Transformer fault diagnosis method

Abstract

The invention provides a transformer fault diagnosis method, which comprises the steps of pairing data of gas dissolved in oil of a transformer in pairs to generate a DGA ratio for transformer fault diagnosis; generating a training data set according to the DGA ratio; accessing a mixed kernel function consisting of a Gaussian kernel function and a polynomial kernel function into a deep mixed learning machine through an automatic encoder, and constructing a deep mixed kernel extreme learning machine; and inputting the training data set into a deep mixed kernel extreme learning machine for training, and generating a transformer fault diagnosis model for transformer fault diagnosis. According to the method, the mixed kernel function and the automatic encoder are introduced, so that the classification precision is improved. In addition, a differential evolution algorithm, a reverse learning strategy and Cauchy variation are integrated into a traditional wild dog algorithm, and an improved wild dog optimization algorithm (IDOA) is formed. By testing a typical test function, it is verified that the IDOA has significant advantages in the aspects of search capability and optimization precision.

Description

Technical Field

[0001] The present application relates to the field of transformer detection, and in particular to a transformer fault diagnosis method. Background Art

[0002] Power transformers are key equipment in power systems, responsible for the rise and fall of voltage. Transformer failures will affect the normal operation of the power system, so improving the accuracy of transformer fault diagnosis is crucial to the stability and reliability of the power system. Transformer dissolved gas analysis (DGA) is a common method for transformer fault diagnosis, but there is a problem of low diagnostic accuracy due to missing coding. In order to improve the accuracy, diagnostic methods based on mathematical models and theories, such as grey correlation theory and fuzzy theory, have been proposed, but the classification accuracy of these methods needs to be improved.

[0003] In recent years, with the development of artificial intelligence, intelligent diagnosis methods with dissolved gas concentration in oil as characteristic quantity have emerged. Among them, artificial neural network and expert system have their own shortcomings. Support vector machine (SVM) is another method, but the selection of its penalty factor and kernel function parameters has a great influence on the classification accuracy.

[0004] The combination of extreme learning machine (ELM) and three-ratio method has improved the diagnostic accuracy to a certain extent, but its learning ability and generalization ability still need to be improved. The kernel extreme learning machine model and deep kernel extreme learning machine prediction model have improved in classification and feature extraction, but the single kernel learning method still has problems in processing high-dimensional complex data. Summary of the invention

[0005] The purpose of the present application is to overcome the above-mentioned problems in the prior art and provide a transformer fault diagnosis method.

[0006] The present application provides a transformer fault diagnosis method, comprising:

[0007] The data of six gases dissolved in transformer oil, H2, CH4, C2H2, C2H4, C2H6 and TH, are paired to generate DGA ratios for transformer fault diagnosis.

[0008] generating a training data set according to the DGA ratio;

[0009] The hybrid kernel function composed of Gaussian kernel function and polynomial kernel function is connected to the deep hybrid learning machine through an autoencoder to construct a deep hybrid kernel extreme learning machine.

[0010] The training data set is input into the deep hybrid kernel extreme learning machine for training, and a transformer fault diagnosis model is generated to perform transformer fault diagnosis.

[0011] Optionally, inputting the training data set into the deep hybrid kernel extreme learning machine for training includes:

[0012] The DOT algorithm is used to optimize the parameters of the deep hybrid kernel extreme learning machine, and the optimization steps include:

[0013] Use Tent chaotic mapping to initialize the parameter population of the DOA algorithm;

[0014] According to the average fitness value of the DOA individuals, an inferior population containing N DOA algorithms is generated, and then iterated, including: generating a mutation vector in each iteration, combining the target vector with the mutation vector through binomial crossover, and then using a greedy algorithm to select individuals to enter the next generation population;

[0015] After the iteration is completed, the output parameters are used as the optimization parameters of the deep hybrid kernel extreme learning machine.

[0016] Optionally, the DOT algorithm is used to optimize the parameters of the deep hybrid kernel extreme learning machine, including:

[0017] The reverse learning strategy and Cauchy mutation are incorporated into the DOT algorithm.

[0018] Optionally, the expression of the hybrid kernel function is as follows:

[0019] K(x,x i )=ν·k RBF +(1-ν)·k poly ,ν∈[0,1];

[0020] Among them, ν is the weight coefficient of the mixing function, which assigns different linear weights to the local kernel function and the global kernel function.

[0021] Optionally, for the deep hybrid kernel extreme learning machine, when the number of training samples is greater than the number of hidden layer nodes, the calculation formula is:

[0022]

[0023] Optionally, the deep hybrid kernel extreme learning machine includes an encoder and a decoder;

[0024] The encoder includes an input layer, a plurality of hidden layers, a kernel mapping layer and a fully connected layer;

[0025] The decoder includes the fully connected layer and an output layer.

[0026] Optionally, the encoder initializes the weights of the hidden layer by compressing the hidden layer and the kernel mapping layer, and stacks the resulting hidden layers according to the weights.

[0027] Optionally, inputting the training data set into the deep hybrid kernel extreme learning machine for training includes:

[0028] The differential evolution algorithm, reverse learning strategy and Cauchy mutation are integrated into the traditional wild dog algorithm to optimize the parameters.

[0029] The beneficial effects of this application are:

[0030] The present application provides a transformer fault diagnosis method, including: pairing the data of six gases H2, CH4, C2H2, C2H4, C2H6 and TH dissolved in the transformer oil in pairs to generate a DGA ratio for transformer fault diagnosis; generating a training data set according to the DGA ratio; connecting a mixed kernel function composed of a Gaussian kernel function and a polynomial kernel function to a deep hybrid learning machine through an autoencoder to construct a deep hybrid kernel extreme learning machine; inputting the training data set into the deep hybrid kernel extreme learning machine for training, generating a transformer fault diagnosis model for transformer fault diagnosis. The present application introduces a mixed kernel function and an autoencoder to improve classification accuracy. In addition, the differential evolution algorithm, the reverse learning strategy and the Cauchy mutation are integrated into the traditional wild dog algorithm to form an improved wild dog optimization algorithm (IDOA). By testing typical test functions, it is verified that IDOA has significant advantages in search capability and optimization accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a transformer fault diagnosis schematic diagram in this application.

[0032] Figure 2 It is a schematic diagram of the eigenvalues ​​and cumulative contribution rates of the correlation coefficient matrix in this application.

[0033] Figure 3 It is a schematic diagram of the DHKELM model in this application.

[0034] Figure 4 It is a schematic diagram of the test function in this application.

[0035] Figure 5 This is a schematic diagram of the optimization comparison of different algorithms in this application. DETAILED DESCRIPTION

[0036] The present application is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present application and implement it.

[0037] The following contents are all examples of specific implementation processes provided for describing in detail the technical solutions to be protected by this application. However, this application may also be implemented in other ways different from the descriptions herein. Persons skilled in the art may, under the guidance of the concepts of this application, adopt different technical means to implement this application. Therefore, this application is not limited to the specific embodiments below.

[0038] like Figure 1 As shown, a transformer fault diagnosis method comprises the following steps:

[0039] S101 pairs the data of six gases, H2, CH4, C2H2, C2H4, C2H6 and TH, dissolved in the transformer oil, to generate a DGA ratio for transformer fault diagnosis.

[0040] Before evaluating the transformer status, it is crucial to select appropriate gas characteristic parameters. In order to obtain more reliable transformer status data, this application considers six gas data: H2, CH4, C2H2, C2H4, C2H6 and total hydrocarbon TH. By pairing these gases in pairs, this application generates 15 DGA ratios that may be used for transformer fault diagnosis, as shown in Table 1. Since the two volume fraction ratios of the same gas have the same distinguishability, this application only lists one of them. Through these gas ratios, this application can more accurately evaluate the status of the transformer.

[0041] Table 1 Dissolved gas ratio in oil

[0042]

[0043] S102 generates a training data set according to the DGA ratio.

[0044] Furthermore, for a given training data set of N (N is 15 in this application) samples:

[0045] T0={(X i ,Y i ),i=1,2,…,N}

[0046] The input feature vector is:

[0047] X i =[X i1 ,X i2 ,...,X in ]∈R n

[0048] The corresponding target vector is:

[0049] t i =[t i1 ,t i2 ,…,tim ]∈R m

[0050] The output vector obtained by the ELM network in this application can be approximated to zero if the number of training samples is equal to the number of neurons in the hidden layer, that is, it satisfies:

[0051]

[0052] The learning process of ELM can be characterized by finding the least squares solution as shown in the formula:

[0053] β=H + T

[0054] Among them, H + is the Moore-Penrose generalized inverse matrix of the hidden layer output matrix H; T is the target output matrix; β is the output weight matrix of the hidden layer.

[0055] S103 connects the hybrid kernel function composed of the Gaussian kernel function and the polynomial kernel function to the deep hybrid learning machine through an autoencoder to construct a deep hybrid kernel extreme learning machine.

[0056] Kernel methods have powerful nonlinear mapping capabilities and are applied to many problems.

[0057] In this application, in order to make ELM obtain stronger learning and generalization capabilities, it is necessary to fuse different types of kernel functions. Introducing the kernel parameter HH + , in order to improve the generalization ability of ELM, the output function of KELM can be expressed as:

[0058]

[0059] Among them, the positive constant C is the penalty parameter and I is the identity matrix.

[0060] The kernel function of KELM (Gaussian kernel function) is defined as:

[0061] HH T =h (xi) h (xj) =k (xi,xj)

[0062] The Gaussian kernel function is a typical local kernel function with better learning ability than the polynomial kernel function. Its expression is:

[0063]

[0064] Where a = 2σ 2 is the exponential parameter of its Gaussian kernel function.

[0065] The polynomial kernel function is used as a global kernel function and is applied to the kernel limit machine. The polynomial kernel function expression is:

[0066] K poly (x,x i )=((x·x i )+q) p

[0067] Among them, q and P are the constant parameter and exponential parameter of the polynomial kernel function respectively.

[0068] The Gaussian kernel function and the polynomial kernel function are independent of each other. The linear weight method is used, and the equation of the new hybrid function is expressed as:

[0069] K(x,x i )=ν·k RBF +(1-ν)·k poly ,ν∈[0,1]

[0070] Among them, ν is the weight coefficient of the mixing function, and different linear weights can be assigned to the local kernel function and the global kernel function. All parameters of the HKELM model will affect its performance. In order to make the model have good learning and generalization capabilities, it is necessary to determine the optimal kernel parameters.

[0071] Furthermore, the deep hybrid kernel extreme learning machine based on the autoencoder in the present application is composed of an input layer, m stacked hidden layers, a kernel mapping layer, a fully connected layer and an output layer.

[0072] The encoder can effectively extract data features by extracting data layer by layer, which is helpful to distinguish easily confused types and improve classification accuracy. The output function of a generalized single hidden layer feedforward neural network with multiple inputs and single outputs can be expressed as:

[0073] f L (x) = H (X) β

[0074] In order to improve the stability of the model, when the number of training samples is greater than the number of hidden layer nodes, the calculation formula is:

[0075]

[0076] like Figure 3 The DHKELM model shown in the figure is composed of a decoder and an encoder. The encoder part initializes the weights of the hidden layer by compressing the hidden layer and the kernel mapping layer. The initialization process is:

[0077] (1) Randomly generate the orthogonal parameter a between the input layer and the hidden layer i and b i.

[0078] (2) Calculate the temporary hidden layer state matrix H t :

[0079]

[0080] and will As a linear activation function g(x).

[0081] (3) Calculate the i+1th implicit weight β i+1 , the calculation formula is:

[0082]

[0083] The resulting hidden layers are stacked up and the output H of the lth layer is i , as the input H of the i+1th layer i +l, its hierarchy is:

[0084] H0=X

[0085] H i+1 =g((β i+1 ) T H i ),i=0,1,…,m-1

[0086] Among them, the expression of the output y of the fully connected layer is:

[0087] Y=g(ω in H m +b in )

[0088] The parameters between the last hidden layer and the output Y of the encoder can be obtained by fine-tuning.

[0089] Finally, the decoder part is set to the traditional neural fully connected layer. The reconstructed data X can be obtained, and the expression is:

[0090]

[0091] The objective function of DHKELM is to minimize the error between the reconstructed data and the input data. The mathematical expression is:

[0092]

[0093] DHKELM needs to use the back-propagation algorithm of the last two layers for training, which greatly reduces the number of parameters that need to be fine-tuned in DHKELM, thereby achieving a significant improvement in efficiency.

[0094] like Figure 2As shown in Figure 1, KPCA is used to reduce the dimension of DGA data, avoiding the problem of linear inseparability of eigenvectors. Therefore, KPCA is used to extract features from 15 eigenvalues.

[0095] like Figure 2 As shown in the figure, the eigenvalues ​​corresponding to the first six principal components are all greater than one, and the cumulative contribution rate has exceeded 95%, which can obtain most of the required feature information, so the dimension of the feature variable after dimensionality reduction is 6.

[0096] S104 inputs the training data set into the deep hybrid kernel extreme learning machine for training, and generates a transformer fault diagnosis model for transformer fault diagnosis.

[0097] In the present application, when training the transformer fault diagnosis model, the population of DOA is initialized using Tent chaotic mapping.

[0098] Due to the randomness, uniformity, and ergodicity of the Tent chaotic sequence, the problem of uneven distribution of the randomly initialized population in the wild dog optimization algorithm was successfully solved. Its mathematical expression is as follows:

[0099]

[0100] Where a is a constant in [0,1]. n+1 >1, then x n+1 The return value is 1; if x n+1 <0, then x n+1 The return value is 0.

[0101] Because the order generated by the Tent chaotic map is poorly distributed at both ends, the Tent chaotic map is improved and its function expression is:

[0102]

[0103] The improved Tent chaotic map has a better ergodicity than before, which is conducive to making the initial population position more evenly distributed in the initial solution space and covering a wider range.

[0104] Furthermore, according to the average fitness value of the wild dogs, an inferior group containing N wild dogs is generated, and each individual consists of an n-dimensional vector.

[0105] Let X i (g) is the i-th individual of the P-th generation, then

[0106]

[0107] Among them, [X min ,X max ] is the search space.

[0108] The value of the jth dimension of the i-th individual in the g-th generation is obtained as follows:

[0109]

[0110] In the formula, i=1,2,…,N; j=1,2,…,n.

[0111] In the gth iteration, the generated mutation vector is:

[0112] V i (g) = X a (g)+F(X b (g)-X c (g))

[0113] Where a, b, c ∈ {1, 2, …, N}; X b (g)-X c (g) is the difference vector, and F is the scaling factor, which is generally selected between [0, 2] to control the impact of the difference vector on the mutation process.

[0114] Furthermore, the target vector and the mutation vector are binomially crossed to ensure the diversity of the population. The expression is as follows:

[0115]

[0116] Among them, the crossover probability R∈[1,2]; j∈[1,2,…,n].

[0117] Furthermore, this application adopts a greedy algorithm to select individuals to enter the next generation population, that is, to generate a new generation of individuals by judging whether the individual is better than the previous one through the fitness function, and the expression is as follows:

[0118]

[0119] Furthermore, integrating reverse learning into DOA can better guide individuals to find the optimal solution. The relevant expression is:

[0120]

[0121]

[0122] in, is the reverse solution of the t-th generation target solution; X new (t+1) is the t+1th generation target solution, u b and l b are the upper and lower bounds, b' is the pseudo information exchange coefficient, and the expression is:

[0123]

[0124] In order to enhance the algorithm's ability to escape from local optimality, the standard Cauchy distribution is introduced into the algorithm to update the target position. The expression is as follows:

[0125] X new (t+1)=C a ·X best (t)

[0126] Integrate reverse learning and Cauchy mutation into DOA by selecting the probability P s To decide which strategy to use to update the target position, the expression is as follows:

[0127]

[0128] Where δ is the adjustment factor. When its value is 0.5, the optimization result is the best.

[0129] In order to ensure that the fitness of the target position obtained by mutation is better than that of the original position, the greedy rule is introduced. When the fitness of the new position is less than that of the original position, the position is updated, otherwise the position is not updated. The update rule is as follows:

[0130]

[0131] Where f(X) represents the fitness value of the x target.

[0132] In this application, the specific selection strategy is: after the DOA algorithm predation strategy ends and the current optimal solution is found, if P s The reverse learning strategy is used to update the target position, otherwise the Cauchy mutation perturbation strategy is selected to update the position; then the fitness is compared to find the wild dog position with the best fitness as the optimal wild dog position.

[0133] like Figure 4 and Figure 5 As shown in the figure, the improved algorithms were tested. It can be seen that in the test process of the two test functions, the five algorithms gradually converged with the increase of the number of iterations, among which IDOA converged the fastest, and reached the optimal fitness at 6 and 4 iterations respectively.

[0134] This application uses kernel principal component analysis (KPCA) to reduce the dimension of gas data and extract effective feature quantities; secondly, the polynomial kernel function is weightedly combined with the Gaussian kernel function to construct a new hybrid kernel function, and the autoencoder is introduced to improve the extreme learning machine to establish the DHKELM model. The reverse learning, Cauchy mutation and differential evolution algorithms are integrated into the wild dog algorithm, and the IDOA performance is tested using two typical test functions, proving that IDOA has stronger stability and optimization ability. IDOA is used to optimize the key parameters of DHKELM and establish the IDOA-DHKELM transformer fault diagnosis model. Finally, the feature quantities extracted by KPCA are used as the input set of the model, and different transformer fault diagnosis models are simulated and verified. The results show that compared with other models, IDOA-DHKELM has higher transformer fault diagnosis accuracy.

Claims

1. A transformer fault diagnosis method, characterized in that: include: The data of six gases dissolved in transformer oil, H2, CH4, C2H2, C2H4, C2H6 and TH, are paired to generate DGA ratios for transformer fault diagnosis. generating a training data set according to the DGA ratio; The hybrid kernel function composed of Gaussian kernel function and polynomial kernel function is connected to the deep hybrid learning machine through an autoencoder to construct a deep hybrid kernel extreme learning machine. The training data set is input into the deep hybrid kernel extreme learning machine for training, and a transformer fault diagnosis model is generated to perform transformer fault diagnosis.

2. The transformer fault diagnosis method according to claim 1, characterized in that: Inputting the training data set into the deep hybrid kernel extreme learning machine for training includes: The DOT algorithm is used to optimize the parameters of the deep hybrid kernel extreme learning machine, and the optimization steps include: Use Tent chaotic mapping to initialize the parameter population of the DOA algorithm; According to the average fitness value of the DOA individuals, an inferior population containing N DOA algorithms is generated, and then iterated, including: generating a mutation vector in each iteration, combining the target vector with the mutation vector through binomial crossover, and then using a greedy algorithm to select individuals to enter the next generation population; After the iteration is completed, the output parameters are used as the optimization parameters of the deep hybrid kernel extreme learning machine.

3. The transformer fault diagnosis method according to claim 2, characterized in that: The DOT algorithm is used to optimize the parameters of the deep hybrid kernel extreme learning machine, including: The reverse learning strategy and Cauchy mutation are incorporated into the DOT algorithm.

4. The transformer fault diagnosis method according to claim 1, characterized in that: The expression of the hybrid kernel function is as follows: K(x,x i )=ν·k RBF +(1-n)·k poly ,ν∈[0,1]; Among them, ν is the weight coefficient of the mixing function, which assigns different linear weights to the local kernel function and the global kernel function.

5. The transformer fault diagnosis method according to claim 1, characterized in that: For the deep hybrid kernel extreme learning machine, when the number of training samples is greater than the number of hidden layer nodes, the calculation formula is:

6. The transformer fault diagnosis method according to claim 1, characterized in that: The deep hybrid kernel extreme learning machine includes an encoder and a decoder; The encoder includes an input layer, a plurality of hidden layers, a kernel mapping layer and a fully connected layer; The decoder includes the fully connected layer and an output layer.

7. The transformer fault diagnosis method according to claim 6, characterized in that: The encoder initializes the weights of the hidden layer by compressing the hidden layer and the kernel mapping layer, and stacks the resulting hidden layers according to the weights.

8. The transformer fault diagnosis method according to claim 1, characterized in that: Inputting the training data set into the deep hybrid kernel extreme learning machine for training includes: The differential evolution algorithm, reverse learning strategy and Cauchy mutation are integrated into the traditional wild dog algorithm to optimize the parameters.

Images