A method for transformer vibration monitoring and fault diagnosis

Abstract

The present invention relates to a transformer vibration monitoring and fault diagnosis method, the method steps include: obtaining a transformer box surface vibration signal under the real-time operation state of the transformer; extracting a time domain spectrum and a time-frequency spectrum from the transformer box surface vibration signal; fusing and reducing the extracted time domain spectrum and time-frequency spectrum to obtain a feature matrix; inputting the feature matrix into an auxiliary generative adversarial network anomaly detection model trained based on historical data of transformer box surface vibration signals to discriminate the fault type and obtain the transformer fault classification. Compared with the prior art, the present invention conducts a comprehensive analysis of the vibration signal from the time domain and the frequency domain, which can not only strengthen the connection between the features of different domains, but also has strong robustness.

Description

Technical Field

[0001] The invention relates to the technical field of transformer fault diagnosis, and in particular to a transformer vibration monitoring and fault diagnosis method. Background Art

[0002] As a key link in power transformation, the transformer is an indispensable core equipment in the power system. Its reliable operation is of great significance to ensure the safety of the entire power supply system. As the transformer's operating life in the power system continues to increase, the clamping device in the transformer will also become loose, and the preload force will be further reduced. The looseness of the winding will affect the power transformer's ability to resist short-circuit impacts. At the same time, the deterioration of the winding is long-term and cumulative. In the early stage, the degree of deterioration may have no effect on the electrical performance of the transformer. By using conventional means, it is impossible to effectively detect changes in the mechanical structure of the power transformer, such as loose windings and loose cores. However, the existence of deterioration weakens the transformer's ability to resist short-circuit shocks and poses a threat to the insulation between turns.

[0003] Research has shown that the vibration signal on the surface of the transformer box contains information about the working status of the power transformer. Whether it is a problem with the mechanical structure of the transformer or an electrical problem, it will have a certain impact on the vibration on the surface of the box. The acceleration sensor adsorbed on the surface of the box can easily collect vibration data, and the collection process will not affect the stable operation of the power transformer. It has been widely used in the industrial field. With the rapid improvement of computer computing power in recent years, the data obtained by various sensors has received more and more attention, and data-driven intelligent diagnosis methods have developed rapidly and gradually become a new trend in fault diagnosis. At present, there is no solution to non-invasively monitor the changes in the mechanical structure of the transformer such as loose windings and loose cores based on the vibration signal on the surface of the transformer box and detect the type of transformer fault. Summary of the invention

[0004] The purpose of the present invention is to realize non-invasive monitoring of the mechanical state of the transformer based on the vibration signal on the surface of the transformer box, and to provide a transformer vibration monitoring and fault diagnosis method based on an auxiliary generative adversarial network.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A transformer vibration monitoring and fault diagnosis method, the method steps comprising:

[0007] Obtain the transformer box surface vibration signal under the real-time operation state of the transformer;

[0008] Extract time domain spectrum and time-frequency spectrum of transformer box surface vibration signal;

[0009] The extracted time domain spectrum and time-frequency spectrum are fused and dimensionally reduced to obtain a feature matrix;

[0010] The feature matrix is ​​input into the auxiliary generative adversarial network anomaly detection model trained based on the historical data of the transformer box surface vibration signal to identify the fault type and obtain the transformer fault classification.

[0011] As a preferred technical solution, the time domain spectrum and time-frequency spectrum of the transformer box surface vibration signal are extracted as follows:

[0012] Perform Gram matrix feature extraction on the transformer vibration signal X of each extracted signal channel to obtain the time domain spectrum coefficient S of the transformer vibration signal;

[0013] The transformer vibration signal X is subjected to continuous wavelet decomposition to obtain the time-frequency energy characteristic coefficient C.

[0014] As a preferred technical solution, the specific steps of performing the Gram-like matrix feature extraction and obtaining the time domain spectrum coefficient S of the transformer vibration signal include:

[0015] Normalize the vibration signal X;

[0016] Calculate the normalized vibration sequence The phase angle of each sequence point in ;

[0017] The cosine value of the sum of the two phase angles is added to the inner product operation to calculate the relationship between the two discrete points in the vector. The calculation formula is as follows:

[0018]

[0019] In the formula, x i ,y i Represents two different vectors, φ i ,φ j Represents the calculated angle of the corresponding vector;

[0020] The time domain spectrum coefficient S of the transformer vibration signal is extracted using a Gram-like matrix.

[0021] As a preferred technical solution, the specific steps of performing continuous wavelet decomposition processing on the transformer vibration signal to obtain the time-frequency energy characteristic coefficient C include:

[0022] For the transformer vibration signal X, the signal is decomposed in multiple scales by translating and scaling wavelet functions, and the time-frequency energy characteristic coefficient C of the transformer vibration signal is extracted.

[0023] As a preferred technical solution, the specific steps of fusing and reducing the dimensionality of the extracted time domain spectrum and time-frequency spectrum include:

[0024] The time domain spectrum coefficient S and the time-frequency energy characteristic coefficient C matrix are concatenated to obtain the characteristic matrix A;

[0025] Calculate the covariance matrix G of the feature matrix A C , and get the covariance matrix G C The characteristic roots and corresponding eigenvectors of

[0026] Calculate the covariance matrix G C The cumulative contribution rate of characteristic roots p;

[0027] When the cumulative contribution rate p of the characteristic root reaches the set value, the corresponding first r eigenvectors are selected in turn to form the column-direction projection matrix W of the characteristic matrix A, and the new characteristic matrix R = W·A after the dimension reduction of the characteristic matrix A is calculated according to the column-direction projection matrix W.

[0028] As a preferred technical solution, the generative network of the auxiliary generative adversarial network anomaly detection model includes a fully connected layer, a transposed convolutional layer and a convolutional layer;

[0029] The random noise Z and the label vector are input into the generator, and the input is converted into a multi-dimensional vector through the embedding layer and the linear unit. The multi-dimensional vector is reconstructed into an image, and then the image size is transformed through the transposed convolution network, and finally the convolution layer and the activation function are used to generate samples;

[0030] The transposed convolutional network adopts batch normalization and ReLU activation function, and the output layer is composed of convolution and Tanh activation function; the Adam algorithm is used to update parameters during network training.

[0031] As a preferred technical solution, the discriminator network of the auxiliary generative adversarial network anomaly detection model includes 3 convolutional layers; except for the first layer, each convolutional layer is accompanied by a batch normalization layer and a Leaky ReLU activation function, and the first convolutional layer is only followed by a Leaky ReLU activation function;

[0032] The discriminator input passes through three convolutional layer channels and then fully connected to output a num_labels+1-dimensional vector; where num_labels represents the number of sample categories;

[0033] Take the first element of the output vector and pass it through the sigmoid activation function to output the authenticity probability of the image;

[0034] The last num_labels elements are activated by the softmax function to output the probability of the corresponding category of the image.

[0035] As a preferred technical solution, the training steps of the auxiliary generative adversarial network anomaly detection model include:

[0036] Initialize data and randomly generate relevant parameters of the GAN network model;

[0037] Fixed the parameters of the generated network G, from the real sample X real From the data distribution, m samples {x 1 ,x 2 ,x 3 ,…,x m};

[0038] Select m random vectors as random noise and input them into the generative network G at the same time as the sample label to obtain m generated samples

[0039] The real sample {x 1 ,x 2 ,x 3 ,…,x m} and generate samples Input into the discriminant network D, calculate the discriminant probability, use the cross entropy function as the loss function, and use the gradient ascent method to update the parameters of the discriminant network D;

[0040] Fix the parameters of the discriminant network D and extract m random vectors {z 1 ,z 2 ,z 3 ,…,z m};

[0041] Taking the discriminant error of the generated sample as the loss function, the gradient descent method is used to update the parameters θ of the generated network G g ;

[0042] After the network parameters are updated, the data G(z)=X under the sample label C is obtained. fake ;

[0043] The discriminant network D is based on X real With X fake If X fake With X real If they are not similar, fix the parameters of the discriminant network D and train and adjust the generative network G. If the output score of the discriminant network D is high, fix the parameters of the generative network G and train to optimize the parameters of the discriminant network D.

[0044] In the process of mutual game, Nash equilibrium is reached, and the output resolution of the authenticity judgment network is 0.5.

[0045] As a preferred technical solution, the training objective function of the auxiliary generative adversarial network includes:

[0046] The cross entropy function L for judging the authenticity of samples S :

[0047] Ls =E[logD(S=real,x)]+E[1-logD(s=real,G(z,y l ))]

[0048] The cross entropy function L for determining sample categories C :

[0049] L c =E[logD(C = C i ,x)]+E[logD(C=C i ,G(z,y l ))]

[0050] In the formula, S = real, x represents the real sample input to the discriminant network D, S = real, G (z, y l ) represents the sample generated by the generative network G and input to the discriminative network D, D(C=C i ,x) indicates that the label is C i The probability that the real sample is classified correctly, D(C=C i ,G(z,y l )) indicates that the label generated by generator G is C i The probability that the generated samples are classified correctly;

[0051] By training the discriminant network D, the cross function L S With L C The sum is maximized, and the generated network G is trained to make L C With L S The difference is the biggest.

[0052] As a preferred technical solution, during the fault diagnosis process:

[0053] If the output probability of the discriminant network D is greater than 0.5 or close to 0.5, the data at this time is a known fault type, and the known type of fault is classified by the auxiliary generative adversarial network anomaly detection model;

[0054] If the output probability of the discriminant network D is far less than 0.5, the input is a new type of fault. At this time, the input data is recorded and the operation and maintenance personnel are prompted.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] 1) The present invention constructs a transformer fault diagnosis model to judge the transformer surface vibration signal under different working conditions, and outputs whether the transformer is normal or has a certain type of fault under the current working condition. The present invention first extracts the time domain spectrum and time-frequency spectrum of the collected transformer box surface vibration signal, fuses the extracted time domain spectrum and time-frequency spectrum to obtain the characteristic matrix of the surface vibration signal, uses historical vibration data to complete the training of the ACGAN-based fault judgment model, and then collects the transformer surface vibration signal in real time, and inputs its characteristic matrix into the trained fault judgment model to achieve non-intrusive detection and fault detection based on vibration signals.

[0057] 2) The present invention matches the collected transformer surface vibration signal to the color space using the corresponding grayscale values ​​of the elements in the Gram-like matrix of the one-dimensional time signal to obtain a two-dimensional image with strong robustness; performs continuous wavelet decomposition processing by translating and scaling the wavelet function, and decomposes the original signal at multiple scales to extract the time-frequency energy characteristics of the transformer vibration signal under different working conditions, and then fuses the feature maps of the two through the 2DPCA method to reduce the dimension to form a feature matrix, which can comprehensively analyze the vibration signal from the time domain and frequency domain, can not only strengthen the connection between the features of different domains, but also supplement the information that may be missed by the single domain features, and has strong robustness.

[0058] 3) When fault samples are scarce, the method proposed in the present invention can also expand the samples of scarce fault types through the ACGAN generation network, thereby improving the accuracy of transformer fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is a flowchart of a fault diagnosis method in an embodiment of the present invention;

[0060] Figure 2 Schematic diagram of the structure of the ACGAN network in an embodiment of the present invention; (a) generator structure, (b) discriminator structure;

[0061] Figure 3 Schematic diagram of a Gram-like matrix, continuous wavelet transform and partial spectrum data after feature fusion corresponding to a fault vibration signal in an embodiment of the present invention; wherein (1), (2) and (3) represent a time domain feature map, a time-frequency feature map and a fusion feature map respectively; (a), (b), (c) and (d) represent a normal signal, a winding looseness, a core looseness and a winding deformation respectively;

[0062] Figure 4 Schematic diagram of the discriminant probability curve of generated samples and real samples in the ACGAN network training process. DETAILED DESCRIPTION

[0063] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0064] Example 1

[0065] The purpose of the present invention is to provide a method for monitoring the mechanical state of a transformer. The method collects vibration signals during operation by installing a vibration sensor on the side wall of the transformer, which can realize non-invasive real-time monitoring of the vibration signals during the operation of the transformer, timely prevent potential fault hazards, and record unknown states to ensure safe and reliable operation of the transformer. In order to achieve the above-mentioned invention purpose, the present invention provides a method for monitoring the mechanical state of a transformer based on an auxiliary generative adversarial network (AuxiliaryClassifier GAN, ACGAN), such as Figure 1 As shown, the method specifically comprises the following steps:

[0066] (1) Obtain the vibration signal of the transformer box surface under the historical operating state of the transformer, denoted as X = {x1, x2, …, x n The Gram matrix feature extraction is performed on the vibration signal of each signal channel to obtain the Gram matrix form of the time domain feature of the vibration signal, and the time domain spectrum coefficient S of the transformer vibration signal is extracted.

[0067] (1.1) Collect the vibration signal on the surface of the transformer box, denoted as X = {x1, x2, …, x n}, perform normalization processing, and the normalization calculation formula is as follows:

[0068]

[0069] (1.2) The normalized vibration sequence is Calculate the angle variable using the arccos(x) function The calculation formula is as follows:

[0070]

[0071] (1.3) Calculate the phase angle of each sequence point, add the cosine value of the sum of the two phase angles to the inner product operation, and calculate the relationship between the two discrete points. The calculation formula is as follows:

[0072]

[0073] In the formula, x i ,y i Represents two different vectors, φ i ,φ jRepresents the angle of the corresponding vector calculated by the arccos(x) function.

[0074] (1.4) The Gram-like matrix of the one-dimensional time signal is as follows. The grayscale values ​​of the elements in the matrix are matched to the color space to obtain a two-dimensional image with strong robustness. The time domain spectrum coefficient S of the transformer vibration signal is extracted. The Gram-like matrix is ​​as follows:

[0075]

[0076] (2) For the transformer vibration signal X = {x1, x2, ..., x n}Continuous wavelet decomposition is performed. Continuous wavelet decomposition can perform multi-scale refinement decomposition of the original signal by translating and scaling the wavelet function, thereby extracting the time-frequency energy characteristic coefficient C of the transformer vibration signal under different working conditions.

[0077] (2.1) The Morlet wavelet basis function is:

[0078]

[0079] Where: f b is the bandwidth; f c is the center frequency.

[0080] According to the basis function ψ(t), we have:

[0081]

[0082] Where: a is the scale expansion factor; b is the time translation expansion factor.

[0083] (2.2) For any continuous signal f(t), its wavelet transform method is:

[0084]

[0085] Where ψ(t) is the Morlet wavelet; * is the conjugate operation; W f (a, b) are wavelet coefficients, and the square of their absolute values ​​is defined as the time-frequency energy distribution spectrum of the signal.

[0086] (3) The time domain spectrum coefficient S of the transformer vibration signal and the time-frequency energy characteristic coefficient C obtained by continuous wavelet transform are:

[0087] S=[s1,s2,…,s n1 ]

[0088] C=[c1,c2,…,c n2 ]

[0089] (4) The two-dimensional feature matrix A can be obtained by concatenating the time domain spectrum coefficient S and the time-frequency energy characteristic coefficient C obtained by continuous wavelet transform through matrix:

[0090] A=[S,C]=[a1,a2,…,a N ]

[0091] (5) Fusion and dimensionality reduction of feature maps using the 2DPCA method

[0092] (5.1) Calculate the average image of the feature matrix A:

[0093]

[0094] (5.2) Center the image and calculate the covariance matrix G of the two-dimensional feature matrix A C , remember G C The characteristic roots of are λ1,λ2,…λ k The corresponding eigenvector is u1u2…u k .

[0095]

[0096] Where: M is the number of vibration signals under different working conditions of the transformer; A i is the feature matrix corresponding to the i-th sample.

[0097] (5.3) Calculate the cumulative contribution rate of the characteristic root:

[0098]

[0099] (5.4) When the cumulative contribution rate p of the characteristic root reaches 0.99, the corresponding first r eigenvectors are selected in turn to form the column-wise projection matrix W of the characteristic matrix A = [w1,w2,…w r ], the new feature matrix R after the original feature matrix A is reduced in dimension is calculated according to the column-wise projection matrix W:

[0100] R=WA

[0101] (6) The new feature matrix R after dimension reduction is divided into a training set and a test set to train the network. According to the idea of ​​dual time scale update rule training (TTRU), different learning rates are set for G and D, which are 0.0003 and 0.0004 respectively.

[0102] (7) Auxiliary generative adversarial network anomaly detection model, which consists of a generative network and a discriminative network

[0103] 7a. The generator structure consists of a fully connected layer, a transposed convolution layer, and a convolution layer. The transposed convolution network uses batch normalization and ReLU activation function, and the output layer consists of convolution and Tanh activation function. The generation process from left to right: the 100-dimensional random noise Z and the 1-dimensional label vector are input into the generation network, converted into a 100-dimensional vector through the embedding layer, and the 100-dimensional vector is transformed into a 128*5*5 vector through the linear unit. After reshaping, it becomes an image of size 5x5 and the number of channels is 128. After that, the image size is changed to 10x10 with the number of channels unchanged through the transposed convolution network. Finally, the convolution layer and tanh activation function are used to generate samples. The Adam algorithm is used to update the parameters during network training.

[0104] 7b. Discriminator model. The network includes 3 convolutional layers. Except for the first convolutional layer, each convolutional layer is accompanied by a batch normalization layer and a Leaky ReLU (alpha=0.2) activation function. The first convolutional layer is only followed by a Leaky ReLU activation function, and the Dropout ratio is 0.25 to alleviate overfitting. The convolution kernel size is 3×3. The step size of the first two convolutional layers is 2, and the step size of the last convolutional layer is 1. After the first two layers, the image size is halved and the number of channels is doubled. After the third convolutional layer, the number of channels is doubled, and the image size remains unchanged. Finally, after full connection, the output is a num_labels+1-dimensional vector, where num_labels represents the number of sample categories. The first element of the vector is passed through the sigmoid activation function to output the probability of the authenticity of the image, and the next num_labels elements are passed through the softmax activation function to output the probability of the corresponding category of the image.

[0105] (8) Training module: During the training process, Z is input into the generator network G together with the sample label C as random noise.

[0106] 8a Initialize the data and randomly generate relevant parameters of the GAN network model;

[0107] 8b Fixed the parameters of the generated network G, from the real sample X real From the data distribution, m samples {x 1 ,x 2 ,x 3 ,…,x m};

[0108] 8c Select m random vectors {z 1 ,z 2 ,z 3 ,…,z m} and the sample label C are simultaneously input into the generative network G by We can get m generated samples

[0109] 8d real sample {x 1 ,x 2 ,x 3 ,…,x m} and generate samples Input into the discriminant network, calculate the discriminant probability, use the cross entropy function as the loss function, and use the gradient ascent method to update the discriminant network parameters θ d , the calculation formula is as follows:

[0110]

[0111]

[0112] Where: D(x i ) is the sample x i The probability of being identified as a real sample by the discriminant network; For sample The probability of being identified as a real sample by the discriminant network; η d is the learning rate of the discriminant network; is the gradient operator.

[0113] 8e Fix the parameters of the discriminant network and extract m random vectors {z 1 ,z 2 ,z 3 ,…,z m};

[0114] 8f uses the discriminant error of the generated sample as the loss function and uses the gradient descent method to update the generated network parameters θ g

[0115]

[0116]

[0117] Where: G(z i ) is a random vector z i The corresponding generated samples; D(G(z i )) is the generated sample G(z i ) The corresponding discriminant network output; η g is the learning rate of the generated network.

[0118] like Figure 4 As shown in Figure 1, it is the discriminant probability curve of generated samples and real samples in the ACGAN network training process.

[0119] (9) Get the data under the label G(z) = X fake , the discriminant network D according to X real With X fake If Xfake With X real If the samples are not similar, the discriminant network will output a lower score, indicating that the generated network G needs to be further optimized. By fixing the parameters of the discriminant network D, the generated network G is trained and adjusted. If the score of the discriminant network D is high, it means that the performance of the discriminant network in distinguishing true from false needs to be improved. At this time, the parameters of the generated network G are fixed and the parameters of the discriminant network D are trained and optimized. In this process of mutual game, Nash equilibrium is reached, and the output of the true and false resolution of the discriminant network is 0.5. The training objective function of the ACGAN network consists of two parts: the cross entropy function L for judging the true and false samples and the cross entropy function L for judging the true and false samples. S And the cross entropy function L for judging the sample category C By training the discriminant network D, the cross function L S With L C The sum is maximized, and the generated network G is trained to make L C With L S The difference is the biggest.

[0120] L s =E[logD(S=real,x)]+E[1-logD(s=real,G(z,y l ))]

[0121] L c =E[logD(C = C i ,x)]+E[logD(C=C i ,G(z,y l ))]

[0122] In the formula, S = real, x represents the real sample input to the discriminant network D, S = real, G (z, y l ) represents the sample generated by the generative network G and input to the discriminative network D, D(C=C i ,x) indicates that the label is C i The probability that the real sample is classified correctly, D(C=C i ,G(z,y l )) indicates that the label generated by generator G is C i The probability that the generated sample is classified correctly.

[0123] (10) The diagnosis module extracts features and fuses the multi-measurement point vibration data obtained in real-time operation, and inputs them into the network model for fault diagnosis. In the process of fault diagnosis, the real-time data features are extracted, and the feature information after dimensionality reduction and fusion is input into the discriminant network D of the ACGAN network model. If the discriminant probability D output is greater than 0.5 or near 0.5, it means that the data at this time is a known fault type, and the known type of fault is classified by the Softmax classifier in the ACGAN network; if the input is a new type of fault, the output probability of the discriminant network D will be very low, far less than 0.5; at this time, the data is recorded, and the operation and maintenance personnel are prompted to conduct a more in-depth study on this data.

[0124] (11) By collecting historical vibration data of the transformer and following the above steps 1-10, a data set is constructed to train the ACGAN network. Samples of scarce fault types can be expanded through the ACGAN generative network.

[0125] (12) By collecting the vibration data of the operating transformer, the features of the original vibration data are extracted through the above steps 1-5, and the data are input into the trained ACGAN network to identify the fault type.

[0126] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A transformer vibration monitoring and fault diagnosis method, characterized in that: The method steps include: Obtain the transformer box surface vibration signal under the real-time operation state of the transformer; Extract time domain spectrum and time-frequency spectrum of transformer box surface vibration signal; The extracted time domain spectrum and time-frequency spectrum are fused and dimensionally reduced to obtain a feature matrix; The feature matrix is ​​input into the auxiliary generative adversarial network anomaly detection model trained based on the historical data of the transformer box surface vibration signal to identify the fault type and obtain the transformer fault classification; The generative network of the auxiliary generative adversarial network anomaly detection model includes a fully connected layer, a transposed convolutional layer and a convolutional layer; The random noise Z and the label vector are input into the generator, and the input is converted into a multi-dimensional vector through the embedding layer and the linear unit. The multi-dimensional vector is reconstructed into an image, and then the image size is transformed through the transposed convolution network, and finally the convolution layer and the activation function are used to generate samples; The transposed convolutional network uses batch normalization and ReLU activation function, and the output layer is composed of convolution and Tanh activation function; the Adam algorithm is used to update parameters during network training; The discriminator network of the auxiliary generative adversarial network anomaly detection model includes three convolutional layers; except for the first layer, each convolutional layer is accompanied by a batch normalization layer and a Leaky ReLU activation function, and the first convolutional layer is only followed by a Leaky ReLU activation function; The discriminator input passes through three convolutional layer channels and then fully connected to output a num_labels+1-dimensional vector; where num_labels represents the number of sample categories; Take the first element of the output vector and pass it through the sigmoid activation function to output the authenticity probability of the image; The last num_labels elements are activated by the softmax function to output the probability of the corresponding category of the image.

2. A transformer vibration monitoring and fault diagnosis method according to claim 1, characterized in that: The time domain spectrum and time-frequency spectrum extracted from the transformer box surface vibration signal are specifically as follows: Perform Gram matrix feature extraction on the transformer vibration signal X of each extracted signal channel to obtain the time domain spectrum coefficient S of the transformer vibration signal; The transformer vibration signal X is subjected to continuous wavelet decomposition to obtain the time-frequency energy characteristic coefficient C.

3. A transformer vibration monitoring and fault diagnosis method according to claim 2, characterized in that: The specific steps of performing the Gram-like matrix feature extraction to obtain the time domain spectrum coefficient S of the transformer vibration signal include: Normalize the vibration signal X; Calculate the normalized vibration sequence The phase angle of each sequence point in ; The cosine value of the sum of the two phase angles is added to the inner product operation to calculate the relationship between the two discrete points in the vector. The calculation formula is as follows: In the formula, x i ,y i Represents two different vectors, φ i ,φ j Represents the calculated angle of the corresponding vector; The time domain spectrum coefficient S of the transformer vibration signal is extracted using a Gram-like matrix.

4. A transformer vibration monitoring and fault diagnosis method according to claim 2, characterized in that: The specific steps of performing continuous wavelet decomposition processing on the transformer vibration signal to obtain the time-frequency energy characteristic coefficient C include: For the transformer vibration signal X, the signal is decomposed in multiple scales by translating and scaling wavelet functions, and the time-frequency energy characteristic coefficient C of the transformer vibration signal is extracted.

5. A transformer vibration monitoring and fault diagnosis method according to claim 2, characterized in that: The specific steps of fusing and reducing the dimensionality of the extracted time domain spectrum and time-frequency spectrum include: The time domain spectrum coefficient S and the time-frequency energy characteristic coefficient C matrix are concatenated to obtain the characteristic matrix A; Calculate the covariance matrix G of the feature matrix A C , and get the covariance matrix G C The characteristic roots and corresponding eigenvectors of Calculate the covariance matrix G C The cumulative contribution rate of characteristic roots p; When the cumulative contribution rate p of the characteristic root reaches the set value, the corresponding first r eigenvectors are selected in turn to form the column-direction projection matrix W of the characteristic matrix A, and the new characteristic matrix R = W·A after the dimension reduction of the characteristic matrix A is calculated according to the column-direction projection matrix W.

6. A transformer vibration monitoring and fault diagnosis method according to claim 1, characterized in that: The training steps of the auxiliary generation adversarial network anomaly detection model include: Initialize data and randomly generate relevant parameters of the GAN network model; Fixed the parameters of the generated network G, from the real sample X real From the data distribution, m samples {x 1 ,x 2 ,x 3 ,…,x m }; Select m random vectors as random noise and input them into the generative network G at the same time as the sample label to obtain m generated samples The real sample {x 1 , x 2 , x 3 ,...,x m } and generate samples Input into the discriminant network D, calculate the discriminant probability, use the cross entropy function as the loss function, and use the gradient ascent method to update the parameters of the discriminant network D; Fix the parameters of the discriminant network D and extract m random vectors {z 1 , z 2 , z 3 , ..., z m }; Taking the discriminant error of the generated sample as the loss function, the gradient descent method is used to update the parameters θ of the generated network G g ; After the network parameters are updated, the data G(z)=X under the sample label C is obtained. fake ; The discriminant network D is based on X real With X fake If X fake With X real If they are not similar, fix the parameters of the discriminant network D and train and adjust the generative network G. If the output score of the discriminant network D is high, fix the parameters of the generative network G and train to optimize the parameters of the discriminant network D. In the process of mutual game, Nash equilibrium is reached, and the output resolution of the authenticity judgment network is 0.

5.

7. A transformer vibration monitoring and fault diagnosis method according to claim 6, characterized in that: The training objective function of the auxiliary generative adversarial network includes: The cross entropy function L for judging the authenticity of samples S : L s =E[logD(s=real,x)]+E[1-logD(s=real,G(z,y l ))] The cross entropy function L for determining sample categories C : L c =E[logD(C=C i ,x)]+E[logD(C=C i ,G(z,y l ))] In the formula, S = real, x represents the real sample input to the discriminant network D, S = real, G (z, y l ) represents the sample generated by the generative network G and input to the discriminative network D, D(C=C i , x) indicates that the label is C i The probability that the real sample is classified correctly, D(C=C i , G(z,y l )) indicates that the label generated by generator G is C i The probability that the generated samples are classified correctly; By training the discriminant network D, the cross function L S With L C The sum is maximized, and the generated network G is trained to make L C With L S The difference is the biggest.

8. A transformer vibration monitoring and fault diagnosis method according to claim 6, characterized in that: During the troubleshooting process: If the output probability of the discriminant network D is greater than 0.5 or close to 0.5, the data at this time is a known fault type, and the known type of fault is classified by the auxiliary generative adversarial network anomaly detection model; If the output probability of the discriminant network D is far less than 0.5, the input is a new type of fault. At this time, the input data is recorded and the operation and maintenance personnel are prompted.

Images