Transformer fault identification method and device and readable storage medium

Abstract

The invention discloses a transformer fault identification method and device and a readable storage medium, and belongs to the field of power system state monitoring. Comprising the steps that voiceprint signals and vibration signals in a normal working mode and a fault mode are made into normal working sample data and fault sample data, and feature components of normal working samples and feature components of fault samples are obtained based on a symplectic geometric mode decomposition algorithm; the feature vectors are used for importing a cyclic generative adversarial network model game to obtain new fault sample feature vectors and extended fault sample feature vectors, and the new fault sample feature vectors and the extended fault sample feature vectors are used for making a training set and a test set; and finally, obtaining a fault identification model with optimal performance, and carrying out real-time detection on the voiceprint signal and the vibration signal of the power transformer. According to the method, fault mode identification and transformer internal and external fault point identification under complex conditions, environments and power system load fluctuation can be realized, and power system maintenance is guided.

Description

Technical Field

[0001] The invention relates to a transformer fault identification method, a device and a readable storage medium, and belongs to the field of power system state monitoring and fault diagnosis. Background Art

[0002] As one of the core equipment of the power system, the power transformer plays an important role in energy conversion and power transmission in the power grid. Its operating status is related to the safety and stability of the entire power system. At present, the total number of power transformers in operation in my country is large and still showing an upward trend. Therefore, accurate diagnosis of whether the transformer has a fault and what kind of fault has occurred is an important measure to ensure the normal operation of the power system and reduce economic losses.

[0003] In the long-term operation, transformers are affected by complex and diverse environments and different working conditions, and are prone to problems such as material degradation, fatigue wear of local components, and reduced insulation performance, which can induce defects such as abnormal vibration and partial discharge. In severe cases, they can cause tripping and other faults, resulting in large-scale power outages in the power grid. With the rapid development of artificial intelligence technology, the use of artificial intelligence technology to deal with the above faults has become a necessary measure to promote the intelligent development of transformer fault diagnosis technology. However, in actual work, since power transformer failures are mostly low-probability and sporadic events, directly collecting fault samples will lead to a lack of fault samples and an unbalanced data set. In addition, the correlation between power transformer failures and different working conditions cannot be effectively obtained through unbalanced data sets. When the training data is highly unbalanced, the classification model usually tends to favor majority class samples and ignores minority class samples, resulting in reduced recognition accuracy of minority class samples, which restricts the training effect of the machine learning model.

[0004] Therefore, the existing power transformer fault diagnosis methods have poor diagnostic effects and low accuracy, and cannot provide effective data references for power system status monitoring, power transformer fault diagnosis and maintenance. Therefore, how to collect targeted target samples under complex working conditions and how to accurately and effectively identify and diagnose power transformer faults under real-time changing power system load trends are important issues that need to be solved. Summary of the invention

[0005] The purpose of the present invention is to provide a transformer fault identification method, device and readable storage medium, collect operating samples of oil-immersed power transformers under complex working conditions, use the symplectic geometric mode decomposition algorithm to extract characteristic components of the voiceprint signal and vibration signal in the operating samples, obtain the characteristic data of the comprehensive transformer operating state and mechanical state, and combine the cyclic generative adversarial network model CycleGAN to perform cyclic adversarial operations on the characteristic vectors of normal working modes and multiple fault modes, comprehensively expand the fault samples, realize the fault mode classification and internal and external fault point identification of transformers under complex conditions, environments and power system load fluctuations, and solve the technical problems described in the background technology.

[0006] In order to solve the above technical problems, the present invention is implemented by adopting the following technical solutions.

[0007] A transformer fault identification method, comprising:

[0008] The soundprint signals and vibration signals of the oil-immersed power transformer in normal working mode and various fault modes under different working conditions are collected respectively;

[0009] The soundprint signals and vibration signals in normal working modes and various fault modes under different working conditions are respectively made into normal working sample data and fault sample data;

[0010] Decomposing the normal working sample data and the fault sample data based on the symplectic geometric mode decomposition algorithm to obtain the characteristic components of the normal working samples and the characteristic components of the fault samples;

[0011] According to the characteristic components of the normal working samples, a characteristic vector of the normal working samples of the transformer is formed; and according to the characteristic components of the fault samples, a characteristic vector of the fault samples of the transformer is formed;

[0012] The feature vectors of normal working samples and faulty samples are introduced into the cyclic generative adversarial network model for game playing to obtain new feature vectors of faulty samples;

[0013] Performing a union operation on the feature vector of the fault sample and the feature vector of the new fault sample to obtain the feature vector of the expanded fault sample;

[0014] Create a training set and a test set based on the feature vectors of the expanded fault samples and the feature vectors of the normal working samples;

[0015] The training set is imported into the stack denoising autoencoder model for training to obtain a fault recognition model, and the fault recognition model is tested using the test set to obtain an optimized fault recognition model;

[0016] Based on the optimized fault identification model, the soundprint signal and vibration signal of the power transformer are detected in real time, and the fault pattern recognition results are used to identify the internal and external fault points of the transformer, thereby guiding the maintenance of the power system.

[0017] Optionally, the multiple failure modes include: partial discharge, low-energy discharge, high-energy discharge, low-temperature overheating, medium-temperature overheating and high-temperature overheating;

[0018] The voiceprint signal is used to detect the internal operating state of the power transformer;

[0019] The vibration signal is used to provide feedback on the assembly connection status of the transformer components.

[0020] Optionally, the normal working sample data and the fault sample data are decomposed based on a symplectic geometric mode decomposition algorithm to obtain characteristic components of the normal working samples and characteristic components of the fault samples, including:

[0021] Signal processing is performed on the normal working sample data and the fault sample data to obtain the normal working sample and the fault sample standardized data. The normal working sample standardized data are the voiceprint signal An i and vibration signal Vn i ; The standardized data of the fault samples are the voiceprint signal Af j and vibration signal Vf j , i is the data volume of normal working samples, j is the data volume of fault samples, and j< <i;

[0022] The normalized data of the normal working samples and the fault samples are reconstructed in phase space to obtain the trajectory matrix X An , X Vn , X Af and X Vf ;

[0023] According to the trajectory matrix X An , X Vn , X Af and X Vf Calculate the corresponding covariance symmetric matrix A=X T X, and construct multiple Hamiltonian matrices M from the corresponding covariance symmetric matrix A;

[0024] The symplectic orthogonal matrix in the symplectic geometric mode decomposition algorithm is used to solve multiple Hamiltonian matrices M respectively to obtain the eigenvalues ​​of normal working sample data and fault sample data. And the eigenvalue The associated eigenvector;

[0025] The characteristic values ​​of normal working sample data and fault sample data The associated eigenvectors are arranged in descending order to obtain the decomposed eigencomponents of normal working samples and fault samples, which are S-An i 、S-Vn i 、S-Af j and S-Vf j ;

[0026] Among them, the trajectory matrix X An The expression is:

[0027] (1)

[0028] Where d is the embedding dimension, τ is the time delay, An is the voiceprint signal in normal working mode, and m is the trajectory matrix X An The number of rows, N is the signal length, and m=N-(d-1)τ;

[0029] The expression of Hamiltonian matrix M is:

[0030] (2)

[0031] Where A is the covariance symmetric matrix, A T is the transposed matrix of A;

[0032] The expression of a symplectic orthogonal matrix is:

[0033] (3)

[0034] Where Q is a symplectic orthogonal matrix, is the transposed matrix of Q, H is a diagonal matrix, H=M 2 , R is the submatrix after matrix transformation; B is an upper triangular matrix, and its eigenvalues ​​are λ1, λ2, λ3,…, λ d , d is the number of eigenvalues ​​contained in matrix B; and the eigenvalues ​​of the symmetric matrix A are ,in , is any eigenvalue in the matrix B; w=1, 2,…, d.

[0035] Optionally, a feature vector of a normal working sample of the transformer is formed according to the feature components of the normal working sample; and a feature vector of a fault sample of the transformer is formed according to the feature components of the fault sample, including:

[0036] The characteristic component S-An of the normal working sample after decomposition i and S-Vn i , combined into the characteristic vector of the normal working sample of the transformer [S-An i , S-Vn i]; and the characteristic component S-Af of the decomposed fault sample j and S-Vf j , combined into the characteristic vector of transformer fault samples [S-Af j , S-Vf j ].

[0037] Optionally, the feature vectors of the normal working samples and the feature vectors of the fault samples of the transformer are introduced into the cyclic generative adversarial network model for game playing to obtain new feature vectors of the fault samples, including:

[0038] The feature vector of the normal working sample [S-An i , S-Vn i ] is processed into the data domain X; and the feature vector of the fault sample [S-Af j , S-Vf j ] is processed into data domain Y;

[0039] Data domain X and data domain Y are based on the generator G, F and discriminator D of the cyclic generative adversarial network model x , D y Play games and obtain new fault sample feature vectors.

[0040] Optionally, the data domain X and the data domain Y are based on the generator G, F and the discriminator D of the cyclic generative adversarial network model x , D y Play games to obtain new fault sample feature vectors, including:

[0041] The generator G is used to map the data in the data domain X to the data domain Y to obtain a sample vector G(x) simulating a fault sample; and the generator F is used to map the data in the data domain Y to the data domain X to obtain a sample vector F(y) simulating a normal working sample;

[0042] Using the discriminator D y Determine the authenticity of the sample vector G(x) of the simulated fault sample. If it is true, output the sample vector G(x) of the simulated fault sample; if it is false, import the sample vector G(x) of the simulated fault sample obtained this time into the generator G and the discriminator D. y The generated adversarial loss function calculates the generator G and the discriminator D y The adversarial loss value is used to update and optimize the generator G and the discriminator D according to the adversarial loss value. y ;

[0043] Using the discriminator D xDetermine the authenticity of the sample vector F(y) that simulates a normal working sample. If it is true, output the sample vector F(y) that simulates a normal working sample. If it is false, import the sample vector F(y) that simulates a normal working sample obtained this time into the generator F and the discriminator D. x The adversarial loss function generated calculates the generator F and the discriminator D x The adversarial loss value is used to update the optimized generator F and discriminator D according to the adversarial loss value. x ;

[0044] The sample vector G(x) of the simulated fault sample that has been output is mapped to the data domain X by using the generator F, so as to obtain the sample vector F[G(x)] of the simulated normal working sample that is cyclic; and the sample vector F(y) of the simulated normal working sample that has been output is mapped to the data domain Y by using the generator G, so as to obtain the feature vector G[F(y)] of the simulated fault sample that is cyclic;

[0045] Through the discriminator D x Determine the authenticity of the sample vector F[G(x)] of the cyclic simulated normal working sample. If it is true, output the sample vector F[G(x)] of the cyclic simulated normal working sample. If it is false, import the sample vector F[G(x)] of the cyclic simulated normal working sample obtained this time into the cycle consistency loss function, calculate the cycle consistency loss value, and update the optimization generator F according to the cycle consistency loss value.

[0046] At the same time, through the discriminator D y Determine the authenticity of the cyclic simulated fault sample feature vector G[F(y)], if it is true, output the cyclic simulated fault sample feature vector G[F(y)], if it is false, import the cyclic simulated fault sample feature vector G[F(y)] obtained this time into the cycle consistency loss function, calculate the cycle consistency loss value, and update the generator G according to the cycle consistency loss value;

[0047] Repeat the above steps until the generator G, F and the discriminator D x , D y When the total loss function is less than the set threshold or meets the upper limit of the number of iterations, the output cyclic simulated fault sample feature vector G[F(y)] is used for a new fault sample feature vector, and cyclically adversarial data in the data domain X and j data in the data domain Y until (ij) new fault sample feature vectors are generated.

[0048] Optionally, the generator G and the discriminator D y The resulting adversarial loss function is:

[0049] (4)

[0050] In the formula, is the adversarial loss function generated by the generator and the discriminator, represents the expectation of data y, The data y in the data domain Y passes through the discriminator D y The output value obtained is represents the expectation of data x, is the natural logarithm function;

[0051] Generator F and Discriminator D x The resulting adversarial loss function is:

[0052] (5)

[0053] The cycle consistency loss function of generator G and F is:

[0054] (6)

[0055] In the formula, is the generator cycle consistency loss function, x~P(x) is the data x is the X domain data that obeys the P distribution, y~Q(y) is the data y is the Y domain data that obeys the Q distribution, is 1 norm;

[0056] The total loss function is:

[0057] (7)

[0058] In the formula, θ is the weight parameter.

[0059] Optionally, a training set and a test set are prepared according to the feature vectors of the expanded fault samples and the feature vectors of the normal working samples, including:

[0060] The feature vector of the normal working sample [S-An i , S-Vn i ] dataset and the feature vector of the expanded fault sample [S-Af i S-Vf i ] to obtain a sampling sample;

[0061] The sample is divided into training set and test set in the ratio of 80%: 20%.

[0062] Optionally, the number of hidden layers of the stack denoising autoencoder model is 4, the activation function is set to tanh, the learning rate is 0.001, and the output layer activation function is softmax.

[0063] A transformer fault identification device, comprising:

[0064] Acquisition and processing module, feature extraction module, game module, fault identification module and application module;

[0065] The acquisition and processing module is used to respectively acquire the soundprint signals and vibration signals of the oil-immersed power transformer in normal working mode and multiple fault modes under different working conditions; and to respectively make the soundprint signals and vibration signals in normal working mode and multiple fault modes under different working conditions into normal working sample data and fault sample data;

[0066] The feature extraction module is used to decompose the normal working sample data and the fault sample data based on the symplectic geometric mode decomposition algorithm to obtain the characteristic components of the normal working samples and the characteristic components of the fault samples; construct the characteristic vector of the normal working samples of the transformer according to the characteristic components of the normal working samples; and construct the characteristic vector of the fault samples of the transformer according to the characteristic components of the fault samples;

[0067] The game module is used to import the feature vector of the normal working sample and the feature vector of the fault sample into the cyclic generative adversarial network model for game, so as to obtain a new feature vector of the fault sample; and perform a union operation on the feature vector of the fault sample and the new feature vector of the fault sample to obtain the feature vector of the expanded fault sample;

[0068] The fault identification module is used to prepare a training set and a test set according to the feature vectors of the expanded fault samples and the feature vectors of the normal working samples; and to import the training set into the stack denoising autoencoder model for training to obtain a fault identification model, and to use the test set to test the fault identification model to obtain an optimized fault identification model;

[0069] The application module is used to perform real-time detection of the soundprint signal and vibration signal of the power transformer based on the optimized fault identification model, and obtain the fault pattern recognition result for identifying the internal and external fault points of the transformer to guide the maintenance of the power system.

[0070] A computer-readable storage medium stores a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned transformer fault identification method.

[0071] Compared with the prior art, the present invention has the following beneficial effects:

[0072] Collect samples of oil-immersed power transformers running under complex working conditions, use the symplectic geometry mode decomposition algorithm to extract characteristic components from the soundprint signal and vibration signal in the running samples, obtain the characteristic data of the comprehensive transformer operating state and mechanical state, and combine the cyclic generative adversarial network model CycleGAN to perform cyclic adversarial operations on the characteristic vectors of normal working modes and multiple fault modes, comprehensively expand the fault samples, realize the fault mode classification and transformer internal and external fault point identification under complex conditions, environments and power system load fluctuations;

[0073] The present invention also uses a stack denoising autoencoder model to reduce noise on the oil-immersed power transformer training data and enhance the features of the voiceprint signal and vibration signal, which is conducive to learning more abstract and complex data feature representations, thereby improving the robustness of the invented method, and ultimately achieving the goal of improving the diagnostic accuracy of the fault identification model and enhancing the deep learning capability, and avoiding the shortcomings of information redundancy in the samples generated by the traditional SMOTE algorithm.

[0074] In addition, the amount of sample data calculation expanded by the method of the present invention is smaller than that of the prior art method, thereby reducing the amount of equipment calculation and balancing the load of the computing equipment; at the same time, the fault diagnosis results are accurately and quickly exported, which can provide online, real-time, accurate and effective reference data for the power monitoring system, which can be used to guide the maintenance of the power system, shorten the troubleshooting time and save maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 is a flow chart of the transformer fault identification method of the present invention;

[0076] Figure 2 is a flow chart of the fault identification process of the present invention through the CycleGAN model;

[0077] Figure 3 is the fault identification confusion matrix of the present invention under situation 1;

[0078] Figure 4 is the fault identification confusion matrix of the present invention under situation 2;

[0079] Figure 5 is the fault identification confusion matrix of the present invention under situation 3;

[0080] Figure 6 is the fault identification confusion matrix of the present invention under situation 4;

[0081] Figure 7 It is a table of structural parameters of the CycleGAN model of the present invention. DETAILED DESCRIPTION

[0082] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0083] Example 1

[0084] In order to overcome the shortcomings of the existing power transformer fault identification method, this embodiment provides a transformer fault identification method, such as Figure 1 and Figure 2 The process shown includes:

[0085] Step 1: respectively collect the soundprint signals and vibration signals of the oil-immersed power transformer in normal working mode and multiple fault modes under different working conditions;

[0086] Step 2: The voiceprint signals and vibration signals in normal working modes and various fault modes under different working conditions are respectively made into normal working sample data and fault sample data;

[0087] Step 3: Decompose the normal working sample data and the fault sample data based on the symplectic geometric mode decomposition algorithm to obtain the characteristic components of the normal working samples and the characteristic components of the fault samples;

[0088] Step 4: construct a feature vector of the normal working sample of the transformer according to the feature components of the normal working sample; and construct a feature vector of the fault sample of the transformer according to the feature components of the fault sample;

[0089] Step 5: Import the feature vector of the normal working sample and the feature vector of the fault sample into the cyclic generative adversarial network model for game, and obtain a new feature vector of the fault sample;

[0090] Step 6: Perform a union operation on the feature vector of the fault sample and the feature vector of the new fault sample to obtain the feature vector of the expanded fault sample;

[0091] Step 7: Create a training set and a test set based on the feature vectors of the expanded fault samples and the feature vectors of the normal working samples;

[0092] Step 8: Import the training set into the stack denoising autoencoder model for training to obtain a fault recognition model, and use the test set to test the fault recognition model to obtain an optimized fault recognition model;

[0093] Step 9: Based on the optimized fault identification model, the voiceprint signal and vibration signal of the power transformer are detected in real time to obtain the fault pattern recognition results to identify the internal and external fault points of the transformer and guide the maintenance of the power system.

[0094] Preferably, the multiple fault modes include: the fault modes mainly include partial discharge (FS1), low energy discharge (FS2), high energy discharge (FS3), low temperature overheating (FS4), medium temperature overheating (FS5), high temperature overheating (FS6), that is, there are 7 operating states in total. ;

[0095] The voiceprint signal is used to detect the internal operating state of the power transformer;

[0096] The vibration signal is used to provide feedback on the assembly connection status of the transformer components.

[0097] Preferably, in step 3, the method of decomposing the normal working sample data and the fault sample data based on the symplectic geometric mode decomposition algorithm to obtain the characteristic components of the normal working samples and the characteristic components of the fault samples includes:

[0098] Step 3.1: Perform signal processing on the normal working sample data and the fault sample data respectively to obtain the normal working sample and the fault sample standardized data. The normal working sample standardized data are An i and Vn i ; The standardized data of the fault samples are Af j and Vf j , i is the data volume of normal working samples, j is the data volume of fault samples, and j< <i;

[0099] Step 3.2: Reconstruct the phase space of the normal working samples and the normalized data of the fault samples to obtain the trajectory matrix X An , X Vn , X Af and X Vf ;

[0100] Step 3.3: According to the trajectory matrix X An , X Vn , X Af and X Vf Calculate the corresponding covariance symmetric matrix A=X T X, and construct multiple Hamiltonian matrices M from the corresponding covariance symmetric matrix A;

[0101] Step 3.4: Use the symplectic orthogonal matrix in the symplectic geometric mode decomposition algorithm to solve multiple Hamiltonian matrices M respectively to obtain the eigenvalues ​​of normal working sample data and fault sample data And the eigenvalue The associated eigenvector;

[0102] Step 3.5: Characteristic values ​​of normal working sample data and fault sample data The associated eigenvectors are arranged in descending order to obtain the decomposed eigencomponents of normal working samples and fault samples, which are S-An i 、S-Vn i 、S-Af j and S-Vf j ;

[0103] Among them, the trajectory matrix X An The expression is:

[0104] (1)

[0105] Where d is the embedding dimension, τ is the time delay, An is the voiceprint signal in normal working mode, and m is the trajectory matrix X An The number of rows, N is the signal length, and m=N-(d-1)τ;

[0106] The expression of Hamiltonian matrix M is:

[0107] (2)

[0108] Where A is the covariance symmetric matrix, A T is the transposed matrix of A;

[0109] The expression of a symplectic orthogonal matrix is:

[0110] (3)

[0111] Where Q is a symplectic orthogonal matrix, is the transposed matrix of Q, H is a diagonal matrix, H=M 2 , R is the submatrix after matrix transformation; B is an upper triangular matrix, and its eigenvalues ​​are λ1, λ2, λ3,…, λ d , d is the number of eigenvalues ​​contained in matrix B; and the eigenvalues ​​of the symmetric matrix A are ,in , is any eigenvalue in the matrix B; w=1, 2,…, d.

[0112] Preferably, in step 4, the method for forming a characteristic vector of a normal working sample of the transformer according to the characteristic components of the normal working sample; and the method for forming a characteristic vector of a fault sample of the transformer according to the characteristic components of the fault sample include:

[0113] Step 4.1: Decompose the characteristic components S-An of the normal working samples i and S-Vn i , combined into the characteristic vector of the normal working sample of the transformer [S-An i, S-Vn i ]; and the characteristic component S-Af of the decomposed fault sample j and S-Vf j , combined into the characteristic vector of transformer fault samples [S-Af j , S-Vf j ].

[0114] Preferably, in step 5, the feature vectors of the normal working samples and the feature vectors of the fault samples of the transformer are introduced into the cyclic generative adversarial network model for game playing to obtain a new feature vector of the fault samples, including:

[0115] Step 5.1: The feature vector of the normal working sample [S-An i , S-Vn i ] is processed into the data domain X; and the feature vector of the fault sample [S-Af j , S-Vf j ] is processed into data domain Y;

[0116] Step 5.2: Data domain X and data domain Y are based on the generator G, F and discriminator D of the cyclic generative adversarial network model x , D y Play games and obtain new fault sample feature vectors.

[0117] Preferably, in step 5.2, the data domain X and the data domain Y are based on the generator G, F and the discriminator D of the cyclic generative adversarial network model. x , D y Methods for playing games and obtaining new fault sample feature vectors include:

[0118] Step 5.2.1: Use generator G to map the data in data domain X to data domain Y to obtain a sample vector G(x) simulating a fault sample; and use generator F to map the data in data domain Y to data domain X to obtain a sample vector F(y) simulating a normal working sample;

[0119] Step 5.2.2: Use the discriminator D y Determine the authenticity of the sample vector G(x) of the simulated fault sample. If it is true, output the sample vector G(x) of the simulated fault sample; if it is false, import the sample vector G(x) of the simulated fault sample obtained this time into the generator G and the discriminator D. y The generated adversarial loss function calculates the generator G and the discriminator D y The adversarial loss value is used to update and optimize the generator G and the discriminator D according to the adversarial loss value. y ;

[0120] Step 5.2.3: Use the discriminator D xDetermine the authenticity of the sample vector F(y) that simulates a normal working sample. If it is true, output the sample vector F(y) that simulates a normal working sample. If it is false, import the sample vector F(y) that simulates a normal working sample obtained this time into the generator F and the discriminator D. x The adversarial loss function generated calculates the generator F and the discriminator D x The adversarial loss value is used to update the optimized generator F and discriminator D according to the adversarial loss value. x ;

[0121] Step 5.2.4: Use the generator F to map the sample vector G(x) of the simulated fault sample that has been output to the data domain X, and obtain the sample vector F[G(x)] of the simulated normal working sample that is cyclic; and use the generator G to map the sample vector F(y) of the simulated normal working sample that has been output to the data domain Y, and obtain the feature vector G[F(y)] of the simulated fault sample that is cyclic;

[0122] Step 5.2.5: Through the discriminator D x Determine the authenticity of the sample vector F[G(x)] of the cyclic simulated normal working sample. If it is true, output the sample vector F[G(x)] of the cyclic simulated normal working sample. If it is false, import the sample vector F[G(x)] of the cyclic simulated normal working sample obtained this time into the cycle consistency loss function, calculate the cycle consistency loss value, and update the optimization generator F according to the cycle consistency loss value.

[0123] At the same time, through the discriminator D y Determine the authenticity of the cyclic simulated fault sample feature vector G[F(y)], if it is true, output the cyclic simulated fault sample feature vector G[F(y)], if it is false, import the cyclic simulated fault sample feature vector G[F(y)] obtained this time into the cycle consistency loss function, calculate the cycle consistency loss value, and update the generator G according to the cycle consistency loss value;

[0124] Step 5.2.6: Repeat the above steps until the generator G, F and the discriminator D x , D y When the total loss function is less than the set threshold or meets the upper limit of the number of iterations, the output cyclic simulated fault sample feature vector G[F(y)] is used for a new fault sample feature vector, and cyclically adversarial data in the data domain X and j data in the data domain Y until (ij) new fault sample feature vectors are generated.

[0125] In this embodiment, when the network training process reaches Nash equilibrium, fault sample expansion is achieved; finally, a stack denoising autoencoder is used for fault identification. The present invention fuses the fault characteristics of different signals of the power transformer, uses a generative adversarial network to expand the number of fault samples to enhance the fault feature information, and solves the problems of low fault detection accuracy and low efficiency of power equipment under sample imbalance.

[0126] Furthermore, the generator G and the discriminator D y The resulting adversarial loss function is:

[0127] (4)

[0128] In the formula, is the adversarial loss function generated by the generator and the discriminator, represents the expectation of data y, The data y in the data domain Y passes through the discriminator D y The output value obtained is represents the expectation of data x, is the natural logarithm function;

[0129] Generator F and Discriminator D x The resulting adversarial loss function is:

[0130] (5)

[0131] The cycle consistency loss function of generator G and F is:

[0132] (6)

[0133] In the formula, is the generator cycle consistency loss function, x~P(x) is the data x is the X domain data that obeys the P distribution, y~Q(y) is the data y is the Y domain data that obeys the Q distribution, is 1 norm;

[0134] The total loss function is:

[0135] (7)

[0136] In the formula, θ is the weight parameter.

[0137] Preferably, in step 7, the method of preparing the training set and the test set according to the expanded feature vectors of the fault samples and the feature vectors of the normal working samples includes:

[0138] Step 7.1: For the feature vector of the normal working sample [S-An i , S-Vn i] dataset and the feature vector of the expanded fault sample [S-Af i S-Vf i ] to obtain a sampling sample;

[0139] Step 7.2: Divide the sample into training set and test set in a ratio of 80%: 20%.

[0140] Preferably, in step 8, the number of hidden layers of the stack denoising autoencoder model is 4, the activation function is set to tanh, the learning rate is 0.001, and the output layer activation function is softmax. In this embodiment, the vibration signal and voiceprint signal data of the oil-immersed power transformer in its normal state (FS0) and a certain fault mode when working are used. The parameters of the network model are set as follows: Figure 7 As shown in the figure, the operation data of oil-immersed power transformers are collected under four specific situations according to complex situations, environments and power system load fluctuations. When the oil-immersed power transformer works similar to situation 1, the vibration signal and voiceprint signal of the power transformer are used to expand the samples and perform fault identification. Among them, the number of samples of FS0 in the normal state is 1000, and the number of samples of FS1-FS6 is 200, that is, the data imbalance ratio is 10:2. The fault identification confusion matrix results obtained when this method was accepted are as follows Figure 3 As shown, it can be seen that the fault identification accuracy of the present invention is about 98.57%; when the oil-immersed power transformer works in approximate situation 2, the power transformer vibration signal and voiceprint signal are used to expand the sample and perform fault identification; wherein, the number of normal state FS0 samples is 1000, and the number of FS1-FS6 samples is 100, that is, the data imbalance ratio is 10:1, and the parameters of the CycleGAN network and SDAE are the same as above, and CycleGAN is used to generate fault samples, and SDAE is used to identify the fault, and the identification result is obtained as follows: Figure 4 As shown, its accuracy is 99.14%. When the oil-immersed power transformer works in similar situations 3 and 4, experimental comparisons are carried out in situations 3 and 4 respectively. In situation 3, only the vibration signal of the power transformer is used for sample expansion and fault identification. In the normal state, the number of FS0 samples is 1000, and the number of samples of FS1-FS6 is 200, that is, the data imbalance ratio is 10:2; in situation 4, only the voiceprint signal is used for sample expansion and fault identification. In the normal state, the number of FS0 samples is 1000, and the number of samples of FS1-FS6 is 200, that is, the data imbalance ratio is 10:2; the sample generation model, recognition method and parameter setting are consistent with the above, and the recognition results are respectively Figure 5 and 6, and its accuracy is 88.86% and 94.00%. It can be concluded that the method of the present invention fuses the voiceprint signal and the vibration signal to enhance the fault feature information of the power transformer compared with the traditional single signal feature extraction method; secondly, the method of the present invention has better popularity than the common dissolved gas analysis DGA method, because the vibration signal and the voiceprint signal only need to be installed outside the transformer, and the gas composition analysis sensor needs to invade the inside of the transformer. In summary, the method of the present invention to expand the data under data imbalance to realize the fault identification method of the power transformer has better effect and applicability, and the identification accuracy can reach more than 98%.

[0141] This embodiment is not limited to the above four situations. Since the working conditions of the transformer are various and mainly depend on its operating state, application scenario and specific design parameters, the operation data of other situations can also be subdivided and collected according to specific scenarios.

[0142] Example 2

[0143] This embodiment provides a transformer fault identification device, including:

[0144] Acquisition and processing module, feature extraction module, game module, fault identification module and application module;

[0145] The acquisition and processing module is used to respectively acquire the soundprint signals and vibration signals of the oil-immersed power transformer in normal working mode and multiple fault modes under different working conditions; and to respectively make the soundprint signals and vibration signals in normal working mode and multiple fault modes under different working conditions into normal working sample data and fault sample data;

[0146] The feature extraction module is used to decompose the normal working sample data and the fault sample data based on the symplectic geometric mode decomposition algorithm to obtain the characteristic components of the normal working samples and the characteristic components of the fault samples; construct the characteristic vector of the normal working samples of the transformer according to the characteristic components of the normal working samples; and construct the characteristic vector of the fault samples of the transformer according to the characteristic components of the fault samples;

[0147] The game module is used to import the feature vector of the normal working sample and the feature vector of the fault sample into the cyclic generative adversarial network model for game, so as to obtain a new feature vector of the fault sample; and perform a union operation on the feature vector of the fault sample and the new feature vector of the fault sample to obtain the feature vector of the expanded fault sample;

[0148] The fault identification module is used to prepare a training set and a test set according to the feature vectors of the expanded fault samples and the feature vectors of the normal working samples; and to import the training set into the stack denoising autoencoder model for training to obtain a fault identification model, and to use the test set to test the fault identification model to obtain an optimized fault identification model;

[0149] The application module is used to perform real-time detection of the soundprint signal and vibration signal of the power transformer based on the optimized fault identification model, and obtain the fault pattern recognition result for identifying the internal and external fault points of the transformer to guide the maintenance of the power system.

[0150] Example 3

[0151] A computer-readable storage medium stores a computer program / instruction. When the computer program / instruction is executed by a processor, the steps of the following transformer fault identification method are implemented, including:

[0152] Step 1: respectively collect the soundprint signals and vibration signals of the oil-immersed power transformer in normal working mode and multiple fault modes under different working conditions;

[0153] Step 2: The voiceprint signals and vibration signals in normal working modes and various fault modes under different working conditions are respectively made into normal working sample data and fault sample data;

[0154] Step 3: Decompose the normal working sample data and the fault sample data based on the symplectic geometric mode decomposition algorithm to obtain the characteristic components of the normal working samples and the characteristic components of the fault samples;

[0155] Step 4: construct a feature vector of the normal working sample of the transformer according to the feature components of the normal working sample; and construct a feature vector of the fault sample of the transformer according to the feature components of the fault sample;

[0156] Step 5: Import the feature vector of the normal working sample and the feature vector of the fault sample into the cyclic generative adversarial network model for game, and obtain a new feature vector of the fault sample;

[0157] Step 6: Perform a union operation on the feature vector of the fault sample and the feature vector of the new fault sample to obtain the feature vector of the expanded fault sample;

[0158] Step 7: Create a training set and a test set based on the feature vectors of the expanded fault samples and the feature vectors of the normal working samples;

[0159] Step 8: Import the training set into the stack denoising autoencoder model for training to obtain a fault recognition model, and use the test set to test the fault recognition model to obtain an optimized fault recognition model;

[0160] Step 9: Based on the optimized fault identification model, the voiceprint signal and vibration signal of the power transformer are detected in real time to obtain the fault pattern recognition results to identify the internal and external fault points of the transformer and guide the maintenance of the power system.

[0161] In summary, the present invention collects operation samples of oil-immersed power transformers under complex working conditions, and uses the symplectic geometric mode decomposition algorithm to extract characteristic components from the soundprint signals and vibration signals in the operation samples, respectively, to obtain characteristic data of the comprehensive transformer operation state and mechanical state, and combines the cyclic generative adversarial network model CycleGAN to perform cyclic adversarial operations on the characteristic vectors of normal working modes and multiple fault modes, comprehensively expand fault samples, and realize fault mode classification and internal and external fault point identification of transformers under complex conditions, environments and power system load fluctuations;

[0162] The present invention also uses a stack denoising autoencoder model to reduce noise on the oil-immersed power transformer training data and enhance the features of the voiceprint signal and vibration signal, which is conducive to learning more abstract and complex data feature representations, thereby improving the robustness of the invented method, and ultimately achieving the goal of improving the diagnostic accuracy of the fault identification model and enhancing the deep learning capability, and avoiding the shortcomings of information redundancy in the samples generated by the traditional SMOTE algorithm.

[0163] In addition, the amount of sample data calculation expanded by the method of the present invention is smaller than that of the prior art method, thereby reducing the amount of equipment calculation and balancing the load of the computing equipment; at the same time, the fault diagnosis results are accurately and quickly exported, which can provide online, real-time, accurate and effective reference data for the power monitoring system, which can be used to guide the maintenance of the power system, shorten the troubleshooting time and save maintenance costs.

[0164] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0165] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0166] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0167] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.

Claims

1. A transformer fault identification method, characterized in that it includes: The soundprint signals and vibration signals of the oil-immersed power transformer in normal working mode and various fault modes under different working conditions are collected respectively; The soundprint signals and vibration signals in normal working modes and various fault modes under different working conditions are respectively made into normal working sample data and fault sample data; Decomposing the normal working sample data and the fault sample data based on the symplectic geometric mode decomposition algorithm to obtain the characteristic components of the normal working samples and the characteristic components of the fault samples; According to the characteristic components of the normal working samples, a characteristic vector of the normal working samples of the transformer is formed; and according to the characteristic components of the fault samples, a characteristic vector of the fault samples of the transformer is formed; The feature vectors of normal working samples and faulty samples are introduced into the cyclic generative adversarial network model for game playing to obtain new feature vectors of faulty samples; Performing a union operation on the feature vector of the fault sample and the feature vector of the new fault sample to obtain the feature vector of the expanded fault sample; Create a training set and a test set based on the feature vectors of the expanded fault samples and the feature vectors of the normal working samples; The training set is imported into the stack denoising autoencoder model for training to obtain a fault recognition model, and the fault recognition model is evaluated using the test set to obtain the fault recognition model with the best performance; Based on the fault identification model with the best performance, the soundprint signal and vibration signal of the power transformer are detected in real time to obtain the fault pattern recognition results, which are used to identify the internal and external fault points of the transformer and guide the maintenance of the power system.

2. The transformer fault identification method according to claim 1 is characterized in that: The multiple failure modes include: partial discharge, low energy discharge, high energy discharge, low temperature overheating, medium temperature overheating and high temperature overheating; The voiceprint signal is used to detect the internal operating state of the power transformer; The vibration signal is used to provide feedback on the assembly connection status of the transformer components.

3. The transformer fault identification method according to claim 1, characterized in that: Based on the symplectic geometric mode decomposition algorithm, the normal working sample data and the fault sample data are decomposed to obtain the characteristic components of the normal working samples and the characteristic components of the fault samples, including: Signal processing is performed on the normal working sample data and the fault sample data to obtain the normal working sample and the fault sample standardized data. The normal working sample standardized data are the voiceprint signal An i and vibration signal Vn i ; The standardized data of the fault samples are the voiceprint signal Af j and vibration signal Vf j , i is the data volume of normal working samples, j is the data volume of fault samples, and j< <i; The normalized data of the normal working samples and the fault samples are reconstructed in phase space to obtain the trajectory matrix X An , X Vn , X Af and X Vf ; According to the trajectory matrix X An , X Vn , X Af and X Vf Calculate the corresponding covariance symmetric matrix A=X T X, and construct multiple Hamiltonian matrices M from the corresponding covariance symmetric matrix A; The symplectic orthogonal matrix in the symplectic geometric mode decomposition algorithm is used to solve multiple Hamiltonian matrices M respectively to obtain the eigenvalues ​​of normal working sample data and fault sample data. And the eigenvalue The associated eigenvector; The characteristic values ​​of normal working sample data and fault sample data The associated eigenvectors are arranged in descending order to obtain the decomposed eigencomponents of normal working samples and fault samples, which are S-An i 、S-Vn i 、S-Af j and S-Vf j ; Among them, the trajectory matrix X An The expression is: （1） Where d is the embedding dimension, τ is the time delay, An is the voiceprint signal in normal working mode, and m is the trajectory matrix X An The number of rows, N is the signal length, and m=N-(d-1)τ; The expression of Hamiltonian matrix M is: （2） Where A is the covariance symmetric matrix, A T is the transposed matrix of A; The expression of a symplectic orthogonal matrix is: （3） Where Q is a symplectic orthogonal matrix, is the transposed matrix of Q, H is a diagonal matrix, H=M 2 , R is the submatrix after matrix transformation; B is an upper triangular matrix, and its eigenvalues ​​are λ1, λ2, λ3,…, λ d , d is the number of eigenvalues ​​contained in matrix B; and the eigenvalues ​​of the symmetric matrix A are ,in , is any eigenvalue in the matrix B; w=1, 2, …,d.

4. The transformer fault identification method according to claim 1, characterized in that: According to the characteristic components of the normal working samples, a characteristic vector of the normal working samples of the transformer is formed; And according to the characteristic components of the fault sample, the characteristic vector of the transformer fault sample is formed, including: The characteristic component S-An of the normal working sample after decomposition i and S-Vn i , combined into the characteristic vector of the normal working sample of the transformer [S-An i , S-Vn i ]; and the characteristic component S-Af of the decomposed fault sample j and S-Vf j , combined into the characteristic vector of transformer fault samples [S-Af j , S-Vf j ].

5. The transformer fault identification method according to claim 1, characterized in that: The feature vectors of the transformer's normal working samples and the feature vectors of the fault samples are introduced into the cyclic generative adversarial network model for game playing to obtain new feature vectors of the fault samples, including: The feature vector of the normal working sample [S-An i , S-Vn i ] is processed into the data domain X; and the feature vector of the fault sample [S-Af j , S-Vf j ] is processed into data domain Y; Data domain X and data domain Y are based on the generator G, F and discriminator D of the cyclic generative adversarial network model x , D y Play games and obtain new fault sample feature vectors.

6. The transformer fault identification method according to claim 5, characterized in that: Data domain X and data domain Y are based on the generator G, F and discriminator D of the cyclic generative adversarial network model x , D y Play games to obtain new fault sample feature vectors, including: The generator G is used to map the data in the data domain X to the data domain Y to obtain a sample vector G(x) simulating a fault sample; and the generator F is used to map the data in the data domain Y to the data domain X to obtain a sample vector F(y) simulating a normal working sample; Using the discriminator D y Determine the authenticity of the sample vector G(x) of the simulated fault sample. If it is true, output the sample vector G(x) of the simulated fault sample; if it is false, import the sample vector G(x) of the simulated fault sample obtained this time into the generator G and the discriminator D. y The generated adversarial loss function calculates the generator G and the discriminator D y The adversarial loss value is used to update and optimize the generator G and the discriminator D according to the adversarial loss value. y ; Using the discriminator D x Determine the authenticity of the sample vector F(y) that simulates a normal working sample. If it is true, output the sample vector F(y) that simulates a normal working sample. If it is false, import the sample vector F(y) that simulates a normal working sample obtained this time into the generator F and the discriminator D. x The adversarial loss function generated calculates the generator F and the discriminator D x The adversarial loss value is used to update the optimized generator F and discriminator D according to the adversarial loss value. x ; The sample vector G(x) of the simulated fault sample that has been output is mapped to the data domain X by using the generator F, so as to obtain the sample vector F[G(x)] of the simulated normal working sample that is cyclic; and the sample vector F(y) of the simulated normal working sample that has been output is mapped to the data domain Y by using the generator G, so as to obtain the feature vector G[F(y)] of the simulated fault sample that is cyclic; Through the discriminator D x Determine the authenticity of the sample vector F[G(x)] of the cyclic simulated normal working sample. If it is true, output the sample vector F[G(x)] of the cyclic simulated normal working sample. If it is false, import the sample vector F[G(x)] of the cyclic simulated normal working sample obtained this time into the cycle consistency loss function, calculate the cycle consistency loss value, and update the optimization generator F according to the cycle consistency loss value. At the same time, through the discriminator D y Determine the authenticity of the cyclic simulated fault sample feature vector G[F(y)], if it is true, output the cyclic simulated fault sample feature vector G[F(y)], if it is false, import the cyclic simulated fault sample feature vector G[F(y)] obtained this time into the cycle consistency loss function, calculate the cycle consistency loss value, and update the generator G according to the cycle consistency loss value; Repeat the above steps until the generator G, F and the discriminator D x , D y When the total loss function is less than the set threshold or meets the upper limit of the number of iterations, the output cyclic simulated fault sample feature vector G[F(y)] is used for a new fault sample feature vector, and cyclically adversarial data in the data domain X and j data in the data domain Y until (ij) new fault sample feature vectors are generated.

7. The transformer fault identification method according to claim 6, characterized in that: Generator G and Discriminator D y The resulting adversarial loss function is: （4） In the formula, is the adversarial loss function generated by the generator and the discriminator, represents the expectation of data y, The data y in the data domain Y passes through the discriminator D y The output value obtained is represents the expectation of data x, is the natural logarithm function; Generator F and Discriminator D x The resulting adversarial loss function is: （5） The cycle consistency loss function of generator G and F is: （6） In the formula, is the generator cycle consistency loss function, x~P(x) is the data x is the X domain data that obeys the P distribution, y~Q(y) is the data y is the Y domain data that obeys the Q distribution, is 1 norm; The total loss function is: （7） In the formula, θ is the weight parameter.

8. The transformer fault identification method according to claim 1, characterized in that: According to the feature vectors of the expanded fault samples and the feature vectors of the normal working samples, a training set and a test set are prepared, including: The feature vector of the normal working sample [S-An i , S-Vn i ] dataset and the feature vector of the expanded fault sample [S-Af i S-Vf i ] to obtain a sampling sample; The sample is divided into training set and test set in the ratio of 80%: 20%.

9. A transformer fault identification device, characterized in that: include: Acquisition and processing module, feature extraction module, game module, fault identification module and application module; The acquisition and processing module is used to respectively acquire the soundprint signals and vibration signals of the oil-immersed power transformer in normal working mode and multiple fault modes under different working conditions; and to respectively make the soundprint signals and vibration signals in normal working mode and multiple fault modes under different working conditions into normal working sample data and fault sample data; The feature extraction module is used to decompose the normal working sample data and the fault sample data based on the symplectic geometric mode decomposition algorithm to obtain the characteristic components of the normal working samples and the characteristic components of the fault samples; construct the characteristic vector of the normal working samples of the transformer according to the characteristic components of the normal working samples; and construct the characteristic vector of the fault samples of the transformer according to the characteristic components of the fault samples; The game module is used to import the feature vector of the normal working sample and the feature vector of the fault sample into the cyclic generative adversarial network model for game, so as to obtain a new feature vector of the fault sample; and perform a union operation on the feature vector of the fault sample and the new feature vector of the fault sample to obtain the feature vector of the expanded fault sample; The fault identification module is used to prepare a training set and a test set according to the feature vectors of the expanded fault samples and the feature vectors of the normal working samples; and to import the training set into the stack denoising autoencoder model for training to obtain a fault identification model, and to use the test set to test the fault identification model to obtain an optimized fault identification model; The application module is used to perform real-time detection of the soundprint signal and vibration signal of the power transformer based on the optimized fault identification model, and obtain the fault pattern recognition result for identifying the internal and external fault points of the transformer to guide the maintenance of the power system.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the steps of the transformer fault identification method described in any one of claims 1-8 are implemented.

Images