Transformer Fault Diagnosis Method Based on Sparrow Search Algorithm Optimized Hybrid Kernel Extreme Learning Machine

Abstract

The invention discloses a transformer fault diagnosis method based on a sparrow search algorithm optimized hybrid kernel extreme learning machine, including the steps of: 1) collecting data of a transformer operating in six states and preprocessing the data; 2) building a prediction model of a hybrid kernel extreme learning machine; 3) determining the population size, discoverer ratio, warning ratio and objective function parameters in the SSA, and initializing the population; 4) selecting the error rate of the training samples of argmin(TrainErrorRate) as the fitness function of the SSA; 5) transmitting the population state information to the HKELM network to obtain different HKELM networks; 6) training the HKELM network to obtain fitness values; 7) comparing the fitness values and updating the population state; 8) judging whether the update process meets the requirements; 9) outputting the optimal parameters and substituting them into the HKELM network, and diagnosing the transformer fault with the optimized network. While ensuring that the network has good generalization performance, this method can greatly improve the stability and learning speed of the forward neural network.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer fault diagnosis, and particularly relates to a transformer fault identification method based on a hybrid kernel function extreme learning machine. Background Technique

[0002] With the continuous development of China's economy, the demand for electricity by the people has become higher and higher. In this situation, the level and capacity of the power grid have also increased accordingly. Therefore, it is particularly important to develop extra-high voltage transformers. Transformers are an important component of power system power transmission and transformation and are the key to realizing long-distance power transmission. Whether the transformer can operate safely and stably is related to the safety of the entire power system. However, when a transformer is operating, it is often affected by various situations, such as weather and climate conditions. Therefore, some faults are inevitable. In order to ensure the stable operation of the transformer, the fault detection of the transformer is very necessary. If a fault occurs and cannot be detected in time, the consequences may be immeasurable, and even lead to the paralysis of the entire power system. Therefore, the fault diagnosis of transformers has always received close attention.

[0003] The dissolved gas analysis method in oil is an effective method for transformer fault diagnosis, which uses the dissolved gas in the oil to diagnose transformer faults. Although the traditional threshold detection is simple to operate, its division is too absolute and needs to be artificially determined. Therefore, in actual applications, this method has a low accuracy rate and large limitations. In recent years, with the rapid development of artificial intelligence, some new algorithms have gradually begun to be applied, such as artificial neural networks and support vector machines. Although these algorithms have been greatly improved compared with the traditional threshold detection method, there are still many deficiencies. The artificial neural network requires a large amount of data for operation, has a long operation time, and is also very easy to fall into the local optimum situation; while the selection of the kernel function of the support vector machine is difficult. Therefore, further research and improvement are still needed. Summary of the Invention

[0004] On the basis of the above background, the present invention proposes a method for optimizing a hybrid kernel extreme learning machine with a sparrow search algorithm for transformer fault diagnosis. By optimizing the parameters of the hybrid kernel function with the sparrow search algorithm, it can not only solve the problem that the algorithm is easy to fall into the local minimum to improve the network performance, but also improve the diagnostic accuracy of the neural network.

[0005] The object of the present invention can be achieved through the following technical solutions, including the following steps:

[0006] Step 1: Data preprocessing.

[0007] Collect data related to transformer failures. The gases dissolved in transformer oil include five types: hydrogen (H2), methane (CH4), acetylene (C2H2), ethylene (C2H4), and ethane (C2H6).

[0008] Normalize the data and map the data to the range [0, 1]. The specific processing formula is as follows:

[0009]

[0010] In the formula, is the content of the characteristic gas after normalization; x ij is the content of the j-th characteristic gas of the i-th sample; x i1 +x i2 +x i3 +x i4 +x i5 is the total content of the five gases H2, CH4, C2H2, C2H4, and C2H6 of the i-th sample.

[0011] Step 2: Establish a hybrid kernel extreme learning machine prediction model.

[0012] 1) The output of the single-hidden-layer feedforward neural network model is:

[0013] F(x) = h(x) × β = H × β

[0014] In the formula, x is the input data vector, F(x) is the network output; h(x), H are the output functions of the hidden layer nodes; β is the output weight vector connecting the hidden layer and the output layer.

[0015] 2) Transform the network training into a problem of solving a linear system. Introduce the regularization coefficient C and the identity matrix I. Then the output weights between the hidden layer and the output are:

[0016]

[0017] 3) The kernel matrix of the kernel extreme learning machine is:

[0018] Ω ELM = HH T = h(x i )h(x j ) = K(x i , x j )

[0019] In the formula, K(x i , x j ) represents the kernel function. In order to obtain strong learning ability and generalization ability, we use the weighted sum of the radial basis kernel function and the polynomial kernel function as the kernel function of the kernel extreme learning machine. At this time, the hybrid kernel function is:

[0020]

[0021] In the formula, λ, m, n, and σ are all parameters of the mixed kernel function.

[0022] At this time, the output of the hybrid kernel extreme learning machine is:

[0023]

[0024] Step 3: Determine the population size, discovery rate, warning rate, and objective function parameters in SSA, and initialize the population.

[0025] The population size can be determined according to the computer's capabilities; the discovery rate is 20%, and the warning rate is 10%; there are five objective function parameters, namely λ, m, n, σ, and C.

[0026] Step 4: Select the error rate of the training samples as the fitness function of SSA;

[0027] The specific fitness function is:

[0028] fitness = argmin(TrainErrorRate)

[0029] In the formula, TrainErrorRate is the error rate of the training set, and argmin refers to the fitness value when the error rate is the lowest.

[0030] Step 5: Transmit the population state information to the HKELM network to obtain different HKELM networks. By calculation, the discoverer is closer to the target position, that is, the parameters to be optimized are obtained; use the training set error rate as the fitness function to find the lowest individual, that is, to minimize fitness.

[0031] Step 6: Compare the fitness values and update the population state.

[0032] Use the training set error rate as the fitness function to find the lowest individual, that is, to minimize the error rate. [[ID=,38]]

[0033] The specific content of updating the population state is:

[0034] In the population, when a predator is detected, the sparrows at the edge of the population will quickly move to a safe area to obtain a better position, and update the positions of the discoverers, followers, and alarmers according to the update formula. The position update formula for the discoverer is as follows:

[0035]

[0036] In the formula, t represents the number of iterations; iter maxDenotes the maximum number of iterations; X i,j Denotes the position information of the i-th sparrow in the j-th dimension; α ∈ (0, 1] is a random number; R2 (R2 ∈ [-1, 1]) and ST (ST ∈ [0.5, 1]) represent the warning value and the safety value respectively; Q is a random number following a normal distribution; L represents a 1×d matrix where each element in the matrix is 1. When the condition R2 < ST is satisfied, the discoverers can conduct extensive foraging operations, and there are no predators around during foraging at this time. On the contrary, when the condition R2 ≥ ST is satisfied, some sparrows in the group have discovered a predator and send a danger alert to the remaining sparrows, so as to ensure that all sparrows quickly transfer to a safe area for foraging.

[0037] The position update formula for followers is as follows:

[0038]

[0039] In the formula, Denotes the current optimal position occupied by the discoverer; Denotes the current globally worst position; n is the population size; A is a 1×d matrix where each element is randomly assigned 1 or -1, and A + = A T (AA T ) -1 .

[0040] When i > n / 2, this indicates that the i-th joiner with a lower fitness value has not obtained food and is in a very hungry state. At this time, it needs to fly to other places to forage to obtain more energy.

[0041] The position update formula for the alarmers is as follows:

[0042]

[0043] In the formula, Denotes the globally current optimal position; β is a step size control parameter, a random number following a normal distribution with a mean of 0 and a variance of 1; f i Is the fitness value of the current sparrow; f g And f w Are the current globally optimal and worst fitness values; ε is the smallest constant to avoid the current fitness value being the worst fitness value and making the denominator zero; K is a random number in the range [-1, 1], which represents both the movement direction of the sparrow and the step size control parameter.

[0044] In the formula, when f i > f g It means that the sparrow is at the edge of the group and is vulnerable to attacks by predators. It shows that the position of the sparrow is the safest at this time. When f satisfiesi = f g When the condition is met, it indicates that the sparrow in the middle at this time perceives the crisis, and in order to prevent the risk of being captured, it approaches the sparrow in the safe position.

[0045] Update the fitness and position information of the sparrow population according to the above formula.

[0046] Step 7: Determine whether the update requirements are met. If so, execute the next step; if not, return to Step 5. Determine whether the particle update process meets the requirements. The first judgment condition is: determine whether the process of updating the population state in Step 6 reaches the maximum number of iterations. The second is: whether the best fitness value of the population is within the given error range. Meeting either condition means meeting the requirements.

[0047] Step 8: Output the optimal parameters and substitute them into the HKELM network. Bring the divided test set into the trained hybrid kernel extreme learning machine, and the output is the accuracy of fault diagnosis.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] The present invention is a transformer fault diagnosis method based on the sparrow search algorithm to optimize the hybrid kernel extreme learning machine. First, the sample data of the dissolved gas concentration in transformer oil and the corresponding fault conclusions are processed to establish a training set and a test set. Then, the weighted sum of the radial basis kernel function and the polynomial kernel function is used to replace the feature mapping of the hidden layer in ELM to establish a diagnostic model of the hybrid kernel extreme learning machine, which improves the generalization ability and diagnostic accuracy of the model. At the same time, the sparrow optimization algorithm is used to optimize the parameters in the model, which improves the optimization efficiency, speeds up the optimization speed, and avoids falling into local optima. Finally, the test samples are input into the trained kernel extreme learning machine to obtain the transformer fault diagnosis results. The whole process is efficient, concise, fast in processing speed, and high in efficiency, and can obtain higher-precision transformer fault diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is the basic structure diagram of the ELM network based on the sparrow search algorithm to optimize the hybrid kernel extreme learning machine according to the invention;

[0051] Figure 2 It is the flowchart of SSA-HKELM in the hybrid kernel extreme learning machine optimized by the sparrow search algorithm according to the invention;

[0052] Figure 3 It is the fault diagnosis result diagram of ELM in the embodiment of the present invention;

[0053] Figure 4 It is the fault diagnosis result diagram of SSA-HKELM in the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0054] To better illustrate the advantages of the present invention, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, rather than all embodiments. All other embodiments obtained by ordinary technicians in this field based on one or more embodiments of the present disclosure without making any creative efforts are within the scope of protection of the present invention.

[0055] In this paper, we analyzed fault gas data from oil-immersed power transformers in substations to verify the effectiveness of the proposed algorithm model in processing and classifying sample data. We collected 207 sets of DGA data with identified fault types and divided them into training and testing samples. 149 sets of training data were used for parameter optimization, and 58 sets of sample data were used to verify the diagnostic accuracy of the optimized model.

[0056] Reference Figure 3 , the method steps of the present invention:

[0057] Step (1): Collect data of the transformer operating in six states, preprocess the data, remove abnormal data, and then divide it into training set and test set;

[0058] Step (2): Build a hybrid kernel extreme learning machine prediction model;

[0059] Step (3): Determine the population size, the proportion of discoverers, the proportion of early warnings, and the objective function parameters in SSA, and initialize the population;

[0060] Step (4): Select the error rate of the training sample argmin(TrainErrorRate) as the fitness function of SSA;

[0061] Step (5): pass the population status information to the HKELM network, obtain different HKELM networks, and compare whether the parameters to be optimized meet the requirements;

[0062] Step (6): Train the HKELM network to obtain the fitness value;

[0063] Step (7): Compare fitness values and update population status;

[0064] Step (8): Determine whether the update process meets the requirements. If so, proceed to the next step. If not, return to step (5);

[0065] Step (9): Output the optimal parameters and bring them into the HKELM network, and use the optimized network to diagnose transformer faults.

[0066] Process the transformer fault data in the training samples. In step 1, normalize the data. The specific processing formula is as follows:

[0067]

[0068] In the formula, is the content of the characteristic gas after normalization; x ij is the content of the j-th characteristic gas of the i-th sample; x i1 +x i2 +x i3 +x i4 +x i5 is the total content of the five gases H2, CH4, C2H2, C2H4, and C2H6 of the i-th sample.

[0069] Some of the processed data is shown in the following table:

[0070]

[0071] A total of 207 groups of complete fault data were selected in the experiment. Seventy percent of the data was used as the training set, and the remaining thirty percent of the data was used as the test set. The specific distribution of the training set and the test set is shown in the following table:

[0072]

[0073] In this embodiment, a Hybrid Kernel Extreme Learning Machine (HKELM) is adopted. The input sample is X i =[x i1 ,x i2 ,…,x ij T , and the output sample is Y i =[y i1 ,y i2 ,…,y ij T .

[0074] As Figure 1 shown, this figure is the structure diagram of the extreme learning machine, which consists of an input layer, a hidden layer, and an output layer. Its learning objective function is: F(x)=h(x)×β = H×β, where h(x)=g(x) is the output of the hidden layer nodes, that is, h(x)=g(w·x + b), and g(x) is the activation function of the hidden layer neurons. The activation function can be a Sigmoid function, a Gaussian function, etc.

[0075] ​​In order to improve the generalization ability of the model and ensure that the model has good classification ability, in this example, the kernel function is added to the ELM algorithm, that is, the kernel function replaces the feature mapping of the hidden layer in ELM to form the kernel-based ELM algorithm.

[0076] The following are several common kernel functions:

[0077] (1) Gaussian kernel function:

[0078]

[0079] (2) Polynomial kernel function:

[0080] K(x,x i )=[(x,x i )+m] n n=1, 2, …N

[0081] (3) Perceptron kernel function:

[0082] K(x,x i )=tanh(ρ(x·x i )+b)

[0083] (4) Linear kernel function:

[0084] K(x,x i )=x·x i

[0085] In the formula, K(x,x i ) is the kernel function, and σ, m, ρ, b are all constants.

[0086] In order to obtain strong learning ability and generalization ability, in this example, we use the weighted sum of the radial basis kernel function and the polynomial kernel function as the kernel function of the kernel extreme learning machine. The formula for the hybrid kernel function is:

[0087]

[0088] In the formula, λ, m, n, σ are all parameters of the hybrid kernel function.

[0089] Finally, the output of the hybrid kernel extreme learning machine is:

[0090]

[0091] In the formula, K(x i ,x j ) is the kernel function, and C is the regularization coefficient.

[0092] In step (7), the specific content of updating the population state is:

[0093] In the population, when a predator is detected, the sparrows at the edge of the population will quickly move towards a safe area to obtain a better position, and the positions of the discoverers, followers, and alarmers are updated according to the update formula. The position update formula for the discoverers is as follows:

[0094]

[0095] In the formula, t represents the number of iterations; iter max represents the maximum number of iterations; X i,j represents the position information of the i-th sparrow in the j-th dimension; α ∈ (0, 1] is a random number; R2 (R2 ∈ [-1, 1]) and ST (ST ∈ [0.5, 1]) represent the early warning value and the safety value respectively; Q is a random number subject to a normal distribution; L represents a 1×d matrix, where each element in the matrix is 1. When the condition R2 < ST is satisfied, the discoverers can perform extensive foraging operations, and there are no predators around during foraging at this time. On the contrary, when the condition R2 ≥ ST is satisfied, some sparrows in the group have detected a predator and send danger alerts to the remaining sparrows, so as to ensure that all sparrows quickly move to a safe area for foraging.

[0096] The position update formula for the followers is as follows:

[0097]

[0098] In the formula, represents the optimal position occupied by the current discoverer; represents the current globally worst position; n is the population size; A is a 1×d matrix, where each element is randomly assigned 1 or -1, and A + = A T (AA T ) -1 .

[0099] When i > n / 2, this indicates that the i-th joiner with a lower fitness value has not obtained food and is in a very hungry state. At this time, it needs to fly to other places to forage to obtain more energy.

[0100] The position update formula for the alarmers is as follows:

[0101]

[0102] In the formula, represents the globally current optimal position; β is the step size control parameter, a random number subject to a normal distribution with a mean of 0 and a variance of 1; f i is the fitness value of the current sparrow; f g and f ware the current global optimal and worst fitness values; ε is the smallest constant to avoid the current fitness value being the worst fitness value and making the denominator zero; K is a random number ranging from [-1, 1], which represents both the movement direction of the sparrow and the step size control parameter.

[0103] In the formula, when f i > f g It means that the sparrow is at the edge of the group and is vulnerable to attacks by predators. It shows that the position of the sparrow is the safest at this time. When f i = f g When the condition is met, it means that the sparrows in the middle have detected a crisis at this time, so they approach the sparrows in the safe position to prevent the risk of being captured.

[0104] In this embodiment, the SSA parameters are set as follows: the population size N = 20, the discoverer ratio PR = 20%, the scout ratio SD = 10%, and the maximum number of iterations T = 50.

[0105] In this example, the proportion of the training set is 70%, which is used to train the kernel extreme learning machine model optimized by SSA to find the optimal parameters. The remaining data is used as the test set to verify the obtained final training model, and the accuracy of fault diagnosis is judged according to the verification results.

[0106] The diagnosis result of this example is as Figure 4 shown, and the accuracy of fault diagnosis is 91.379%. When directly using ELM for training and learning, the diagnosis result of the model is as Figure 3 shown, and the accuracy is only 84.483%. Comparing the two models, the HKELM model optimized by SSA has a 6.896% higher fault diagnosis accuracy than the ELM model.

[0107] To sum up, in this paper, the polynomial kernel function and the radial basis kernel function are first combined to obtain a mixed kernel function, and then the mixed KELM is constructed using this mixed kernel function. The parameters of the mixed KELM are optimized by the sparrow search algorithm to obtain the improved SSA-HKELM model. Applying the mixed kernel function can improve the learning ability and generalization ability of KELM. Using the SSA algorithm to optimize the parameters of the mixed KELM improves the optimization efficiency, speeds up the optimization speed, avoids falling into local optima, optimizes the network structure of the mixed KELM, and finally improves the accuracy of transformer fault diagnosis.

[0108] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of various equivalent modifications or substitutions within the technical scope disclosed by the present invention, and these modifications or substitutions should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A transformer fault diagnosis method based on the sparrow search algorithm to optimize the hybrid kernel extreme learning machine, characterized in that, It includes the following steps: 1) Collect data on the operation of the transformer in six states. The gases dissolved in the transformer oil include five types: hydrogen (H2), methane (CH4), acetylene (C2H2), ethylene (C2H4), and ethane (C2H6). Normalize the data and divide the processed data into a training set and a test set according to a certain proportion; 2) Build a hybrid kernel extreme learning machine prediction model; 3) Determine the population size, discoverer ratio, scout ratio, and objective function parameters of the SSA, and initialize the population; 4) Select the error rate of the training samples of argmin(TrainErrorRate) as the fitness function of the SSA; 5) Transmit the population state information to the HKELM network to obtain different HKELM networks, and compare whether the parameters to be optimized meet the requirements; 6) Train the HKELM network to obtain the fitness value; 7) Compare the fitness values and update the population state; 8) Judge whether the update process meets the requirements. If it meets, execute the next step. If it does not meet, return to step 5); 9) Output the optimal parameters and substitute them into the HKELM network, and use the optimized network to diagnose transformer faults; The steps for building the hybrid kernel extreme learning machine in step 2) are as follows: The output of the single-hidden-layer feedforward neural network model is: F(x) = h(x)×β = H×β where x is the input data vector, F(x) is the network output; h(x) and H are the output functions of the hidden layer nodes; β is the output weight vector connecting the hidden layer and the output layer; Transform the network training into a problem of solving a linear system. Introduce the regularization coefficient C and the identity matrix I, then the output weight between the hidden layer and the output is: The kernel matrix of the kernel function extreme learning machine is: Ω ELM = HH T = h(x i )h(x j ) = K(x i , x j ) where, K(x i , x j ) represents a kernel function, and the weighted sum of the radial basis kernel function and the polynomial kernel function is used as the kernel function of the kernel extreme learning machine. At this time, the hybrid kernel function is: where λ, m, n, σ are all parameters of the hybrid kernel function; At this time, the output of the hybrid kernel extreme learning machine is: In step 3), the specific contents of determining the population size, discoverer ratio, scout ratio, and objective function parameters in the SSA are: The population size can be determined according to the computer's ability; the discoverer ratio is 20%, and the scout ratio is 10%; there are five objective function parameters, namely λ, m, n, σ, C.

2. The method according to claim 1, characterized in that, The specific content of step 1) is: Collect data related to transformer faults. The gases dissolved in the transformer oil include five types: hydrogen (H2), methane (CH4), acetylene (C2H2), ethylene (C2H4), and ethane (C2H6). Normalize the data and map the data to between [0,1]. The specific processing formula is as follows: In the formula, is the content of the characteristic gas after normalization; x ij is the content of the j-th characteristic gas in the i-th sample; x i1 +x i2 +x i3 +x i4 +x i5 is the total content of the five gases H2, CH4, C2H2, C2H4, and C2H6 in the i-th sample.

3. The method according to claim 1, characterized in that, In step 4), the expression of the selected fitness function is: fitness = argmin(TrainErrorRate) where TrainErrorRate is the error rate of the training set, and argmin refers to the fitness value when the error rate is the lowest.

4. The method according to claim 1, characterized in that In step 6), the specific content of the fitness value obtained by training is: It is found through calculation that the discoverer is closer to the target position, that is, the parameters to be optimized are obtained; use the training set error rate as the fitness function to find the lowest individual, that is, make fitness obtain the minimum value.

5. The method according to claim 1, characterized in that, In step 7), the specific content of updating the population state is as follows: In the population, when a predator is detected, the sparrows at the edge of the population will quickly move to a safe area to obtain a better position. According to the update formula, the positions of the discoverers, followers, and alarmers are updated. The position update formula for the discoverers is as follows: where t represents the number of iterations; iter max represents the maximum number of iterations; X i,j represents the position information of the i-th sparrow in the j-th dimension; α ∈ (0, 1] is a random number; R2 and ST represent the early warning value and the safety value respectively, where R2 ∈ [-1, 1] and ST ∈ [0.5, 1]; Q is a random number that follows a normal distribution; L represents a 1×d matrix, where each element in the matrix is 1; when the condition R2 < ST is met, the discoverers can perform extensive foraging operations, and there is no predator around during foraging at this time; on the contrary, when the condition R2 ≥ ST is met, some sparrows in the group have detected a predator and send a danger alarm to the remaining sparrows, so as to ensure that all sparrows quickly move to a safe area to forage; The position update formula for the followers is as follows: In the formula, represents the optimal position occupied by the current discoverer; represents the current globally worst position; n is the population size; A is a 1×d matrix, where each element is randomly assigned 1 or -1, and A + = A T (AA T ) -1 ; When i > n / 2, this indicates that the i-th joiner with a lower fitness value has not obtained food and is in a very hungry state. At this time, it needs to fly to other places to forage to obtain more energy; The position update formula for the alarmers is as follows: In the formula, represents the globally current optimal position; β is a step size control parameter, a random number following a normal distribution with a mean of 0 and a variance of 1; f i is the fitness value of the current sparrow; f g and f w are the current globally optimal and worst fitness values; ε is the smallest constant to avoid the current fitness value being the worst fitness value and making the denominator zero; K is a random number ranging from [-1, 1], which represents both the movement direction of the sparrows and the step size control parameter; In the formula, when f i > f g it means that the sparrow is at the edge of the group and is vulnerable to attacks by predators, indicating that the position of the sparrow is the safest at this time. When f i = f g the condition is met, it means that the sparrow in the middle has detected a crisis at this time, so it approaches the sparrow in the safe position to prevent the risk of being captured; According to the above formula, the fitness and position information of the sparrow population are updated.

6. The method according to claim 1, wherein The judgment conditions in step 8) include: 1) Whether the process of updating the population state reaches the maximum number of iterations; 2) Determine whether the best fitness value of the population reaches the minimum value; Meeting any of the above conditions meets the requirements.

7. The method according to claim 1, wherein The specific content in step 9) is as follows: Output the optimal parameters and bring them into the HKELM network. Bring the divided test set into the trained hybrid kernel extreme learning machine, and the output is the accuracy rate of fault diagnosis.

Images