Metal	0.5548	3.4017	885.86	29.76	29.57	130.12	−175.23	Fault

reports the region-wise statistical metrics used for fault labeling. Severe deviations are particularly noted in the Partial Discharge & Floating Metal region, with a very high MAPE of 130.12% and an extremely negative R² of −175.23, clearly indicating a critical abnormality.

3.6.1. Engineering Interpretation & Recommendation

In Case Study 5, significant deviations were identified in four of the five diagnostic regions. The most severe anomaly was observed in the high-frequency Partial Discharge & Floating Metal region, characterized by an extremely high MAPE (130.12%) and a highly negative R² (−175.23). Additional deviations were detected in the Core Issues, Winding Bulk Displacement, and Winding Deformation regions, indicating widespread mechanical and dielectric stress.

The DWT-SVM classifier using db4 wavelet features consistently identified these faults with 80% cross-validated accuracy, even under field-realistic conditions. Based on the dispersion and severity of these anomalies, it is strongly recommended that the transformer be subjected to extensive offline diagnostics, including Dissolved Gas Analysis (DGA), insulation resistance measurement, and comprehensive visual inspection. For continuous health monitoring, the DWT-SVM (db4) methodology should be standardized for this transformer type.

3.6.2. Root Cause Analysis Based on FRA Standards

The observed abnormalities in the Partial Discharge & Floating Metal region (frequency > 1 MHz) are consistent with critical dielectric degradation or the presence of loose metallic particles. According to IEC 60076-18 [40] and CIGRÉ WG A2.26 [41], such severe high-frequency disturbances typically point toward internal corona activity, floating conductors, or compromised insulation interfaces. Furthermore, mild but cumulative deviations in the lower frequency bands (Core and Winding Deformation regions) suggest accumulated mechanical stresses, possibly due to prolonged thermal cycling or mechanical shocks during transportation. Accordingly, the transformer should be promptly decommissioned for internal inspection, re-tensioning of structural elements, insulation reinforcement, and post-maintenance FRA validation.

While the db4 wavelet demonstrated superior diagnostic consistency, it is acknowledged that using a single wavelet type may limit the capture of certain frequency features. Alternative wavelets such as coif5 and sym5 were also evaluated, revealing minor variations in classification performance, but db4 offered the best trade-off between sensitivity and interpretability.

4. Discussion

The five case studies presented in Section 3 provide a comprehensive evaluation of the proposed DWT-SVM-based fault detection framework for transformer winding analysis. Each case involved real-world FRA measurements from operational transformers with varying ratings and fault profiles.

Across all cases, the diagnostic strategy—combining multi-level wavelet decomposition (particularly db4 mother wavelet) with statistical deviation labeling and SVM classification—demonstrated strong potential for condition monitoring applications. However, several trends and considerations emerge from the cross-validated performance results that merit further discussion.

Firstly, while initial non-validated results often suggested near-perfect fault discrimination, the introduction of 5-fold cross-validation revealed a more realistic diagnostic accuracy in the range of 40% to 80%, depending on case complexity. In particular, Case Studies 1, 2, and 5 maintained relatively high cross-validated accuracy (around 80%), indicating strong generalization when clear deviations were present in the frequency response data. In contrast, Case Studies 3 and 4 exhibited reduced accuracies (approximately 40%), largely attributed to more subtle mechanical and dielectric anomalies that were harder to isolate based on wavelet features alone.

Secondly, the db4 wavelet consistently outperformed coif5 and sym5 across all case studies, reinforcing its suitability for FRA signal decomposition in transformer diagnostics. Its compact support, excellent localization properties, and proven robustness against noise made it the preferred mother wavelet for this application.

Thirdly, the region-based statistical deviation labeling approach proved critical for improving diagnostic interpretability. By focusing classification on specific frequency regions tied to known physical phenomena (e.g., core clamping in low-frequency bands; inter-turn shorts in high-frequency bands), the model achieved enhanced fault localization compared to global FRA signature comparisons.

Additionally, the GUI developed for this framework significantly streamlined the diagnostic workflow. Practical testing across all five transformers demonstrated that complete diagnosis—including feature extraction, classification, and automated PDF reporting—could be performed within 8–12 s per transformer on a standard engineering workstation. No operational bottlenecks were observed, confirming the tool’s suitability for real-time or near-real-time deployment in substation environments.

Nonetheless, some limitations were observed. The drop in classification performance under subtle fault scenarios (as seen in Case Studies 3 and 4) suggests that future work should explore data augmentation techniques, ensemble classifiers, or hybrid feature sets (e.g., combining time-frequency and statistical-domain features) to enhance fault separability.

Finally, all engineering interpretations and root cause analyses remained compliant with IEC 60076-18 and CIGRÉ WG A2.26 standards, ensuring that diagnostic conclusions are technically defensible and actionable for utilities.

5. Conclusions

This work developed and validated a discrete wavelet transform (DWT) and support vector machine (SVM) framework for fault detection in power transformer windings based on frequency response analysis (FRA) measurements. Five case studies were analyzed, covering different fault scenarios, including core issues, winding bulk displacement, winding deformation, inter-turn shorts, and partial discharge phenomena.

Across all five case studies, the DWT-SVM approach using the db4 mother wavelet demonstrated the most reliable classification performance under cross-validation. Although classification accuracy varied depending on the specific fault region and case study, the proposed scheme consistently outperformed traditional methods such as WT-SVM and DWT-FFA, confirming its robustness and diagnostic value in practical applications.

Specifically, the DWT-SVM framework achieved cross-validated classification accuracies ranging from 40% to 80% across the five real-world cases, whereas conventional WT-SVM and DWT-FFA methods achieved lower accuracies of approximately 50% and 16.67%, respectively. Statistical evaluation also revealed that regions labeled as faulted exhibited substantial deviation metrics, including MAPE values exceeding 18% and R² coefficients as low as –175.23. These quantitative findings reinforce the superior diagnostic fidelity and generalization capability of the proposed framework compared to traditional FRA analysis techniques.

The region-based statistical deviation labeling methodology enabled a systematic distinction between normal operational deviations and true fault signatures, improving diagnostic precision. Furthermore, engineering interpretations and root cause analyses were aligned with IEC 60076-18 and CIGRÉ WG A2.26 standards, ensuring that the framework’s findings are technically defensible and applicable to field practice. The developed MATLAB-based diagnostic tool, which integrates automated feature extraction, classification, reporting, and interpretation modules, facilitates efficient condition assessment and supports maintenance decision-making processes. Based on the validation results across the five case studies, the DWT-SVM (db4) scheme is recommended for standard adoption in transformer health monitoring workflows, with future enhancements focused on online FRA integration and broader fault-type coverage.

For future real-time deployments, the DWT-SVM framework’s computational efficiency can be enhanced through model compression techniques, parallel processing, and selective feature pruning, ensuring suitability for online FRA-based condition monitoring without compromising diagnostic accuracy. Future work may explore advanced signal decomposition techniques such as Empirical Wavelet Transform (EWT) to potentially enhance feature extraction and fault detection sensitivity.