Converter transformer fault diagnosis method based on optimized S synchronous extraction transformation and DFAE
By optimizing the S-synchronous extraction transformation and DFAE model, the problem of insufficient feature learning in the fault diagnosis method of the converter transformer in complex environments is solved, high-precision and robust fault diagnosis is achieved, and the operational reliability of the converter transformer is improved.
Patent Information
- Application Number
- CN202510580491.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-06
- Publication Date
- 2025-09-12
AI Technical Summary
Existing converter transformer fault diagnosis methods have insufficient information feature learning capabilities when faced with complex field environments and noise interference, resulting in low diagnostic accuracy, strong dependence on labeled data, and poor robustness and generalization.
A fault diagnosis method based on optimized S-sync extraction transform and deep fusion autoencoder (DFAE) was adopted. By improving the sync extraction operator and designing the wavelet activation function, a DFAE model was constructed. The model was trained using unlabeled data and fine-tuned with a small amount of labeled data to enhance feature learning ability.
Effectively extracting fault features from the vibration signals of commutation transformers improves the accuracy and robustness of fault diagnosis, enables accurate identification of different fault categories in complex environments, and reduces dependence on labeled data.
Smart Images
Figure CN120632550A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of converter transformer fault diagnosis, and in particular to a converter transformer fault diagnosis method based on optimized S-synchronous extraction transformation and DFAE. Background Art
[0002] As a crucial component of ultra-high voltage direct current (UHVDC) transmission systems, the safe and stable operation of converter transformers (CTs) is crucial for the reliable operation of the entire transmission system. However, due to the long-term harsh operating environments in which CTs operate, coupled with the multiple impacts of high voltage, high current, and heavy mechanical loads, their operation inevitably faces severe technical challenges, making their inspection and maintenance complex and costly. A CT failure can potentially cause severe power outages, posing a significant threat to the normal operation and steady development of society and the economy. Therefore, developing fast, accurate, and cost-effective fault diagnosis methods for CTs is of great practical significance for improving the stability and safety of power systems.
[0003] With the rapid development of intelligent sensing technology, the widespread use of high-precision, large-scale sensors has provided new breakthroughs in fault detection for converter transformers. This more comprehensive and accurate monitoring data has significantly improved the accuracy of fault diagnosis, leading to the emergence of numerous data-driven converter transformer fault diagnosis methods that significantly enhance both the accuracy and efficiency of diagnosis. However, many traditional data-driven converter transformer fault diagnosis methods rely on supervised learning, which means that model training effectiveness and diagnostic accuracy often depend on the availability of labeled data. However, due to the complexity and unique characteristics of converter transformer field operations, collecting sufficient labeled data is challenging, severely limiting the widespread adoption and application of these methods.
[0004] To overcome the limitations of supervised learning-based commutator fault diagnosis methods, several semi-supervised learning-based commutator fault diagnosis methods have emerged in recent years. Semi-supervised methods can leverage both labeled data and large amounts of unlabeled data for model training. However, existing semi-supervised methods face numerous challenges: 1) The complex and harsh working environment of commutator sites, rife with strong noise interference, poses a significant challenge to extracting effective fault information from semi-supervised commutator fault diagnosis methods; 2) Traditional semi-supervised commutator fault diagnosis methods lack the ability to learn information features, resulting in low fault diagnosis accuracy; and 3) Traditional semi-supervised commutator fault diagnosis methods still rely heavily on a small amount of labeled data, resulting in poor model robustness and generalization. Summary of the Invention
[0005] To solve the above problems, the present invention aims to provide a fault diagnosis method for commutation transformers based on optimized S-synchronous extraction transformation and DFAE.
[0006] To achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions:
[0007] The commutation transformer fault diagnosis method based on optimized S-synchronous extraction transformation and DFAE includes the following steps:
[0008] S1. Constructing a converter transformer fault dataset: collecting vibration signals of actual converter transformers under different fault types and dividing them into a training set and a test set;
[0009] S2, design variance parameter correction and adaptive parameters;
[0010] S3. Improved synchronous extraction transformation to highlight fault characteristics: The improved synchronous extraction operator is used to further improve the energy concentration of the time-frequency representation to highlight the fault characteristics;
[0011] S4. Build DFAE model;
[0012] S5. Train the DFAE model: Use the time-frequency graph data constructed by the training set and the test set to complete model training and evaluation;
[0013] S6. Fault diagnosis: The time-frequency diagram constructed by the collected vibration signals of the operating converter transformer is input into the fault diagnosis model. The model will output the predicted fault type to realize the converter transformer fault diagnosis.
[0014] Preferably, step S2 is to calculate the variance parameter according to the energy concentration index to optimize the optimal parameter set of the S transform, and convert the vibration signal into a time-frequency diagram with the best time-frequency representation.
[0015] Preferably, step S4 designs a wavelet activation function based on Morlet wavelet and uses it as the activation function of the convolutional autoencoder, thereby constructing a DFAE model based on the basic network.
[0016] Preferably, step S1 specifically includes the following steps:
[0017] The acceleration sensor is used to collect vibration signals of the actual commutation transformer under different fault types such as normal operation, abnormal winding looseness, and abnormal tap changer looseness. 300 to 500 groups of signals are collected under each fault type. Each group of signals can be expressed as x(n) = (x1, x2, ..., x N ), where x i represents the i-th sampling point of a set of signals x(n), 1≤i≤N, N is the total number of sampling points of a set of signals. The collected signals are divided into training set and test set in a ratio of 7:3. The training set contains a large amount of unlabeled data and a small amount of labeled data, while the test set only contains labeled data, forming the converter transformer fault data set.
[0018] Preferably, step S2 specifically includes the following steps:
[0019] Step S2.1: Correct the variance parameter
[0020] Modify σ to a multivariate function of f so that the change mode and rate of change of the window width can be flexibly adjusted through different parameters. The specific definition is as follows:
[0021]
[0022] Among them, the variance σ is the width of the control window function, α, β, and γ are all parameters that control the window width, defined as the set Ψ = {α, β, γ}; among them, α is used to control the rate of change of the window width, and β and γ define the width of the window, and the optimized S transform is obtained:
[0023]
[0024] Where ω(n-τ, f) is the Gaussian window function; N is the total number of sampling points of a set of signals;
[0025] Step S2.2: Design adaptive parameters
[0026] First, based on simulation experiments, we obtain the empirical value ranges of each parameter: α∈[0.5, 3.5], β∈[5, 12], and γ∈[1, 2.5]. Then, we propose the concept of energy concentration index based on this range to adaptively select the optimal parameter set and obtain the best time-frequency representation of the vibration signal. Specifically:
[0027] The optimized S-transform result S(τ, f) is standardized under each parameter set Ψ, and the calculation process is:
[0028]
[0029] in Indicates S Ψ The normalized result of (τ, f) is then calculated for each The corresponding energy concentration EC is calculated as follows:
[0030]
[0031] The larger the EC value, the higher the time-frequency energy concentration. Therefore, the parameter set corresponding to the maximum EC value is the optimal parameter set Ψ opt ;
[0032] Step S2.3: Obtain the optimal time-frequency representation
[0033] For the original vibration signal, calculate its optimized S transform under the optimal parameter set to obtain the best time-frequency representation result That is, the time-frequency diagram with the highest time-frequency energy concentration is obtained. The calculation process is expressed as:
[0034]
[0035] Preferably, step S3 specifically includes the following steps:
[0036] Step S3.1: Calculation The derivative of :
[0037]
[0038] Among them, f0(τ, f) is the ideal instantaneous frequency trajectory of the signal,
[0039] Step S3.2: Calculate the improved synchronous extraction operator:
[0040]
[0041] Where δ is the Dirichlet function, Re|·| is the real part operation, and Δf is the frequency interval;
[0042] Step S3.3: Use the improved synchronization extraction operator to extract the time-frequency coefficients on the instantaneous frequency ridge, eliminate redundant time-frequency coefficients, and obtain a time-frequency graph with highly concentrated time-frequency energy. The specific calculation process is as follows:
[0043]
[0044] Preferably, step S4 includes the following steps:
[0045] Step 4.1: Build a wavelet convolutional autoencoder
[0046] The real part of the Morlet wavelet is used to construct the wavelet activation function, which is expressed as:
[0047]
[0048] Where t is the time variable of Morlet wavelet;
[0049] After using the wavelet activation function as the activation function of the hidden layer of the convolutional autoencoder, the wavelet convolutional autoencoder is obtained. The output of the i-th neuron in the hidden layer can be expressed as:
[0050]
[0051] Among them, k is the size of the local receptive field, W ij is the weight of the convolution kernel, z j is the input time-frequency graph or the feature of the previous layer output, a i and b i are the scale factor and translation factor of the wavelet activation function, is the wavelet activation function;
[0052] Step 4.2: Build the DFAE model
[0053] The DFAE model is constructed by the following network structure: input layer, first-layer denoising autoencoder, second-layer denoising autoencoder, contraction autoencoder and output layer.
[0054] Preferably, in step S5, when training the model, a training strategy of unsupervised pre-training and supervised fine-tuning is adopted.
[0055] Further preferably, the unsupervised pre-training process is as follows:
[0056] First, with the goal of minimizing the model’s loss function, the model is pre-trained using a large amount of unlabeled data in the training set. The model’s loss function is expressed as:
[0057]
[0058] in, Represents the time-frequency diagram containing noise information, y represents the output reconstructed image, λ represents the shrinkage factor, is the Jacobian matrix The F-norm of .
[0059] Further preferably, the supervised fine-tuning training process is as follows:
[0060] Modify the model's loss function to the following formula and train the model using a small amount of labeled data in the training set:
[0061]
[0062] Where C represents the number of fault types of the converter transformer, represents the one-hot encoding of the true label of the input sample, represents the model's predicted probability for each fault category.
[0063] Compared with the prior art, the present invention has the following advantages:
[0064] The present invention's converter transformer fault diagnosis method based on optimized S-synchronous extraction transformation and DFAE can effectively extract fault feature information from converter transformer vibration signals, enhance the model's feature learning capabilities, and fully utilize unlabeled data to effectively implement converter transformer fault diagnosis. On the one hand, the proposed optimized S-synchronous extraction transformation can convert the original vibration signal into a time-frequency diagram with higher energy concentration, thereby better highlighting the fault characteristics; on the other hand, the proposed DFAE model can more effectively capture the differential characteristics of signals of different fault categories, and combined with the advantages of different autoencoders, greatly improve the accuracy of the converter transformer semi-supervised method. Specifically:
[0065] First, the optimized S synchronous extraction transform can flexibly adjust the window width change mode and rate of change by correcting the variance parameter into a multivariate function of frequency. At the same time, the energy concentration index can measure the time-frequency energy concentration under different parameter sets, so as to adaptively select the optimal parameter set. The improved synchronous extraction operator obtained in this way can further improve the energy concentration of the time-frequency graph and better highlight the fault characteristic information.
[0066] Secondly, by designing a wavelet activation function and using it as the activation function of the hidden layer of the convolutional autoencoder, the time-frequency locality and focal characteristics of the wavelet function are fully utilized to effectively capture the differential characteristics of signals of different fault categories. The DFAE constructed based on this network combines the advantages of the denoising autoencoder and the contraction autoencoder, better realizes the fault information extraction, and improves the accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0067] Figure 1 Schematic diagram of the collected normal / abnormal vibration signals of the commutation transformer;
[0068] Figure 2 The figure is a processing result diagram of the vibration signal after the variance parameter correction and adaptive parameter design in step S2;
[0069] Figure 3 is the result graph after processing in step S3;
[0070] Figure 4 This is the network structure diagram of the DFAE model;
[0071] Figure 5 Figure 2 is the fault diagnosis result diagram of different fault diagnosis models. DETAILED DESCRIPTION
[0072] The purpose of the present invention is to provide a method for diagnosing converter transformer faults based on optimized S-synchronous extraction transformation and DFAE. The present invention will be further described below with reference to specific embodiments.
[0073] The DFAE of the present invention is a deep fusion autoencoder.
[0074] Example 1
[0075] The commutation transformer fault diagnosis method based on optimized S-synchronous extraction transformation and DFAE includes the following steps:
[0076] S1. Constructing a converter transformer fault dataset: collecting vibration signals of actual converter transformers under different fault types and dividing them into a training set and a test set;
[0077] The specific steps include:
[0078] The acceleration sensor is used to collect vibration signals of the actual commutation transformer under different fault types such as normal operation, abnormal winding looseness, and abnormal tap changer looseness. 300 to 500 groups of signals are collected under each fault type. Each group of signals can be expressed as x(n) = (x1, x2, ..., x N ), where x i represents the i-th sampling point of a set of signals x(n), 1≤i≤N, N is the total number of sampling points of a set of signals. The collected signals are divided into training set and test set in a ratio of 7:3. The training set contains a large amount of unlabeled data and a small amount of labeled data, while the test set only contains labeled data, forming the converter transformer fault data set.
[0079] In this embodiment, the acquisition frequency of the sensor is set to 10kHz, 400 sets of signals are collected under each fault type, and the total number of sampling points N of a set of signals is set to 1024; Figure 1 As shown;
[0080] S2. Design variance parameter correction and adaptive parameters: Calculate the variance parameters based on the energy concentration index to optimize the optimal parameter set of the S transform and convert the vibration signal into a time-frequency diagram with the best time-frequency representation;
[0081] The present invention proposes a time-frequency analysis method that optimizes the S synchronous extraction transform to extract features from vibration signals containing converter transformer fault information. The S transform is a time-frequency analysis method that can convert vibration signals into time-frequency diagrams to achieve effective extraction of time-frequency features. However, the width transformation of its window function has local frequency dependence, resulting in a lack of flexibility when processing actual signals. The synchronous extraction transform is a post-processing process of the short-time Fourier transform, which can improve the energy divergence problem of traditional time-frequency analysis methods and improve energy concentration, but it also inherits the shortcomings of the short-time Fourier transform's fixed time-frequency resolution. Accordingly, the present invention corrects the variance parameters of the S transform, proposes an adaptive parameter design method to find the optimal parameter set, and improves the synchronous extraction operator of the synchronous extraction transform, and finally applies it to the optimized S transform to achieve full mining and effective extraction of fault information. This step first completes the variance parameter correction and adaptive parameter design of the S transform, thereby obtaining a time-frequency diagram with the best time-frequency representation. As Figure 2 The specific steps are as follows:
[0082] Step S2.1: Correct the variance parameter
[0083] In the S transform, the variance σ is used to control the width of the window function and is defined as f represents frequency. This design solves the problem of fixed time-frequency resolution of the short-time Fourier transform, but still has the disadvantage of local frequency dependence. To address this problem, the present invention modifies σ to a multivariate function of f, so that the change mode and rate of change of the window width can be flexibly adjusted by different parameters. The specific definitions are as follows:
[0084]
[0085] Among them, α, β, and γ are parameters that control the window width, which are defined as the set Ψ = {α, β, γ}; where α is used to control the rate of change of the window width, and β and γ define the width of the window, and the optimized S transform is obtained:
[0086]
[0087] Where ω(n-τ, f) is the Gaussian window function; N is the total number of sampling points of a set of signals;
[0088] Step S2.2: Design adaptive parameters
[0089] In order to determine the optimal value of the parameter Ψ, first, the empirical value range of each parameter is obtained based on simulation experiments: α∈[0.5, 3.5], β∈[5, 12], γ∈[1, 2.5]. Then, based on this, the concept of energy concentration index is proposed to adaptively select the optimal parameter set and thus obtain the best time-frequency representation of the vibration signal. Specifically:
[0090] The optimized S-transform result S(τ, f) is standardized under each parameter set Ψ, and the calculation process is:
[0091]
[0092] in Indicates S Ψ The normalized result of (τ, f) is then calculated for each The corresponding energy concentration EC is calculated as follows:
[0093]
[0094] The larger the EC value, the higher the time-frequency energy concentration. Therefore, the parameter set corresponding to the maximum EC value is the optimal parameter set Ψ opt ;
[0095] Step S2.3: Obtain the optimal time-frequency representation
[0096] For the original vibration signal, calculate its optimized S transform under the optimal parameter set to obtain the best time-frequency representation result That is, the time-frequency diagram with the highest time-frequency energy concentration is obtained. The calculation process is expressed as:
[0097]
[0098] S3. Improved synchronous extraction transformation to highlight fault characteristics: The improved synchronous extraction operator is used to further improve the energy concentration of the time-frequency representation to highlight the fault characteristics; based on the time-frequency diagram obtained by optimizing the S transform, the synchronous extraction operator of the synchronous extraction transformation is improved to further improve the energy concentration of the time-frequency representation and enhance the time-frequency readability of the time-frequency diagram. Figure 3 The specific steps are as follows:
[0099] Step S3.1: Calculation The derivative of :
[0100]
[0101] Where f0(τ, f) is the ideal instantaneous frequency trajectory of the signal;
[0102] Step S3.2: Calculate the improved synchronous extraction operator:
[0103]
[0104] Where δ is the Dirichlet function, Re|·| is the real part operation, and Δf is the frequency interval;
[0105] Since the improved synchronization extraction operator is obtained based on the optimal parameter set calculated by the adaptive parameter design method, it can better extract and characterize the instantaneous frequency of the original signal and improve the energy concentration.
[0106] Step S3.3: Use the improved synchronization extraction operator to extract the time-frequency coefficients on the instantaneous frequency ridge, eliminate redundant time-frequency coefficients, and obtain a time-frequency graph with highly concentrated time-frequency energy. The specific calculation process is as follows:
[0107]
[0108] The above calculations convert the original vibration signal into a time-frequency diagram with higher energy concentration, thereby better highlighting the fault characteristics and improving the accuracy of fault diagnosis. The resulting time-frequency diagram is then used as input for subsequent fault diagnosis model training and testing.
[0109] S4. Construct DFAE model: Design a wavelet activation function based on Morlet wavelet and use it as the activation function of the convolutional autoencoder to construct the DFAE model based on this basic network;
[0110] This embodiment proposes a DFAE model as a fault diagnosis model for converters. Convolutional autoencoder is an unsupervised network structure specifically for processing image data. The core idea is to learn the effective feature representation of input data through encoding and decoding operations on the input data, while removing useless information. However, the activation function of the hidden layer of the traditional convolutional autoencoder usually adopts the ReLU function, which often faces the problem of "dead neurons" and is difficult to establish an accurate mapping relationship between various fault types and the original input data. To address this problem, this embodiment designs a wavelet activation function based on Morlet wavelet, and uses it as the activation function of the hidden layer of the convolutional autoencoder, and then constructs DFAE, so that the difference signal characteristics can be described with different resolutions, thereby improving the model fault recognition accuracy.
[0111] Step 4.1: Build a wavelet convolutional autoencoder
[0112] The Morlet wavelet is expressed as a complex exponential function in the time domain and has the shape of a Gaussian window in the frequency domain. In actual use, complex wavelets are prone to phase distortion when restoring signals. Therefore, this embodiment uses the real part of the Morlet wavelet to construct a wavelet activation function, which is expressed as:
[0113]
[0114] Where t is the time variable of Morlet wavelet;
[0115] After using the wavelet activation function as the activation function of the hidden layer of the convolutional autoencoder, the wavelet convolutional autoencoder is obtained. The output of the i-th neuron in the hidden layer can be expressed as:
[0116]
[0117] Among them, k is the size of the local receptive field, W ij is the weight of the convolution kernel, z j is the input time-frequency graph or the feature of the previous layer output, a i and b i are the scale factor and translation factor of the wavelet activation function, is the wavelet activation function;
[0118] Step 4.2: Build the DFAE model
[0119] Denoising autoencoders, based on autoencoders, introduce a corruption process at the input layer to remove noise from the input data. Contracting autoencoders, on the other hand, enhance the model's ability to learn internal information features by adding a compression term to the loss function. This paper proposes a DFAE model, using a wavelet convolutional autoencoder as the model's foundational network and combining the advantages of both the denoising and contracting autoencoders.
[0120] like Figure 4 As shown, the DFAE model constructed in this embodiment consists of the following network structure: input layer, first layer denoising autoencoder, second layer denoising autoencoder, contraction autoencoder and output layer;
[0121] S5. Training the DFAE model: The time-frequency graph data constructed by the training set and the test set are used to complete the model training and evaluation. When training the model, the unsupervised pre-training and supervised fine-tuning training strategies are adopted, making this method a semi-supervised fault diagnosis model. It can make full use of unlabeled data for model training and extract fault features from the input and output of the data respectively to improve the accuracy of fault diagnosis.
[0122] The unsupervised pre-training process is as follows:
[0123] First, with the goal of minimizing the model’s loss function, the model is pre-trained using a large amount of unlabeled data in the training set. The model’s loss function is expressed as:
[0124]
[0125] in, Represents the time-frequency diagram containing noise information, y represents the output reconstructed image, λ represents the shrinkage factor, is the Jacobian matrix The F-norm of .
[0126] After the model is pre-trained, a Softmax operation is added after the output layer of the model, and the model loss function is modified. The model is fine-tuned using a small amount of labeled data in the training set, and the accuracy of the model is evaluated using the data in the test set. The training process of the supervised fine-tuning is as follows:
[0127] Modify the model's loss function to the following formula and train the model using a small amount of labeled data in the training set:
[0128]
[0129] Where C represents the number of fault types of the converter transformer, represents the one-hot encoding of the true label of the input sample, represents the model's predicted probability for each fault category.
[0130] S6. Fault diagnosis: After the DFAE model training is completed, the time-frequency diagram constructed by the collected vibration signals of the operating converter transformer is input into the fault diagnosis model. The model will output the predicted fault type to realize the fault diagnosis of the converter transformer, such as Figure 5 As shown, it can be seen that the method proposed in the present invention is significantly superior to other methods in terms of fault diagnosis accuracy. AE refers to autoencoder, DAE (ReLU) refers to denoising autoencoder with ReLU as activation function, CAE (ReLU) refers to contraction autoencoder with ReLU as activation function, DAE (Morlet) refers to denoising autoencoder with Morlet wavelet as activation function, and CAE (Morlet) refers to contraction autoencoder with Morlet wavelet as activation function. This shows that the method proposed in the present invention can effectively solve the fault diagnosis problem of converter transformers and provide reliable technical guarantee for the safe operation of converter transformers.
Claims
1. A commutation transformer fault diagnosis method based on optimized S-synchronous extraction transformation and DFAE is characterized by: The following steps are involved: S1. Constructing a converter transformer fault dataset: collecting vibration signals of actual converter transformers under different fault types and dividing them into a training set and a test set; S2, design variance parameter correction and adaptive parameters; S3. Improved synchronous extraction transformation to highlight fault characteristics: The improved synchronous extraction operator is used to further improve the energy concentration of the time-frequency representation to highlight the fault characteristics; S4. Build DFAE model; S5. Train the DFAE model: Use the time-frequency graph data constructed by the training set and the test set to complete model training and evaluation; S6. Fault diagnosis: The time-frequency diagram constructed by the collected vibration signals of the operating converter transformer is input into the fault diagnosis model. The model will output the predicted fault type to realize the converter transformer fault diagnosis.
2. The method for commutation transformer fault diagnosis based on optimized S-synchronous extraction transformation and DFAE according to claim 1 is characterized in that: Step S2 is to calculate the variance parameter based on the energy concentration index to optimize the optimal parameter set of the S transform, and convert the vibration signal into a time-frequency diagram with the best time-frequency representation.
3. The converter transformer fault diagnosis method based on optimized S-synchronous extraction transformation and DFAE according to claim 1 is characterized by: In step S4, a wavelet activation function is designed based on Morlet wavelet, and used as the activation function of the convolutional autoencoder, and the DFAE model is constructed based on this basic network.
4. The method for commutation transformer fault diagnosis based on optimized S-synchronous extraction transformation and DFAE according to claim 1 is characterized in that: Step S1 specifically includes the following steps: The acceleration sensor is used to collect vibration signals of the actual commutation transformer under different fault types such as normal operation, abnormal winding looseness, and abnormal tap changer looseness. 300 to 500 groups of signals are collected under each fault type. Each group of signals can be expressed as x(n) = (x1, x2, ..., x N ), where x i represents the i-th sampling point of a set of signals x(n), 1≤i≤N, N is the total number of sampling points of a set of signals. The collected signals are divided into training set and test set in a ratio of 7:
3. The training set contains a large amount of unlabeled data and a small amount of labeled data, while the test set only contains labeled data, forming the converter transformer fault data set.
5. The converter transformer fault diagnosis method based on optimized S-synchronous extraction transformation and DFAE according to claim 1 is characterized in that: Step S2 specifically includes the following steps: Step S2.1: Correct the variance parameter Modify σ to a multivariate function of f so that the change mode and rate of change of the window width can be flexibly adjusted through different parameters. The specific definition is as follows: Among them, the variance σ is the width of the control window function, α, β, and γ are all parameters that control the window width, defined as the set Ψ = {α, β, γ}; among them, α is used to control the rate of change of the window width, and β and γ define the width of the window, and the optimized S transform is obtained: Where ω(n-τ,f) is the Gaussian window function; N is the total number of sampling points of a set of signals; Step S2.2: Design adaptive parameters First, based on simulation experiments, we obtain the empirical value ranges of each parameter: α∈[0.5,3.5], β∈[5,12], γ∈[1,2.5]. Then, based on this, we propose the concept of energy concentration index to adaptively select the optimal parameter set and thus obtain the best time-frequency representation of the vibration signal. Specifically: The optimized S-transform result S(τ,f) is standardized under each parameter set Ψ, and the calculation process is: Among them S Ψ (τ,f) represents S Ψ The normalized result of (τ,f) is then calculated for each S Ψ The energy concentration EC corresponding to (τ,f) is calculated as follows: The larger the EC value, the higher the time-frequency energy concentration. Therefore, the parameter set corresponding to the maximum EC value is the optimal parameter set Ψ opt ; Step S2.3: Obtain the optimal time-frequency representation For the original vibration signal, calculate its optimized S transform under the optimal parameter set to obtain the best time-frequency representation result That is, the time-frequency diagram with the highest time-frequency energy concentration is obtained. The calculation process is expressed as:
6. The method for commutation transformer fault diagnosis based on optimized S-synchronous extraction transformation and DFAE according to claim 1 is characterized in that: Step S3 specifically includes the following steps: Step S3.1: Calculation The derivative of : Where f0(τ,f) is the ideal instantaneous frequency trajectory of the signal, Step S3.2: Calculate the improved synchronous extraction operator: Where δ is the Dirichlet function, Re|·| is the real part operation, and Δf is the frequency interval; Step S3.3: Use the improved synchronization extraction operator to extract the time-frequency coefficients on the instantaneous frequency ridge, eliminate redundant time-frequency coefficients, and obtain a time-frequency graph with highly concentrated time-frequency energy. The specific calculation process is as follows:
7. The method for commutation transformer fault diagnosis based on optimized S-synchronous extraction transformation and DFAE according to claim 1 is characterized in that: Step S4 includes the following steps: Step 4.1: Build a wavelet convolutional autoencoder The real part of the Morlet wavelet is used to construct the wavelet activation function, which is expressed as: Where t is the time variable of Morlet wavelet; After using the wavelet activation function as the activation function of the hidden layer of the convolutional autoencoder, the wavelet convolutional autoencoder is obtained. The output of the i-th neuron in the hidden layer can be expressed as: Among them, k is the size of the local receptive field, W ij is the weight of the convolution kernel, z j is the input time-frequency graph or the feature of the previous layer output, a i and b i are the scale factor and translation factor of the wavelet activation function, is the wavelet activation function; Step 4.2: Build the DFAE model The DFAE model is constructed by the following network structure: input layer, first-layer denoising autoencoder, second-layer denoising autoencoder, contraction autoencoder and output layer.
8. The method for commutation transformer fault diagnosis based on optimized S-synchronous extraction transformation and DFAE according to claim 1 is characterized in that: In step S5, when training the model, a training strategy of unsupervised pre-training and supervised fine-tuning is adopted.
9. The method for diagnosing converter transformer faults based on optimized S-synchronous extraction transformation and DFAE according to claim 8 is characterized in that: The unsupervised pre-training process is as follows: First, with the goal of minimizing the model’s loss function, the model is pre-trained using a large amount of unlabeled data in the training set. The model’s loss function is expressed as: in, represents the time-frequency diagram containing noise information, y represents the output reconstructed image, λ represents the shrinkage factor, is the Jacobian matrix The F-norm of .
10. The method for commutation transformer fault diagnosis based on optimized S-synchronous extraction transformation and DFAE according to claim 8, characterized in that: The training process of the supervised fine-tuning is as follows: Modify the model's loss function to the following formula and train the model using a small amount of labeled data in the training set: Where C represents the number of fault types of the converter transformer, represents the one-hot encoding of the true label of the input sample, represents the model's predicted probability for each fault category.