Transformer fault diagnosis method based on Fourier convolution double-current denoising network

Abstract

The invention relates to a transformer fault diagnosis method, and discloses a transformer fault diagnosis method based on a Fourier convolution double-current denoising network, and the method comprises the steps: 1, collecting transformer vibration data to form a data set, carrying out the normalization processing, and dividing the data set into a training set, a test set and a verification set; 2, performing wavelet decomposition on the training set, and reconstructing the training set into a high-frequency signal and a low-frequency signal; a third step of inputting the high-frequency signal with noise into a two-layer Fourier convolution module to extract key frequency domain features, and completing adaptive denoising in combination with residual connection and an SE attention mechanism; 4, for the reconstructed low-frequency signals, extracting time domain features of the signals through two-layer convolution in combination with batch normalization and an activation function; 5, carrying out feature fusion on the extracted time domain and frequency domain, and carrying out global average pooling; and 6, performing hyper-parameter optimization by using the verification set, and inputting the test set into the trained model to complete fault diagnosis. The method can effectively solve the problems of misjudgment, missed judgment and the like of a diagnosis model due to the fact that the transformer is in a high-noise environment for a long time and fault features are easy to submerge.

Description

[0001] This invention relates to the field of transformer fault diagnosis research, specifically to a transformer fault diagnosis technology based on a Fourier convolution dual-stream denoising network. Background Technology

[0002] As a crucial piece of equipment in the power system, the stability of transformer performance directly affects the safety and reliability of the entire power grid. Because transformers often bear high loads and operate in complex environments, any potential fault or anomaly can lead to equipment shutdown, causing power supply interruptions and even affecting the power system in a wider area. Therefore, timely and accurate detection of transformer health status, especially early fault diagnosis, is of great significance for power system maintenance. To improve transformer service life and reduce the risk of sudden outages, it is essential to establish an accurate and reliable transformer fault diagnosis model.

[0003] With the rapid development of modern industry, traditional machine learning methods, such as expert systems, artificial neural networks (ANNs), wavelet transforms, and random forests (RF), have been widely applied to fault diagnosis of transformer enclosure vibration signals. While these methods have achieved certain results, they still rely excessively on expert experience and manual feature extraction, leading to high labor costs and weak generalization ability. To address these issues, deep learning methods, such as convolutional neural networks (CNNs), generative adversarial networks (GANs), and deep residual networks (ResNets), have been widely used in transformer fault diagnosis. Although these methods can automatically extract features, in practical applications, transformers are often located in outdoor environments with significant noise, making weak fault signals easily masked by noise, thus making fault feature extraction more difficult. This is especially true under complex mechanical coupling conditions, where existing methods struggle to cope effectively, leading to misjudgments and missed diagnoses in the diagnostic model, affecting the accuracy and precision of fault diagnosis.

[0004] To address the aforementioned problems, this invention proposes a transformer fault diagnosis method based on a Fourier convolutional dual-stream denoising network. This method combines traditional signal processing techniques with deep learning, enhancing the feature extraction capability. Furthermore, through layer-by-layer denoising, it effectively solves the problem of misjudgment and missed judgment of fault samples caused by the difficulty in extracting fault features due to noise interference. Summary of the Invention

[0005] The purpose of this invention is to propose a transformer fault diagnosis method based on a Fourier convolution dual-stream denoising network, which effectively solves the problem of fault sample misjudgment and omission caused by the difficulty in extracting fault features due to noise interference, and improves the accuracy of the diagnostic model.

[0006] This invention employs the following technical solution: a transformer fault diagnosis method based on a Fourier convolution dual-stream denoising network, comprising the following steps:

[0007] The first step is to collect transformer vibration data to form a dataset, perform normalization processing, and divide it into training set, test set, and validation set;

[0008] The second step is to perform wavelet decomposition on the training set to reconstruct it into high-frequency and low-frequency signals.

[0009] The third step involves inputting the noisy high-frequency signal into a two-layer Fourier convolution module to extract key frequency domain features, and then combining residual connections and SE attention mechanism to complete adaptive denoising.

[0010] The fourth step is to extract the temporal features of the reconstructed low-frequency signal by combining two layers of convolution with batch normalization and activation functions.

[0011] The fifth step is to fuse the extracted time-domain and frequency-domain features and perform global average pooling.

[0012] The sixth step is to optimize the hyperparameters using the validation set and input the test set into the trained model to complete the fault diagnosis.

[0013] Furthermore, the specific steps of the first step are as follows:

[0014] Step 1: Collect transformer vibration data to form a dataset;

[0015] Specifically, label sample data for four fault types—loose transformer windings, loose iron core, deformed windings, and deformed iron core—as well as their normal operating conditions were collected.

[0016] Step 2: Normalize the collected sample data;

[0017] Step 3: Divide the normalized fault dataset into a training set, a test set, and a validation set;

[0018] Furthermore, the specific steps of the second step are as follows:

[0019] Step 1: Perform wavelet transform on the training data and decompose it into three levels of components with different frequencies;

[0020] Step 2: Reconstruct the first two levels of components into high-frequency signals and the last level of components into low-frequency signals, dividing them into two branches: high-frequency and low-frequency.

[0021] Furthermore, the specific steps of the third step are as follows:

[0022] Step 1: Input the reconstructed high-frequency signal into the Fourier convolution module to extract key frequency domain features, and perform identity mapping on some frequency domain features through residual connections to preserve the original information and enhance feature representation ability.

[0023] Specifically, first, the high-frequency signal is input into a 3×3 convolutional layer, and features are extracted using batch normalization and the ReLU activation function. Second, the extracted time-domain features are converted into frequency-domain features through Fourier transform, and multiplication is performed in the frequency domain to complete frequency-domain convolution. Then, inverse Fourier transform is applied to convert the frequency-domain features back to the time domain. Finally, a 1×1 convolutional layer is used to further process and fuse the frequency-domain information.

[0024] Step 2: Input the features with frequency domain information into the Fourier convolution module again for further extraction to enhance the frequency domain representation of fault features;

[0025] Step 3: Fuse the original features after identity mapping with the output features;

[0026] Step 4: Use the SE attention mechanism to obtain the weight of each channel;

[0027] Specifically, the output features containing frequency information are processed by global average pooling and then input into two fully connected layers. The layers are then normalized using the Sigmoid activation function to obtain the weight value for each channel, which is then used to weight and adjust the output feature map.

[0028] Step 5: Apply the obtained weights to the soft thresholding function to complete adaptive soft thresholding denoising;

[0029] Furthermore, the specific steps of the fourth step are as follows:

[0030] Step 1: Input the reconstructed low-frequency signal into a convolutional layer with a kernel size of 3×3, and extract the temporal features of the low-frequency signal by combining batch normalization and ReLU activation function;

[0031] Step 2: Input the extracted features back into a 3×3 convolutional layer to further mine deeper feature representations;

[0032] Furthermore, the specific steps of the fifth step are as follows:

[0033] Step 1: By adding elements one by one, high-frequency features containing frequency domain information and low-frequency features containing time domain information are fused.

[0034] Step 2: Use global average pooling to process the fused features to obtain a global average representation;

[0035] Furthermore, the specific steps of the sixth step are as follows:

[0036] Step 1: Fine-tune and optimize the model hyperparameters using the validation set data, repeat the iteration, and save the model;

[0037] Step 2: Input the test set data into the optimal classifier to complete the fault diagnosis. Attached Figure Description

[0038] Figure 1 This is an overall flowchart of the method involved in this invention;

[0039] Figure 2 This is the flowchart of the Fourier convolution module;

[0040] Figure 3 This is a confusion matrix diagram of the classification results in the embodiment;

[0041] Figure 4 This is a classification visualization diagram of the training process in the embodiment;

[0042] Figure 5 This is a radar chart for noise analysis in the embodiment. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The specific embodiments described herein are merely illustrative and are not intended to limit the invention.

[0044] The first step is to collect transformer vibration data to form a dataset, perform normalization processing, and divide it into training, testing, and validation sets. The specific steps are as follows:

[0045] S101: Collect transformer vibration data to form a dataset;

[0046] Specifically, a total of 5,000 samples of transformer enclosure vibration data were collected, including 1,000 samples of transformer winding loosening, 1,000 samples of iron core loosening, 1,000 samples of winding deformation, 1,000 samples of iron core deformation, and 1,000 samples of normal operation status. The time step of each sample group was 1,024.

[0047] S102: The collected sample data is normalized by compressing the original data to the range [0,1] to eliminate numerical scale differences, thereby improving the stability and convergence speed of the model. The specific formula is as follows:

[0048]

[0049] In the formula: x represents the normalized data, x represents the original data, min(x) represents the minimum value in the original signal, and max(x) represents the maximum value in the original signal;

[0050] S103: Divide the normalized fault dataset into a training set, a test set, and a validation set;

[0051] Specifically, the dataset is divided in a 7:2:1 ratio, with 70% for training, 20% for testing, and 10% for validation, to ensure the model's generalization ability.

[0052] The second step is to perform wavelet decomposition on the training set to reconstruct it into high-frequency and low-frequency signals. The specific steps are as follows:

[0053] S201: As Figure 1 As shown, wavelet transform is performed on the training data, and it is decomposed into three levels of components with different frequencies. The specific formula is as follows:

[0054]

[0055] In the formula: x(t) is the original signal, cA n (k) represents the low-frequency approximation coefficients of the nth-level wavelet decomposition, reflecting the overall trend of the signal, cD n (k) represents the high-frequency detail coefficients of the nth-level wavelet decomposition, reflecting the high-frequency components in the signal, such as abrupt changes and noise. n (k) represents a low-pass filter used to extract low-frequency information from the signal, h n (k) represents a high-pass filter used to extract high-frequency information from the signal, and m and k represent time step indices. Therefore, the formula for the three-level decomposition is as follows:

[0056] x(t) = cA1 + cD1

[0057] cA1=cA2+cD2

[0058] cA2=cA3+cD3

[0059] The first-level decomposition is cA1 and cD1, the second-level decomposition is cA2 and cD2, and the third-level decomposition is cA3 and cD3.

[0060] S202: Reconstruct the first two stages of components into high-frequency signals and the last stage into low-frequency signals, dividing them into high-frequency and low-frequency branches. The high-frequency branch signal is x. hight (t), the low-frequency branch signal is x low (t);

[0061] The third step involves inputting the noisy high-frequency signal into a two-layer Fourier convolutional module to extract key frequency domain features, and then combining residual connections and SE attention mechanisms to complete adaptive denoising. The specific steps are as follows:

[0062] S301: As Figure 2As shown, the reconstructed high-frequency signal is input into the Fourier convolution module to extract key frequency domain features, and some frequency domain features are subjected to identity mapping through residual connections to preserve the original information and enhance the feature expression capability.

[0063] Specifically, firstly, the high-frequency signal is input into a convolutional layer with a kernel size of 3×3, capturing local details using a smaller receptive field. This operation can extract short-term dependencies and enhance the local pattern representation of the signal. Simultaneously, batch normalization (BN) is used to eliminate internal covariate bias, accelerate model convergence, and improve stability. Furthermore, the ReLU activation function is used to introduce a non-linear transformation, avoiding the gradient vanishing problem and enhancing the expressive power of important features, resulting in a feature map X. conv The specific formula is as follows:

[0064] X conv =ReLU(BN(W*x) hight (t)+b))

[0065] In the formula: W represents the weight of the 3×3 convolution kernel, * represents the convolution operation, and b represents the bias term;

[0066] Secondly, the extracted time-domain features are mapped to the frequency domain using Fourier Transform (FFT). Fourier Transform can transform the original signal from a time dimension to a frequency dimension, analyzing different frequency components of the signal. This allows the model to focus on periodic features and the changing patterns of different frequency components. This process enhances the ability to perceive the spectrum of fault signals. The specific formula is as follows:

[0067] F(X conv )=F(X conv )

[0068] In the formula: F(X) conv Let F(.) denote the frequency domain signal, and F(.) denote the Fourier transform. The specific formula for the Fourier transform is shown below:

[0069]

[0070] In the formula: u represents the frequency index, N represents the signal length, j represents the imaginary unit, j 2 =-1;

[0071] Next, a product operation is performed in the frequency domain, which is called frequency domain convolution. Unlike traditional convolution, frequency domain convolution is equivalent to a weighted operation on the signal spectrum. It can efficiently extract global features and enhance the relationship between different frequency components. It can avoid redundant calculations in time domain convolution, thereby improving computational efficiency. The specific formula is shown below:

[0072] Fconv (u)=F(X conv )·W f (u)

[0073] In the formula: F conv (u) represents the frequency domain feature map, W f (u) represents the convolution kernel (learnable parameter) in the frequency domain, and · represents element-wise multiplication;

[0074] Then, an inverse Fourier transform (IFFT) is performed to convert the features processed by frequency domain convolution back to the time domain signal. This allows the features modulated by frequency information to be recovered in the time domain, enabling the model to combine the feature representation capabilities of the frequency and time domains and enhance the fault feature extraction capability. The specific formula is shown below:

[0075] X IFFT =F -1 (F conv )

[0076] In the formula: X IFFT F represents the signal after the inverse Fourier transform. -1 (.) denotes the inverse Fourier transform, where the formula for the inverse Fourier transform is shown below:

[0077]

[0078] Finally, a 1×1 convolutional layer is used to further process the features converted back to the time domain to obtain the feature map P;

[0079] S302: The features with frequency domain information are input again into the Fourier convolution module for further extraction, enhancing the frequency domain representation of the fault features, and obtaining feature map P1;

[0080] S303: Merge the original feature P after the identity mapping with the output feature P1 to obtain the fused feature map P2, as shown in the following formula:

[0081] P2 = P + P1

[0082] S304: Use the SE attention mechanism to obtain the weight of each channel;

[0083] Specifically, firstly, Global Average Pooling (GAP) is performed on the output features containing frequency information, compressing the global features of each channel into a scalar representation G. NThen, the pooled features undergo feature transformation through two layers of fully connected (FC) networks, using non-linear activation functions (such as ReLU) to enhance feature representation. Subsequently, the output is normalized using the sigmoid activation function to generate attention weights for each channel. Finally, these weights are redistributed to the corresponding feature channels to enhance important features and suppress irrelevant features, thereby improving the model's feature extraction capability and robustness. The specific formula is shown below:

[0084] A N =Sigmoid(W1ReLU(W2G) N ))

[0085] In the formula: A N G represents the attention coefficient of the Nth channel, W1 and W2 represent the weights generated by the two fully connected layers, and G... N This represents the global average of the absolute values ​​of the features in each channel;

[0086] S305: By applying the obtained weights to the soft thresholding function, adaptive soft thresholding denoising is completed, resulting in the denoised feature P3. The specific formula is shown below:

[0087] λ N =G N A N

[0088] P3 = sign(P2)max(|P2|-λ) N ,0)

[0089] In the formula: λ N P3 represents the adaptive soft threshold, and P3 represents the feature map after adaptive denoising.

[0090] The fourth step involves extracting the temporal features of the reconstructed low-frequency signal using a combination of two convolutional layers, batch normalization, and activation functions. The specific steps are as follows:

[0091] S401: The reconstructed low-frequency signal is input into a convolutional layer with a kernel size of 3×3. Combined with batch normalization and ReLU activation function, the temporal features of the low-frequency signal are extracted.

[0092] S402: The extracted features are input again into a 3×3 convolutional layer to further mine deeper feature representations and obtain low-frequency features P4;

[0093] The fifth step involves fusing the extracted time-domain and frequency-domain features and performing global average pooling. The specific steps are as follows:

[0094] S501: By adding elements one by one, high-frequency features containing frequency domain information and low-frequency features containing time domain information are fused to obtain a fused feature map P. 34 The specific formula is as follows:

[0095] P 34 =P3+P4

[0096] S502: Use global average pooling to process the fused features to obtain a global average representation;

[0097] The sixth step is to optimize hyperparameters using the validation set and input the test set into the trained model to complete fault diagnosis. The specific steps are as follows:

[0098] S601: Use validation set data to fine-tune and optimize the model hyperparameters, repeat the iteration, and save the model;

[0099] S602: Input the test set data into the optimal classifier to complete the fault diagnosis.

[0100] Example:

[0101] The experimental model of this invention runs on Python 3.11, Windows 11, Intel(R) Core(TM) i7-13700H, and RTX 4060 GPU.

[0102] A total of 5000 transformer tank vibration data samples were collected, including 1000 samples of transformer winding loosening, 1000 samples of core loosening, 1000 samples of winding deformation, 1000 samples of core deformation, and 1000 samples of normal operation. The time step of each sample group was 1024. The collected sample data were normalized and the dataset was divided into a 7:2:1 ratio, with 70% for training, 20% for testing, and 10% for validation, as detailed in Table 1, to ensure the generalization ability of the model.

[0103] Table 1. Description of Transformer Tank Fault Data Set

[0104]

[0105] To verify the effectiveness of the proposed diagnostic model, this study compares and evaluates it with several existing methods, including one-dimensional convolutional neural networks (1D-CNN), support vector machines (SVM), and residual networks (ResNet). Furthermore, several common semi-supervised and self-supervised learning methods, including Wasserstein GAN with gradient penalty (WGAN-GP), SimCLR, and BYOL, are introduced to further measure the model's adaptability. In terms of performance evaluation, in addition to traditional diagnostic accuracy, this study also introduces metrics such as recall, precision, and F1 score to achieve a comprehensive evaluation of the model. To ensure the reliability of the experimental results, each experiment was repeated 10 times, and the average value was used as the final reference result, with 50 iterations. The specific results of the comparative experiments are shown in Table 2. The experimental results show that the method involved in this invention achieves the best performance in all evaluation metrics, with a diagnostic accuracy of 98.75%, which is 17.98% higher than that of 1D-CNN. The main reason for this significant improvement is that 1D-CNN does not employ a denoising strategy, making it difficult to effectively extract key fault features in high-noise environments, thus affecting diagnostic accuracy. The confusion matrices for various faults are as follows: Figure 3 As shown, the results further validate the superiority of the method, with the diagnostic accuracy of each fault type exceeding 96%, and the identification accuracy of normal operating status reaching 100%.

[0106] Table 2 Comparison results of different models

[0107]

[0108]

[0109] To evaluate the reliability of the model training process, this study uses the T-SNE algorithm to perform visualization analysis on samples from each fault category in the training set. 100 samples are randomly selected from each fault category, and the number of iterations is set to 50. The visualization results are as follows: Figure 4 As shown, after 50 iterations, various fault samples have been clearly separated, indicating that the model can effectively distinguish different fault types, thus further verifying the stability and reliability of the proposed method.

[0110] To verify the robustness of our proposed method in high-noise environments, Gaussian white noise of -4dB, -2dB, 0dB, 2dB, and 4dB was added to the data, and the results were compared with existing methods. The experiment was set to 50 iterations, and 10 sets of experiments were conducted, with the average value taken. The experimental results are as follows: Figure 5As shown, our method exhibits optimal performance across all noise levels. This result demonstrates that our method can effectively overcome the impact of noise interference on fault feature extraction, reduce misjudgments and omissions of fault samples caused by noise, and thus improve the accuracy and robustness of fault diagnosis.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; therefore, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A transformer fault diagnosis method based on a Fourier convolutional dual-stream denoising network, characterized in that, Includes the following steps: The first step is to collect transformer vibration data to form a dataset, perform normalization processing, and divide it into training set, test set, and validation set; The second step is to perform wavelet decomposition on the training set to reconstruct it into high-frequency and low-frequency signals. The third step involves inputting the noisy high-frequency signal into a two-layer Fourier convolution module to extract key frequency domain features, and then combining residual connections and SE attention mechanism to complete adaptive denoising. The fourth step is to extract the temporal features of the reconstructed low-frequency signal by combining two layers of convolution with batch normalization and activation functions. The fifth step is to fuse the extracted time-domain and frequency-domain features and perform global average pooling. The sixth step is to optimize the hyperparameters using the validation set and input the test set into the trained model to complete the fault diagnosis.

2. The transformer fault diagnosis method based on a Fourier convolutional dual-stream denoising network according to claim 1, characterized in that, The specific steps of the first step are as follows: Step 1: Collect transformer vibration data to form a dataset; Specifically, label sample data for four fault types—loose transformer windings, loose iron core, deformed windings, and deformed iron core—as well as their normal operating conditions were collected. Step 2: Normalize the collected sample data; Step 3: Divide the normalized fault dataset into training set, test set and validation set.

3. The transformer fault diagnosis method based on a Fourier convolutional dual-stream denoising network according to claim 1, characterized in that, The specific steps of the second step are as follows: Step 1: Perform wavelet transform on the training data and decompose it into three levels of components with different frequencies; Step 2: Reconstruct the first two components into high-frequency signals and the last component into low-frequency signals, dividing them into high-frequency and low-frequency branches.

4. The transformer fault diagnosis method based on a Fourier convolutional dual-stream denoising network according to claim 1, characterized in that, The specific steps of the third step are as follows: Step 1: Input the reconstructed high-frequency signal into the Fourier convolution module to extract key frequency domain features, and perform identity mapping on some frequency domain features through residual connections to preserve the original information and enhance feature representation ability. Specifically, first, the high-frequency signal is input into a 3×3 convolutional layer, and features are extracted using batch normalization and the ReLU activation function. Second, the extracted time-domain features are converted into frequency-domain features through Fourier transform, and multiplication is performed in the frequency domain to complete frequency-domain convolution. Then, inverse Fourier transform is applied to convert the frequency-domain features back to the time domain. Finally, a 1×1 convolutional layer is used to further process and fuse the frequency-domain information. Step 2: Input the features with frequency domain information into the Fourier convolution module again for further extraction to enhance the frequency domain representation of fault features; Step 3: Fuse the original features after identity mapping with the output features; Step 4: Use the SE attention mechanism to obtain the weight of each channel; Specifically, the output features containing frequency information are processed by global average pooling and then input into two fully connected layers. The layers are then normalized using the Sigmoid activation function to obtain the weight value for each channel, which is then used to weight and adjust the output feature map. Step 5: Apply the obtained weights to the soft threshold function to complete the adaptive soft thresholding denoising.

5. The transformer fault diagnosis method based on a Fourier convolutional dual-stream denoising network according to claim 1, characterized in that, The specific steps of the fourth step are as follows: Step 1: Input the reconstructed low-frequency signal into a convolutional layer with a kernel size of 3×3, and extract the temporal features of the low-frequency signal by combining batch normalization and ReLU activation function; Step 2: Input the extracted features back into a 3×3 convolutional layer to further mine deeper feature representations.

6. The transformer fault diagnosis method based on a Fourier convolutional dual-stream denoising network according to claim 1, characterized in that, The specific steps of the fifth step are as follows: Step 1: By adding elements one by one, high-frequency features containing frequency domain information and low-frequency features containing time domain information are fused. Step 2: Use global average pooling to process the fused features to obtain a global average representation.

7. The transformer fault diagnosis method based on a Fourier convolutional dual-stream denoising network according to claim 1, characterized in that, The specific steps of the sixth step are as follows: Step 1: Fine-tune and optimize the model hyperparameters using the validation set data, repeat the iteration, and save the model; Step 2: Input the test set data into the optimal classifier to complete the fault diagnosis.

Images