Changbing Luo1, Hongwen Liu2*, Guochao Qian2, Zhengqiang He1, Kun Yang2, Wangjun Wu1, Weizhao Yang1, Jianzhong Yao1, Min Zhang1, Lilong Liu1
1Diqing Power Supply Bureau, Yunnan Power Grid Co., Ltd., Diqing, 674400, China
2Electric Power Research Institute of Yunnan Power Grid Co., Ltd., Kunming 650214, China
*Corresponding Author’s Email: liuhongwen_2025@126.com
Abstract: Transformers are core equipment in power transmission and distribution systems, and the condition of their windings significantly impacts their reliable operation. Therefore, transformer winding fault diagnosis holds great significance. Frequency response analysis (FRA) has been widely accepted as a powerful tool for transformer winding fault diagnosis. However, most existing FRA-based methods extract limited feature information, which restricts the accuracy and application prospects of winding fault diagnosis methods. Therefore, this paper proposes a transformer winding fault diagnosis method based on multi-scale two-dimensional recurrent features and the EMA-MobileNetV3 network model. First, a multi-scale, signed, and asymmetric Recurrence Plot (MSRP) transformation technique is proposed to convert frequency response curves into two-dimensional images, effectively enhancing the multi-scale information of winding fault features. Subsequently, an efficient multi-scale feature attention MobileNetV3 model is proposed for winding fault diagnosis, enabling rapid and accurate identification of winding states. An experimental platform for transformer winding fault diagnosis is then established, where frequency response data under different states are obtained through frequency response testing, and winding fault diagnosis is performed, including fault type classification and fault severity assessment. The results show that the proposed method achieves an accuracy rate of 97.50% and 96.25% in terms of type and severity, outperforming other methods. Finally, the trained model was applied to field transformers to validate the method’s effectiveness on actual transformers.
Keywords: Transformer winding, fault diagnosis, MSRP, frequency response analysis (FRA), EMA-MobileNetV3
- Introduction
As a core infrastructure component of power systems, power transformers play a critical role in the transmission and distribution of energy. Their operational reliability directly impacts the quality of power transmission across the entire grid [1,2]. Statistical analysis of fault cases indicates that approximately 30% of transformer failures originate from internal windings [3,4]. Therefore, accurately identifying and assessing the condition of transformer windings is of great significance for enhancing the safety and reliability of power systems.
Currently, the primary methods for detecting the condition of transformer windings include dissolved gas analysis (DGA) [5], vibration analysis [6], and frequency response analysis (FRA). FRA exhibits higher sensitivity to transformer winding faults compared to other methods [7] and is also one of the most widely used methods for on-site detection. The FRA method involves injecting a sinusoidal signal with a scanning frequency into the transformer winding and measuring the response signal in the frequency domain to construct the frequency response characteristic [8]. Research indicates that transformer windings can be described as a two-port circuit network composed of resistance, capacitance, and inductance in the high-frequency range, and their frequency response characteristic (Vout/Vin expressed in decibels) can characterize the operational state of the winding [3]. For FRA, standards such as IEC 60076-18-2012 and C57.149-2018 have been established, which indicate that correlation coefficients can be used to identify the state of the windings [9], [10], but they can only provide a general assessment of the windings and cannot accurately distinguish the type and severity of winding faults. Therefore, in addition to the correlation coefficients mentioned in the standards, domestic and international scholars have conducted extensive research on feature extraction and fault identification methods for FRA results. In terms of feature extraction, domestic and international scholars have studied various aspects, including statistical features, waveform features, and image features. Specifically. Wang et al. used a fast relaxation method to fit the characteristic parameters of the frequency response curve, thereby identifying the state of the transformer winding [11]. In some studies, the root mean square (RMS) and standard deviation features of the amplitude-frequency curve were extracted to judge the state of the winding, and it was found that these feature values have a certain degree of resistance to external noise interference[12]. Yousof et al converted the frequency response curve into a Nyquist plot and discussed the correspondence between the Nyquist plot and the transformer winding state[13]. Zhao et al performed binary processing on healthy and faulty response curves and extracted feature coefficients based on the similarity of image matrices to achieve the identification of three common winding fault types[14]. Aljohani et al converted the frequency response curve into a polar coordinate diagram and applied digital image processing (DIP) techniques to extract image features such as geometry, texture, and invariant moments, overcoming the limitation of traditional FRA feature extraction methods, which provide relatively limited information[15]. Therefore, converting the frequency response curve into a two-dimensional image and then extracting features from the image is currently a more advantageous method.
However, most current feature extraction methods, including numerical statistical features and image features, still compress frequency response features into a single value or array using defined metrics, which limits the extraction of feature information from frequency response curves and results in suboptimal diagnostic accuracy [16].
Additionally, to further extract FR curve features and improve diagnostic accuracy, some machine learning models [17-19] have been introduced into transformer fault identification based on frequency response curves. Artificial neural networks (ANN) were applied to the intelligent diagnosis of transformer winding frequency response, enabling the identification of fault severity and location[17]. A multi-layer perceptron (MLP) was used to interpret the characteristics in the frequency response curve and detect transformer winding short-circuit faults. Support vector machine (SVM) methods have shown great promise in transformer winding fault diagnosis[18]. SVM was introduced into transformer FRA diagnosis. This study diagnosed three types of transformer fault states and achieved good results[19]. Zhou et al applied SVMs to detect faults and their severity[20]. Bigdeli et al applied machine learning techniques including probabilistic neural networks, k-nearest neighbor algorithms, decision trees, and SVMs to identify transformer winding fault types. The results indicated that SVMs were more accurate in fault classification[21]. However, SVM is inherently a binary classifier. As the number of fault types increases, it faces numerous diagnostic challenges [22]. Additionally, although these intelligent methods have achieved some success in transformer winding fault diagnosis, their performance remains heavily dependent on manually defined numerical feature metrics, resulting in diagnostic results being highly constrained by the selection and quality of prior features.
Excitingly, some methods combining data visualization and deep learning have been introduced into the field of transformer FRA diagnosis [22,23]. a Hilbert visualization method was proposed to further visualize the frequency response features extracted by the sliding window and combine them with a pre-trained CNN network to achieve winding fault location[22]. Similarly, the Gram angle field visualizes the frequency response feature value sequence extracted by the sliding window and uses a basic CNN network for winding amplitude classification [23]. The lightweight network MobileNet-V2 is also used to identify images after frequency response visualization, significantly improving the accuracy of fault type identification[24],. It is worth noting that these visualization methods all use sliding windows to extract defined metric features, and the frequency response information of the winding still has certain limitations. ShuffleNet is used as a diagnostic model to diagnose images obtained by simple transformations of the frequency response curve[25]. a simple 2D-CNN model is proposed to identify interturn short-circuit faults, achieving high accuracy[26].
As summarized in the literature review, statistical and numerical indicators are widely used as tools for extracting features from frequency response test results. However, the application of these methods requires complex mathematical calculations and experienced personnel. Most visualization methods remain limited to implementing defined metric feature sequences, resulting in restricted information richness within the images. Additionally, such images often contain a significant amount of unused space or image symmetry redundancy. When training CNN networks with such images, the subtle features of the FR curves in the images are difficult to fully identify. The CNN networks employed are relatively basic, making it challenging to effectively extract subtle features from frequency responses, and their computational efficiency and performance require further consideration, thereby limiting the accuracy and field application prospects of winding fault diagnosis methods.
To address the issues identified in the aforementioned methods, this paper proposes a transformer-based winding fault diagnosis method for transformers, utilizing multi-scale, signed, and asymmetric recurrence plot feature images combined with Efficient Multi-scale Attention—MobileNetV3. In summary, the main innovations of this paper are as follows:
1) Innovatively introducing the MSRP method to transform one-dimensional frequency response curves into two-dimensional images, achieving full visualization of frequency response features and changes, and enhancing the expression of winding fault features. 2) Proposing a winding diagnosis model named EMA-MobileNetV3 to achieve multi-scale extraction of fault information and precise diagnosis. Through comparative experiments, the performance of the proposed model in winding fault type identification and severity assessment is demonstrated.
The rest of the paper is structured as follows. First, the key tools and models used in this study are introduced in Section Ⅱ. The winding fault experimental platform and data acquisition are described in Section Ⅲ. Classification results and analysis are presented in Section Ⅳ. Section Ⅴ validates the effectiveness of the proposed method on field transformers. Finally, the summary of this work is presented in Section Ⅵ.
- PROPOSED DIAGNOSIS METHOD
2.1 Overview
To achieve high-precision diagnosis of winding faults, this study proposes an advanced winding fault identification method, as shown in Figure. 1, with the following main steps.
Step1：Obtain frequency response data of the winding in healthy and different defect states through the experimental platform and frequency response tester, and normalize all frequency response curves to [0,1].
Step2：Divide the frequency response curve into two equal parts, apply the SRP transformation to each part to obtain MSRP feature images, and construct a winding status database. The MSRP-transformed feature images provide additional features that cannot be observed in 1-D frequency response data, enabling intuitive identification and classification of fault modes.
Step3：The EMA-MobileNetV3, which integrates the efficient multi-scale attention mechanism, is used as an efficient feature extractor to obtain more distinguishable features of fault types and severity from the input MSRP images. The extracted representative feature vectors are then used for classification to obtain the final transformer winding fault classification and identification results.
Step4：Apply the trained model to field transformers to diagnose the type and severity of winding faults.
Figure. 1 Main process framework of the proposed method
2.2 Overview fault feature enhancement based on MS recurrence plot
A Recurrence Plot (RP) is a method for visualizing the periodicity of trajectories in phase space, and has been successfully applied in fields such as motor bearing signal processing and fault diagnosis. This method preserves the correlation information and nonlinear characteristics between amplitude and phase in the original one-dimensional signal, and contains richer phase space information, enabling the identification of minute local differences in one-dimensional time series [27,28] .Feature information based on two-dimensional recursive graphs has been proven to effectively improve the classification accuracy of models. However, it has disadvantages such as trend confusion and information symmetry.
This study innovatively introduces an improved RP method—multi-scale, signed, and asymmetric Recurrence Plot (MSRP)—into transformer winding fault diagnosis [29], addressing issues such as feature scale differences, information redundancy, and trend confusion that exist in traditional RP methods. The specific steps are as follows.
This study innovatively introduces an improved RP method—multi-scale, signed, and asymmetric Recurrence Plot (MSRP)—into transformer winding fault diagnosis, addressing issues such as feature scale differences, information redundancy, and trend confusion that exist in traditional RP methods. The specific steps are as follows：
1) First, for the original frequency response curve X′i = {x′1, x′2, …, x′n}，normalize the data X′ to the range [0,1] using the following formula:
The primary purpose of normalizing data is to eliminate the adverse effects caused by differences in units or orders of magnitude between different variables, thereby ensuring the accuracy and reliability of subsequent calculations.
2) Asymmetric design. Since the original RP image is symmetrical along the main diagonal and contains redundant information, asymmetric RP is proposed. Specifically, the normalized frequency response curve X is divided into two parts.
3) Multi-scale design. Select an appropriate state space dimension d to control the number of state points, and an appropriate embedding delay 𝜏 control the interval between adjacent points in the state space, thereby extracting multi-scale fault features in the frequency response and enhancing the expressive power of the RP method under limited image size.
4) Perform a phase space transformation on each 1-D frequency response subcurve Xi = {x1, x2, …, xm} from the initial reconstruction, then obtain the reconstruction matrix Xm(d, 𝜏). The specific representation is as follows
where d is the embedding dimension, 𝜏 is the Embedded delay , Xm(d, r) is related to d and r and represents a frequency response series of trajectories in the phase space, and xm+(d-1)r represents the value of the series X at point m+(d-1)r.
5) The corresponding image features are obtained by calculating the Euclidean distance between states.
where Xi and Xj are the points of the phase space transformation and ||•|| is the Euclidean paradigm.
6) Distance matrix symbolization processing. To solve the problem of trend confusion, the extracted symbol mask is multiplied by RP to produce a signed RP, generating a symbol-enhanced distance matrix S.
7) SRP coding is performed independently on each subcurve, and the results are summarized into an image set in its own half-interval. Then, each half image is merged to create an asymmetric image. In this way, the symmetry of the image is reduced, information redundancy is lowered, and the feature expression ability of the image is improved, thereby achieving the representation and fusion of multi-scale features.
8) Finally, to avoid information loss when converting one-dimensional signals to two-dimensional signals, each element in the distance matrix is multiplied by 255 and mapped to different RGB values according to different values, ultimately generating a color texture map. This MSRP image can effectively retain the complete feature information of the original time series.
The conversion of MSRP is shown in Figure. 2. This method can effectively reveal the similarity, chaos, and non-stationarity of the frequency response curve, enhance the response characteristics of winding faults, and provide a basis for subsequent winding fault identification.
Figure. 2 Steps for converting a MSRP image into a frequency response curve
2.3 Multi-scale feature extraction and diagnosis of windings based on EMA-MobileNetV3
2.3.1 MobileNetV3
MobileNetV3[30] is a residual neural network based on a linear bottleneck optimization strategy, consisting of multiple linear bottleneck blocks and efficient layers. Pointwise convolution is a 1×1 ordinary convolution used for feature channel expansion and compression. Deep convolution is a convolution that uses one convolution kernel on each input channel.
1) Bneck block: MobileNetV3 uses separable convolutions to divide conventional convolutions into depth convolutions and point convolutions, as shown in Figure. 3.
The steps for extracting multi-scale features using Bneck are as follows: First, the input features are passed through a “1×1” convolution to extract more high-dimensional feature information using depth-separable convolution, and global information is obtained through the SE attention block. Finally, a point convolution operation is applied to convert the feature map back to the same dimension as the input. The features extracted by the Bneck block in MobileNetV3 are shown in the following formula:
In this context, x represents the input feature，F(x)denotes the output feature, fNL s the nonlinear activation function, D refers to the depth-separable convolution, P signifies the pointwise convolution，and C denotes the standard convolution layer. SE represents the squeezing and excitation module. s denotes the convolution stride. Add(•) denotes an addition operation.
Figure. 3 Depthwise separable convolution diagram
2) Efficient-Layer: In order to preserve high-dimensional features as much as possible and reduce inference latency, MobileNetV3 modified the end bneck, moving the last convolutional layer after the final average pooling layer. This set of final features is now calculated with a spatial resolution of 1×1. The efficient layer reduces computational costs with almost no loss of accuracy. Finally, the features output by the Efficient-Layer are fed into a fully connected layer for fault classification.
However, the traditional MobileNetV3 still has the following issues: 1) The compression of model size and depth limits the channel capacity and receptive field diversity of its feature maps, weakening the network’s ability to model cross-scale semantic associations. This makes it difficult to fully capture fine-grained, long-range dependent features in feature images and poses challenges in understanding the relationships between different regions in an image; 2) Although multi-scale features within Bneck are extracted using different convolution kernels, the SE attention mechanism only performs global compression-excitation operations along the channel dimension, ignoring spatial interactions. This leads to significant loss of discriminative information in complex scenes.
2.3.2 Efficient Multi-Scale Attention Module
The EMA module [31] was introduced to replace the SE attention in the Bneck block, with the following objectives: first, to recalibrate the channel weights in each parallel branch by encoding global information; second, to establish inter-dimensional dependencies by capturing cross-dimensional interactions, thereby effectively enhancing the model’s ability to learn recursive image features with subtle differences; and finally, to fuse multi-scale feature information through cross-spatial learning methods, thereby enhancing the model’s ability to represent important features. EMA is a lightweight and efficient multi-scale attention mechanism that significantly improves model performance without significantly increasing the number of model parameters. The module structure of EMA is shown in Figure. 4. The EMA attention mechanism achieves this through multi-scale feature interaction (simultaneously capturing local and global context) and cross-spatial information fusion (establishing efficient information flow channels across different spatial dimensions).
The specific calculation process of the EMA module is as follows:
First, for a given input feature X, X is divided into g groups of sub-features in the channel dimension to learn different information, so that each sub-feature map can learn different feature representations in a targeted manner, thereby enhancing the model’s ability to perceive multi-scale features of MSRP images.
Subsequently, to extract the attention weights of the group feature maps, three parallel processing paths were adopted: two branches based on 1×1 convolutions and one branch based on 3×3 convolutions. Specifically, in the 1×1 branch, a two-dimensional global average pooling operation is used to encode channel information, generating two tensors. These two tensors are then adjusted through a Sigmoid activation function to obtain the corresponding encoding weights, thereby acquiring precise spatial information. The 1×1 branch aims to capture long-range dependencies between channels and preserve interactions between channels. The 3×3 branch aggregates local spatial structural information using a 3×3 convolution kernel to capture multi-scale features of the MSRP image of the winding. This branch helps capture richer local spatial information and enhances fault feature representation capabilities.
Finally, the EMA module utilizes a cross-spatial learning mechanism to aggregate information from different branches, enabling cross-dimensional interaction. Specifically, a two-dimensional global average pooling operation is performed on the feature maps output by the 1×1 convolution branch and the 3×3 convolution branch to encode global spatial information. Subsequently, the pooled features are reconstructed and normalized using the Softmax function. Finally, feature information from the two branches is fused through element-wise multiplication operations to generate an output feature map containing complete spatial position information. This cross-spatial learning method not only captures pixel-level pairwise relationships but also highlights the global contextual information of all pixels, enabling the EMA module to effectively construct short-range and long-range dependencies without reducing the channel dimension, thereby enhancing the fault feature extraction capability of the MobileNetV3 network. The structure of EMA-MobileNetV3 is shown in Figure. 5.
Figure. 4 The EMA attention model structure diagram
Figure. 5 EMA-MobileNetV3 structure
2.3.3 Focal Loss function
Winding fault data is typically unbalanced. To mitigate the impact of sample category imbalance on the performance of the recognition model, the focal loss function is used in place of the cross-entropy loss function to reduce the weight of easily classified samples and improve the recognition accuracy of unbalanced samples. The loss function expression is as follows.
In summary, the EMA-MobileNetV3 proposed in this work incorporates an efficient multi-scale attention mechanism, which not only captures pixel-level long-range dependencies without reducing the number of channels, but also enhances fine-grained feature interactions across space and channels. This enables the suppression of background noise and significantly highlights the weak fault features and their severity in the frequency response of transformer windings.
- WINDING DIGNOSTIC EXPERIMENTAL PLATFORM
3.1 Test platform setup
This study established a transformer winding fault testing platform and used Megger FRAX99 to test the winding frequency response curve, with matching impedance of 50Ω for both input and output. The basic parameters of the experimental transformer are shown in Table 1. Considering the difficulty of the experiment and the repeatability of the test, the fault simulation was performed on the high-voltage (HV) winding. The test diagram is shown in Figure. 6.
Table 1 Main Parameters of Experimental Transformer
| parameter | number |
| core height | 530 mm |
| Low-voltage winding- height | 474 mm |
| Low-voltage winding- outer radius | 89mm |
| Low-voltage winding- number of turns per coil | 16 |
| High-voltage winding- height | 474 mm |
| High-voltage winding- outer radius | 162 mm |
| High-voltage winding- number of turns per coil | 32 |
Figure. 6 Test Schematic
(a) AD (b) SCV (c) SC
Figure. 7 Fault simulation
3.2 Fault simulation testing
The definitions of fault types, degree, and location in this study are shown in Table 2. The fault simulation diagram is shown in Figure. 7.
1) Axial displacement
Axial displacement faults in the winding were simulated by inserting shims of different heights, as shown in Figure 8(a). The severity of the AD fault is given by.
where, Δh is the axial displacement height，and h is the total height of the winding.
Axial displacement faults of varying severity were simulated in the upper, middle, and lower sections of the experimental transformer winding, with three fault levels defined, as shown in Table 2. Each fault type was tested ten times, resulting in a total of 90 experimental groups.
The amplitude-frequency characteristic curve obtained from the test is shown in Figure. 8. “AD-T-1” indicates a level 1 severity axial displacement fault in the upper part of the winding, and the other curves follow suit.
2) Series capacitance variation
In this study, SCV was used to simulate the radial insulation degradation and local warping of windings commonly occurring in transformers[20]. The fault simulation was conducted by paralleling capacitors ranging from 100 to 680 pF between the windings, as illustrated in Figure.8(b). Capacitors of different sizes were connected in parallel between the upper, middle, and lower windings of the test transformer, with three fault levels set, as detailed in Table 3. Each fault type was tested ten times, resulting in a total of 180 experiments.
The frequency response curve obtained from the test is shown in Figure.8(b). “SCV-M-220” indicates a severe bulging and warping fault of level 1 in the middle of the winding. The meanings of other curves are similar.
3) Short circuit
In this study, fault simulation was performed using short-circuited coil windings, with the fault simulation diagram shown in Figure.8(c). Different numbers of coil windings were short-circuited at the upper, middle, and lower parts of the experimental transformer winding, with three fault levels set, as shown in Table 4. Each fault type was tested ten times, resulting in a total of 90 experiments.
Table 2 Winding axial displacement fault setting
| Location | degree of AD | Degree of fault | Location | degree of AD | Degree of fault | Location | degree of AD | Degree of fault |
| Top | 1% | Level 1 | Middle | 1% | Level 1 | Bottom | 1% | Level 1 |
| 2% | Level 2 | 2% | Level 2 | 2% | Level 2 | |||
| 3% | Level 3 | 3% | Level 3 | 3% | Level 3 |
Table 3 Winding Drumming and Warping Fault Settings
| Location | Capacitance（pF） | Degree of fault | Location | Capacitance（pF） | Degree of fault | Location | Capacitance（pF） | Degree of fault |
| Top | 100&220 | Level 1 | Middle | 100&220 | Level 1 | Bottom | 100&220 | Level 1 |
| 330&470 | Level 2 | 330&470 | Level 2 | 330&470 | Level 2 | |||
| 560&680 | Level 3 | 560&680 | Level 3 | 560&680 | Level 3 |
Table 4 Fault setting for short circuit between pancake windings
| Location | SC count | Degree of fault | Location | SC count | Degree of fault | Location | SC count | Degree of fault |
| Top | Level 1 | Level 1 | Middle | Level 1 | Level 1 | Bottom | Level 1 | Level 1 |
| Level 2 | Level 2 | Level 2 | Level 2 | Level 2 | Level 2 | |||
| Level 3 | Level 3 | Level 3 | Level 3 | Level 3 | Level 3 |
(a) AD faults (b) SCV faults (c) SC faults
Figure. 8 Frequency Response Curve
3.3 Fault simulation testing
This study converts and visualizes the FRA curves obtained from the experimental platform tests using MSRP to enhance winding fault characteristics. Taking a fault in the middle region of the winding as an example, the MSRP images of different winding fault types and fault severities are shown in Figure. 9 , Figure. 10 , and Figure. 11.
(a) 1% AD (b) 2% AD (c) 3% AD
Figure. 9 Images of different fault levels in the middle part of the AD defect
(a) 220pF (b) 470pF (c) 680pF
Figure. 10 Images of different fault levels in the middle part of the SCV defect
(a) SC-1 (b) SC-2 (c) SC-3
Figure. 11 Images of different fault degrees of the middle part of the SC defect
As shown in Figure. 9 to Figure.10, the differences in the distribution of color blocks in images with different fault types provide critical reference value for analyzing and understanding image content. For the three typical faults, as the severity of the fault increases, both the color depth and area of multiple local regions in the MSRP image undergo corresponding changes. Therefore, by converting the frequency response curve into a two-dimensional image using MSRP, more direct and visual fault information is provided, enabling clearer identification of the patterns between different fault types and severity levels of winding faults, thereby significantly enriching the characteristics of winding faults.
- EXPERIMENTAL RESULTS AND DISCUSSION
4.1 Evaluation criteria
In this study, accuracy, precision, recall, and the comprehensive evaluation metric F1-score were used as evaluation metrics for the model, as follows:
where, TP denotes true positive cases, TN denotes true negative cases, FP denotes false positive cases, and FN denotes false negative cases. Accuracy measures the proportion of samples correctly predicted by the classifier, i.e., the ratio of true positives to true negatives among all samples. Precision refers to the proportion of samples actually classified as winding faults among all samples classified as winding faults.
4.2 Network training and fault identification results
In this paper, we propose a winding fault identification method based on MSRP feature enhancement and EMA-MobileNetV3 for classifying winding fault frequency response feature maps. The winding fault identification experiments used a computer equipped with an AMD Ryzen 9 7945HX with Radeon Graphics, GeForce RTX4060 GPU, and 16 GB RAM running a 64-bit Windows operating system. The software environment was Python 3.11 using the PyTorch 2.0.0 library and CUDA 11.8.
In the experimental dataset, there are 360 fault samples and 40 healthy samples, totaling 400 samples. The ratio of the training set to the test set is set to 8:2. The entire network training process is set to 50 epochs, with 16 samples per batch, an initial learning rate of 0.001, and a cosine annealing strategy is used to adjust the learning rate. During training, the Adam optimizer was used to tune the network parameters, and the focus loss function was employed to calculate the model loss value. The training set data underwent image enhancement preprocessing techniques such as random flipping and normalization to improve quality and reduce overfitting [32].
The confusion matrix is an effective tool for analyzing the recognition rates of various types and degrees of faults. Figure. 12 visualizes the final classification results of the ema-MobileNetV3 model after training using a confusion matrix, including fault type identification and severity assessment, where the vertical and horizontal labels represent the actual winding fault type and the predicted fault type or severity, respectively. When windings experience faults of different types and severity levels, the frequency response curves correspondingly change, which further results in distinct MSRP images. Therefore, the feature maps extracted by the EMA-MobileNetV3 ensure that images from different winding states can be classified into different categories As shown in Figure. 12. The MSRP+EMA-MobileNetV3 classification model proposed in this study has only a small number of label prediction errors in the SCV fault type and degree 1 and degree 2 severity levels, while the recognition accuracy for other fault types and severity levels is 100%. This is because, under different fault locations and types, degree 2 and the two preceding severity levels may exhibit certain similarities, making recognition more challenging. For winding fault type identification and severity assessment, the evaluation parameters of the proposed method all exceed 96%, as shown in Table 5. This demonstrates that the winding diagnosis method proposed in this study exhibits strong performance.
(a) fault type (b) fault extent
Figure. 12 Confusion matrices of MSRP+EMA-MobileNetV3 in fault type and extent recognition
Table 5 The evaluating indicator of MSRP+EMA-MobileNetV3 in fault type and extent recognition
| diagnosis | Accuracy | Precision | Recall | F1 |
| Fault type | 97.50% | 97.57% | 97.50% | 97.46% |
| Fault extent | 96.25% | 97.50% | 97.50% | 97.50% |
4.3 Comparative experiment
1) Comparative analysis between different models
To comprehensively evaluate the performance of the network model proposed in this study for transformer winding fault diagnosis, it is crucial to conduct a comparative analysis among different classification models. Such comparisons not only highlight the superiority of the model proposed in this paper but also provide a reference for selecting the most suitable fault identification method for this application scenario. Specifically, we compare the network model proposed in this paper with other algorithms. Five network models—ShufflenetV2, MobileNetV3, ECA-MobileNetV3, and ResNet34—were selected for comparison with the method proposed in this paper. This is because these four networks encompass a wide range of network structures, from lightweight to deep, and these network models have been thoroughly validated and applied in classification experiments. These four models and the model proposed in this paper were trained on the winding MSRP image dataset, and the training process for fault type identification is shown in Figure. 13. The accuracy of all models increased with the number of iterations and eventually converged. The model proposed in this study converged the fastest, reaching convergence at 30 iterations, with a final training convergence accuracy of 98.125%, significantly higher than other models.
(a) training accuracy (b) training loss
Figure. 13 Different models training accuracy and loss curves
For winding fault type identification, the comparison of evaluation metrics for each model is shown in Table 6. Among these, to assess the model’s inference speed, test time was also included as an evaluation metric. As shown in Table 6. The fault type identification accuracy of EMA-MobileNetV3 is 97.5%, the highest among all models, with Precision, Recall, and F1 scores all closest to 1, indicating strong generalization capabilities of the model. The classification accuracy rates of ShufflenetV2, MobileNetV3, ECA-MobileNetV3, and ResNet34 are 88.75%, 90.00%, 93.75%, and 95.00%, respectively. Compared to the baseline MobileNetV3 and ECA-MobileNetV3 models, the classification accuracy has improved by 7.5% and 3.75%, respectively. For the assessment of winding fault severity, the comparison of test results for each model is shown in Table 7. The proposed model achieves accuracy, precision, recall, and F1 scores of 96.25%, 97.50%, 97.50%, and 97.50%, respectively, which are also higher than those of other deep learning models. Overall, this method can effectively identify winding fault types and severity with high reliability and accuracy. This is primarily due to the proposed model introducing the EMA attention mechanism, which enhances cross-spatial-channel fine-grained feature interaction and mapping, improves the efficiency and capability of multi-scale feature information fusion, and refines the representation of winding fault features.
In terms of testing time, the proposed EMA-MobileNetV3 model significantly outperforms popular recognition methods such as ShufflenetV2 and ResNet34. Compared to MobileNetV3 and ECA-MobileNetV3, the proposed method only slightly increases the inference time. Specifically, the type recognition test time for ECA-ConvNeXt is 1.27 seconds, which is 130 milliseconds and 60 milliseconds longer than MobileNetV3 and ECA-MobileNetV3, respectively. However, regardless of whether it is type or severity recognition, the model can still complete winding fault classification within 1.5 seconds. Therefore, EMA-MobileNetV3 achieves a significant improvement in winding recognition accuracy and other metrics while maintaining fast inference time. The primary reason is that the EMA attention mechanism establishes efficient information flow channels between different spatial dimensions without significantly increasing computational overhead or parameter counts, thereby enhancing model performance.
Table 6 Comparison of test results of each model for fault tpye
| network models | Accuracy | Precision | Recall | F1 score | test time(s) |
| ShufflenetV2 | 88.75% | 88.10% | 89.50% | 88.79% | 2.17 |
| MobileNetV3 | 90.00% | 90.70% | 89.30% | 90.00% | 1.14 |
| ECA-MobileNetV3 | 93.75% | 94.87% | 92.73% | 93.79% | 1.21 |
| ResNet34 | 95.00% | 94.56% | 95.52% | 95.04% | 3.92 |
| proposed | 97.50% | 97.57% | 97.50% | 97.46% | 1.27 |
Table 7 Comparison of test results of each model for fault extent
| network models | Accuracy | Precision | Recall | F1 score | test time(s) |
| ShufflenetV2 | 93.75% | 94.87% | 92.73% | 93.79% | 2.21 |
| MobileNetV3 | 90.00% | 89.50% | 90.80% | 90.14% | 1.18 |
| ECA-MobileNetV3 | 95.00% | 94.56% | 95.52% | 95.04% | 1.24 |
| ResNet34 | 93.75% | 93.21% | 94.42% | 93.81% | 5.86 |
| proposed | 96.25% | 97.50% | 97.50% | 97.50% | 1.33 |
Considering the slight increase in testing time, which resulted in a significant improvement in fault type and severity identification accuracy to 97.50% and 96.25%, respectively, this trade-off is reasonable. The above results confirm the effectiveness of EMA-MobileNetV3 in maintaining fast testing time and high identification accuracy.
2) Comparative analysis between different coding characteristics
To further validate the superiority of the proposed method, particularly the advantages of the MSRP image transformation method in identifying transformer winding fault types and severity, this work conducted a database comparison of different 1-D transformed two-dimensional images: traditional recurrence plot images [28], Hilbert ID images [16], and GAF images [23]. It is worth noting that both RP and GAF images exhibit characteristics of information redundancy and symmetry. The comparison results of various metrics are clearly presented in Table 8. The MSRP transformation method introduced in this work demonstrates significant advantages in identifying transformer winding fault types and severity. Specifically, in fault type classification, the MSRP achieves an accuracy rate of 97.50%, which is 6.25%, 7.50%, and 3.75% higher than RP (91.25%), Hilbert ID (90.00%), and GAF (93.75%), respectively. In fault severity identification, MSRP also achieved the highest performance, reaching 96.25%, which is 3.75% higher than RP and GAF (92.50%). In summary, this work introduces the MSRP image transformation method to effectively reveal the similarity, chaos, and non-stationarity of frequency response curves based on RP, significantly enriching the local feature differences of windings and achieving multi-scale feature enhancement of winding fault responses. This highly aligns with the outstanding multi-scale feature extraction capabilities of EMA-MobileNetV3.
Table 8 Comparison of test results of each Conversion feature for fault type and extent
| Conversion feature | Work content | Accuracy | Precision | Recall | F1 score |
| RP | Type | 91.25% | 91.00% | 91.50% | 91.25% |
| Extext | 92.50% | 92.86% | 92.15% | 92.50% | |
| Hilbert ID | Type | 90.00% | 91.20% | 88.90% | 90.03% |
| Extext | 88.75% | 89.47% | 88.06% | 88.76% | |
| GAF | Type | 93.75% | 92.85% | 94.60% | 93.72% |
| Extext | 92.50% | 93.25% | 91.75% | 92.51% | |
| MSRP
(our) |
Type | 97.50% | 97.57% | 97.50% | 97.46% |
| Extext | 96.25% | 97.50% | 97.50% | 97.50% |
- APPLICATION CASES
An 110kV on-load tap-changing transformer produced by Xi’an Xibian Zhongte Electrical Equipment Co., Ltd. was tested at a certain substation to verify the applicability of the proposed method in actual engineering scenarios, rather than adding a new test set. The winding frequency response curves obtained from the testing are shown in Figure.14. Then, the frequency response curves were converted into images using MSRP to enhance the features, as shown in Figure.15. And the image was input into the pre-trained EMA-MobileNetV3 recognition model. The detection results indicated that there was an axial displacement fault in the upper region of the winding. As shown in Figure. 16. The actual fault condition of the transformer winding was identified, confirming that an axial displacement fault had indeed occurred in the upper region of the winding. The output results matched the actual fault condition, further validating the effectiveness of the method proposed in this paper and its applicability in real-world testing scenarios.
(a) Field Testing (b) Test frequency response curve
Figure. 14 Field testing and result for 110kV on-load voltage regulator transformer
Figure. 15 MSRP feature enhance
Figure. 16 Field fault winding
- CONCLUSION
This work proposes a transformer winding fault diagnosis method based on a combination of multi-scale, asymmetric, and signed recursive two-dimensional image transformation and the EMA-MoblieNetV3 intelligent classification model, aiming to improve the accuracy of winding fault diagnosis. The main conclusions are as follows.
1) This work innovatively introduces a two-dimensional MSRP method to convert winding frequency response data, achieving multi-scale enhancement of winding fault features and providing an effective feature data foundation for subsequent fault diagnosis.
2) A novel intelligent classification method, EMA-MobileNetV3, is proposed to precisely identify winding faults. This method uses an efficient multi-scale attention enhancement (EMA) module to assign importance weights to multi-scale features across different spatial scales. This enhances its ability to extract fault features. EMA-MobileNetV3 improves the accuracy of traditional MobileNetV3 recognition while avoiding significant time delays. Compared to other classification methods, the proposed method achieves test accuracies of 97.50% and 96.25% for recognizing winding faults by type and scale, respectively.
In summary, the method proposed in this paper can accurately and effectively classify transformer winding faults while maintaining good computational efficiency. In the future, this algorithm can be deployed on edge devices to achieve real-time fault diagnosis of windings. It also provides valuable reference for fault detection and identification in other types of electrical equipment.
Acknowledgements
This research was funded by the Electric Power Research Institute of Yunnan Power Grid Co., Ltd.·(Project No. YNKJXM20222300) and Diqing Power Supply Bureau of Yunnan Power Grid Co., Ltd. (Project No. YNKJXM20222327).
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, author-ship, and/or publication of this article.
Data Sharing Agreement
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.
References
[1] L. Zhou, et al., “Detection for Axial Displacement Fault of Winding in Autotransformer Based on Dynamic Frequency Division of FRA and HHO-RF,” in IEEE Sensors Journal, vol. 24, no. 20, pp. 3261932629, Oct. 15, 2024, doi: 10.1109/JSEN.2024.3448267.
[2] A. Kumar, B. R. Bhalja, and G. B. Kumbhar, “Novel Technique for Location Identification and Estimation of Extent of Turn-to-Turn Fault in Transformer Winding,” IEEE Trans. Ind. Electron., vol. 70, no. 7, pp. 7382–7392, Jul. 2023.
[3] F. Ren, Z. Kang, S. Ji and Q. Li, “High-Frequency Ladder Network Synthesis of Transformer Winding for Its Mechanical Condition Assessment,” in IEEE Transactions on Industrial Electronics, vol. 70, no. 6, pp. 6261-6271, June 2023.
[4] S. Yuan, L. Zhou, F. Cai, D. Wang, L. Wang, and L. Huang, “A Lightweight Optimization Study for Large Capacity Dry-Type Onboard Traction Transformer of EMU Based on CFD and RSM,” in IEEE Transactions on Transportation Electrification, vol. 9, no. 2, pp. 2941–2954, June 2023, doi: 10.1109/TTE.2022.3224503.
[5] M. Badawi et al., “Reliable Estimation for Health Index of Transformer Oil Based on Novel Combined Predictive Maintenance Techniques,” IEEE Access, vol. 10, pp. 25954–25972, 2022, doi: 10.1109/ACCESS.2022.3156102.
[6] W. Song et al., “Magnetostrictive Vibration Characteristics of Amorphous Alloy Transformer With Three-Dimensional Wound Core,” in IEEE Access, vol. 12, pp. 43958-43967, 2024, doi: 10.1109/ACCESS.2024.3380456.
[7] X. Ouyang, Y. Zheng, S. Pan, and J. Luo, “Toward a Comprehensive Evaluation on the Online Methods for Monitoring Transformer Turn-to-Turn Faults,” IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, vol. 71, no. 2, 2024.
[8] Saleh A , Yasin K , Abderrahmane B ,et al.A Review of Frequency Response Analysis Methods for Power Transformer Diagnostics[J].Energies, 2016, 9(12):879.DOI:10.3390/en9110879.
[9] IEEE Guide for the Application and Interpretation of Frequency Response Analysis for Oil-Immersed Transformers,” IEEE Std C57.149-2024 (Revision of IEEE Std C57.149-2012), pp. 1–55, 2025, doi: 10.1109/IEEESTD.2025.10907879.
[10] Power Industry Standards of the People’s Republic of China, Response Analysis Method for Winding Deformation of Power Transformers, Standard DL/T 911-2016, Feb. 2016.
[11] D. Wang et al., “Defect Diagnosis for Autotransformer Winding in High-Speed Railway Based on Transfer Function Pole Distribution,” IEEE TRANSACTIONS ON TRANSPORTATION ELECTRIFICATION, vol. 10, no. 3, 2024.
[12] S. M. A. N. Al-Ameri, et al., “Understanding the influence of power transformer faults on the frequency response signature using simulation analysis and statistical indicators,” IEEE Access, vol. 9, pp. 70935–70947, 2021.
[13] Yousof, M. F. M. , Ekanayake, C. , & Saha, T. K. . (2015). Frequency response analysis to investigate deformation of transformer winding. IEEE Transactions on Dielectrics and Electrical Insulation, 22(4), 2359-2367.
[14] Z. Zhao, C. Yao, C. Li, and S. Islam, “Detection of Power Transformer Winding Deformation Using Improved FRA Based on Binary Morphology and Extreme Point Variation,” IEEE Transactions on Industrial Electronics, vol. 65, no. 4, pp. 3509–3519, 2018.
[15] O. Aljohani and A. Abu-Siada, “Application of DIP to Detect Power Transformers Axial Displacement and Disk Space Variation Using FRA Polar Plot Signature,” IEEE Trans. Ind. Inf., vol. 13, no. 4, pp. 1794–1805, Aug. 2017.
[16] J. Duan, Y. He, and X. Wu, “Serial transfer learning (STL) theory for processing data insufficiency: Fault diagnosis of transformer windings,” International Journal of Electrical Power & Energy Systems, vol. 130, p. 106965, Sep. 2021, doi: 10.1016/j.ijepes.2021.106965.
[17] A. J. Ghanizadeh and G. B. Gharehpetian, “ANN and cross-correlation based features for discrimination between electrical and mechanical defects and their localization in transformer winding,” in IEEE Transactions on Dielectrics and Electrical Insulation, vol. 21, no. 5, pp. 2374-2382, Oct. 2014, doi: 10.1109/TDEI.2014.004364.
[18] A. Moradzadeh and K. Pourhossein, “Early Detection of Turn-to-Turn Faults in Power Transformer Winding: An Experimental Study,” in 2019 International Aegean Conference on Electrical Machines and Power Electronics (ACEMP) & 2019 International Conference on Optimization of Electrical and Electronic Equipment (OPTIM), 2019, pp. 199–204. doi: 10.1109/ACEMP-OPTIM44294.2019.9007169.
[19] Liu J , Zhao Z , Tang C ,et al.Classifying Transformer Winding Deformation Fault Types and Degrees using FRA based on Support Vector Machine[J].IEEE Access, 2019, PP(99):1-1.DOI:10.1109/ACCESS.2019.2932497.
[20] L. Zhou, T. Lin, X. Zhou, S. Gao, Z. Wu, and C. Zhang, “Detection of Winding Faults Using Image Features and Binary Tree Support Vector Machine for Autotransformer,” IEEE Trans. Transp. Electrific., vol. 6, no. 2, pp. 625–634, Jun. 2020.
[21] Bigdeli,,Mehdi,Siano,,Pierluigi,Alhelou,,Hassan,& Haes.(2021).Intelligent Classifiers in Distinguishing Transformer Faults Using Frequency Response Analysis.IEEE ACCESS,9,13981-13991.
[22] X. Wu, Y. He, C. Wang, W. Wu, C. Wang, and J. Duan, “Hilbert ID Considering Multi-Window Feature Extraction for Transformer Deep Vision Fault Positioning,” IEEE Access, vol. 8, pp. 91276–91286, 2020, doi: 10.1109/access.2020.2991844.
[23] S. Yang, Z. Li, X. Yang, and H. Wu, “Fault Identification Method of Transformer Winding based on GramianAngular Difference Field and Convolutional Neural Network,” EEENG, vol. 17, no. 8, pp. 837–847, Oct. 2024, doi: 10.2174/0123520965272942231009050206.
[24] J. Duan, Y. He, X. Wu, H. Zhang, and W. Wu, “Anti-Interference Deep Visual Identification Method for Fault Localization of Transformer Using a Winding Model,” Sensors, vol. 19, no. 19, p. 4153, Sep. 2019, doi: 10.3390/s19194153.
[25] J. Duan, Y. He, and X. Wu, “Serial transfer learning (STL) theory for processing data insufficiency: Fault diagnosis of transformer windings,” International Journal of Electrical Power & Energy Systems, vol. 130, p. 106965, Sep. 2021, doi: 10.1016/j.ijepes.2021.106965.
[26] A. Moradzadeh, B. Mohammadi-Ivatloo, and A. P. Aguiar, “Turn-to-Turn Short Circuit Fault Localization in Transformer Winding via Image Processing and Deep Learning Method,” IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, vol. 18, no. 7, 2022.
[27] B. M. Mathunjwa, Y. T. Lin, C. H. Lin, M. F. Abbod, and J. S. Shieh, “ECG arrhythmia classification by using a recurrence plot and convolutional neural network,” Biomedical Signal Processing and Control, vol. 64, 2021, Art no. 102262.
[28] X. Liu, L. Xia, J. Shi, L. Zhang, L. Bai, and S. Wang, “A Fault Diagnosis Method of Rolling Bearing Based on Improved Recurrence Plot and Convolutional Neural Network,” IEEE Sensors Journal, vol. 23, no. 10, pp. 10767-10775, 2023.
[29] Y. Wei, M. Du, T. Li, X. Zheng, and C. Ji, “Feature-fused residual network for time series classification,” Journal of King Saud University – Computer and Information Sciences, vol. 36, no. 10, p. 102227, Dec. 2024, doi: 10.1016/j.jksuci.2024.102227.
[30] A. Howard, M. Sandler, G. Chu, and L. C. Chen, “Searching forMobileNetV3,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2019,pp. 1314–1324.
[31] D. Ouyang et al., “Efficient Multi-Scale Attention Module with Cross-Spatial Learning,” in ICASSP 2023 – 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Rhodes Island, Greece: IEEE, Jun. 2023, pp. 1–5. doi: 10.1109/icassp49357.2023.10096516.
[32] D. Wang et al., “Defect Identification for Auxiliary Power Cable in EMUs Based on TFDR and GAF-DCN,” IEEE Trans. Transp. Electrific., pp. 1–1, 2024, doi: 10.1109/TTE.2024.3441690.