Transformer fault diagnosis method, device and equipment and storage medium

Abstract

The invention discloses a transformer fault diagnosis method, device and equipment and a storage medium, and relates to the technical field of fault diagnosis, and the method comprises the following steps: obtaining historical insulating oil chromatography fault data of a transformer, carrying out the feature extraction of the historical insulating oil chromatography fault data, and obtaining a plurality of groups of feature data corresponding to different fault types; optimizing a kernel parameter and a regularization coefficient in the kernel extreme learning machine KELM model through an improved sparrow search algorithm ISSA to obtain an ISSA-KELM model; the ISSA-KELM model is trained through the multiple sets of feature data and the corresponding fault types, and a fault diagnosis model is obtained; and inputting multiple groups of characteristic data of the insulating oil chromatographic data of the transformer into the fault diagnosis model to obtain a corresponding fault type. Compared with a traditional fault diagnosis method, the method has the advantage that the accuracy is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and particularly to a transformer fault diagnosis method, device, equipment and storage medium. Background Art

[0002] Fault diagnosis of transformer equipment is very important for system safety and stable power supply. Fault diagnosis can accurately identify power network and equipment faults that have occurred or may occur. Transformer insulating oil chromatographic data usually refers to the composition data of dissolved gases in a transformer oil sample analyzed by gas chromatography. These data can help detect whether there are faults or potential equipment problems in the transformer. By analyzing the concentration and ratio of dissolved gases, the type of electrical fault inside the transformer can be reflected, such as overheating, partial discharge or arc discharge, etc. The fault diagnosis method based on dissolved gas data can analyze faults during the operation of the transformer, detect early problems in a timely manner, and track their development to prevent serious accidents. This method makes maintenance shift from a regular mode to condition monitoring, improving the operation and maintenance efficiency of the transformer. Therefore, it is of great significance to study the fault diagnosis method based on dissolved gas data in insulating oil.

[0003] Currently, common fault diagnosis methods mainly include rule-based expert systems, various statistical analysis methods, deep learning algorithms, etc. Although these methods are effective in specific situations, they still have problems such as low fault diagnosis accuracy and insufficient recognition ability. Compared with common fault diagnosis methods, the data processing ability of machine learning algorithms is more powerful, the feature extraction ability is also more efficient than traditional methods, and the non-linear fitting ability of machine learning algorithms is very excellent, which can effectively solve many problems.

[0004] Kernel extreme learning machine (KELM) is a machine learning algorithm proposed by combining extreme learning machine (ELM) and kernel function. KELM not only retains the advantages of ELM, but also can optimize various performances of ELM, such as generalization and robustness. KELM also shows powerful performance when dealing with large-scale data sets.

[0005] The application of kernel extreme learning machine in fault diagnosis is becoming more and more mature, but the selection of its parameters affects the fault diagnosis effect. The sparrow search algorithm is an intelligent optimization algorithm with advantages such as simple parameter setting and fast convergence speed. The sparrow search algorithm can be used to optimize the parameter selection of the kernel extreme learning machine. However, the existing sparrow search algorithm has the disadvantages of low population diversity and easy local convergence in iteration, resulting in low fault diagnosis accuracy of the fault diagnosis method. Summary of the Invention

[0006] Based on the defects existing in the above-mentioned prior art, the present invention provides a transformer fault diagnosis method, device, equipment and storage medium, which solves the problems that the existing sparrow search algorithm has low population diversity and is prone to fall into the optimum locally during iteration, resulting in low fault diagnosis accuracy of the fault diagnosis method.

[0007] The present invention adopts the following technical solutions:

[0008] In the first aspect, the present invention provides a transformer fault diagnosis method, including the following steps:

[0009] Obtain the historical insulation oil chromatogram fault data of the transformer, and perform feature extraction on the historical insulation oil chromatogram fault data to obtain multiple groups of feature data corresponding to different fault types;

[0010] Optimize the kernel parameters and regularization coefficients in the kernel extreme learning machine KELM model through the improved sparrow search algorithm ISSA to obtain the ISSA-KELM model; wherein, the random population initialization strategy in the original sparrow search algorithm is replaced by chaotic mapping, the optimal individual position after population initialization is updated by the whale algorithm, and the position of the predator during the iteration process is updated by the golden sine algorithm to obtain ISSA;

[0011] Use multiple groups of feature data as input and the corresponding fault types as output to train the ISSA-KELM model to obtain a fault diagnosis model;

[0012] Perform feature extraction on the transformer insulation oil chromatogram data collected in real time, and input the extracted multiple groups of feature data into the fault diagnosis model to obtain the corresponding fault types.

[0013] Preferably, the optimization of the kernel parameters and regularization coefficients in the kernel extreme learning machine KELM model through the improved sparrow search algorithm ISSA includes the following steps:

[0014] Take the combination of the kernel parameters and the regularization coefficients as the sparrow individuals of ISSA, and perform multiple iterations on the sparrow individuals;

[0015] Initialize the relevant parameters of ISSA, including the sparrow population size, the maximum number of iterations, the proportion of discoverers, the proportion of predators, and the proportion of perceivers;

[0016] Initialize the sparrow population through chaotic mapping;

[0017] Update the optimal individual positions of the discoverers, predators, and perceivers in the initialized sparrow population through the whale algorithm; recalculate the new fitness value of each sparrow after updating the position, and compare it with the original fitness value; if the new fitness value of the current sparrow is greater than the original fitness value, then take the new position value as the best fitness value;

[0018] Determine whether the maximum number of iterations or the solution accuracy is reached. If so, return the sparrow position information of the best fitness value, which is the optimal combination of kernel parameters and regularization coefficients; otherwise, update the position of the discoverer. If the current population size is less than half of the total number of individuals and the number of iterations is less than one-fifth of the maximum number of iterations, update the position of the predator through the golden sine algorithm.

[0019] Update the position of the perceivers. If the position of an individual exceeds the boundary, adjust it according to the boundary control.

[0020] Output the optimized optimal solution, that is, the optimal combination of kernel parameters and regularization coefficients.

[0021] Preferably, when the current number of iterations is greater than half of the total number of individuals or less than half of the total number of individuals and greater than one-fifth of the maximum number of iterations, update the position of the predator using the predator update formula in the original sparrow search algorithm.

[0022] Preferably, the random population initialization strategy in the original sparrow search algorithm is replaced by chaotic mapping, as follows:

[0023] x k+1 = cos(k cos -1 (x k ), x k ∈[0,1];

[0024] In the formula, x k+1 is the (k + 1)-dimensional recurrence value, and x k is the k-dimensional variable value.

[0025] Preferably, the position of the optimal individual after population initialization is updated by the whale algorithm, as follows:

[0026]

[0027] In the formula, is the position of sparrow i in the j-th dimension at the (t + 1)-th iteration, is the position of sparrow i in the j-th dimension at the t-th iteration, is the current global optimal position, e is the natural constant, b is the shape parameter, and l is a random number in [0,1].

[0028] Preferably, the position of the predator during the iteration process is updated by the golden sine algorithm, as follows:

[0029]

[0030] In the formula, is the position of sparrow i at the (t + 1)-th iteration, is the position of sparrow i at the t-th iteration, is the individual optimal position, x1 and x2 are obtained from the golden section coefficient t, and r1 and r2 are both random numbers.

[0031] Preferably, the multiple groups of the characteristic data include gas component concentrations, gas component ratios, and time series characteristics. Before performing feature extraction on the historical insulating oil chromatogram fault data, it is necessary to map the insulating oil chromatogram fault data at different time points to the same time axis and perform linear interpolation on the mapped insulating oil chromatogram fault data.

[0032] In a second aspect, the present invention provides a transformer fault diagnosis device, including:

[0033] An acquisition module, configured to collect and acquire the historical insulating oil chromatogram fault data of a transformer, and perform feature extraction on the historical insulating oil chromatogram fault data to obtain multiple groups of characteristic data corresponding to different fault types;

[0034] An optimization module, configured to optimize the kernel parameters and regularization coefficients in the kernel extreme learning machine KELM model through the improved sparrow search algorithm ISSA to obtain the ISSA-KELM model; wherein, the chaotic mapping is used to replace the random population initialization strategy in the original sparrow search algorithm, the optimal individual position after population initialization is updated through the whale algorithm, and the position of the predator during the iterative process is updated through the golden sine algorithm to obtain ISSA;

[0035] A training module, configured to train the ISSA-KELM model with multiple groups of characteristic data as inputs and the corresponding fault types as outputs to obtain a fault diagnosis model;

[0036] A diagnosis module, configured to perform feature extraction on the real-time collected insulating oil chromatogram data of the transformer, input the extracted multiple groups of characteristic data into the fault diagnosis model, and obtain the corresponding fault type.

[0037] In a third aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the above-mentioned transformer fault diagnosis method is implemented.

[0038] In a fourth aspect, the present invention provides a computer-readable storage medium. The storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned transformer fault diagnosis method is implemented.

[0039] Compared with the prior art, the above at least one technical solution adopted by the present invention can achieve the following beneficial effects:

[0040] In the present invention, the improved sparrow search algorithm ISSA is used to optimize the kernel parameters and regularization coefficients in the kernel extreme learning machine KELM model to obtain the ISSA-KELM model. First, the chaotic mapping is used to replace the random population initialization strategy in the original sparrow search algorithm, which can enrich the population diversity and improve the effectiveness and flexibility of initialization. Then, the whale algorithm is used to update the position of the optimal individual after population initialization, which can perturb the current optimal solution and improve the ability of the original algorithm to jump out of the local optimum. Finally, the golden sine algorithm is used to update the position of the predator during the iteration process, effectively balancing the ability of global exploitation and local exploration, further increasing the search range of the optimal solution, and avoiding the local optimum being easily trapped in the original algorithm. Compared with the traditional fault diagnosis method, the present invention greatly improves the accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0042] Figure 1 It is a flowchart of a transformer fault diagnosis method of the present invention;

[0043] Figure 2 It is a schematic diagram of the convolutional neural network of the present invention;

[0044] Figure 3 It is a convolution process diagram of an embodiment of the present invention;

[0045] Figure 4 It is a schematic diagram of the fully connected layer structure of the present invention;

[0046] Among them, Figure 4 (a) of : Before adding Dropout, Figure 4 (b) of : After adding Dropout;

[0047] Figure 5 It is a schematic diagram of the improved sparrow search algorithm framework of the present invention;

[0048] Figure 6 It is a schematic diagram of the kernel extreme learning machine model structure of the present invention;

[0049] Figure 7 It is a fitness change curve graph of the present invention;

[0050] Figure 8 It is a schematic diagram of the training result of the ISSA-KELM training set of the present invention;

[0051] Figure 9 Schematic diagram of the diagnostic results of the ISSA-KELM test set of the present invention;

[0052] Figure 10 Schematic diagram of the diagnostic results of the KELM test set of the present invention. Specific implementation manner

[0053] Next, in combination with the accompanying drawings in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0054] 1. Explanation and illustration of the embodiment. In order to enable those skilled in the art to fully understand how the present invention is specifically implemented, this part is an explanatory embodiment that expands and explains the technical solution of the claim.

[0055] To solve the defects existing in the prior art, the present invention proposes a fault diagnosis method based on CNN-ISSA-KELM. First, the original data is preprocessed to remove irrelevant and redundant data, feature extraction is performed through a convolutional neural network, and the feature data is divided into a training set and a test set. Then, the training set is input into the KELM model for training. At the same time, the Chebyshev chaotic map is used to improve the population initialization in the sparrow search algorithm, the whale algorithm is used to improve the position of the optimal individual, and the golden sine algorithm is used for optimization to improve the position of the predator. The improved sparrow search algorithm is used to optimize two important parameters of the KELM model, thereby improving the accuracy of the model, and this model is used to diagnose faults. Finally, the accuracy and effectiveness of the proposed model are verified through test cases.

[0056] To improve the accuracy of the diagnostic model, the present invention first uses a convolutional neural network (CNN) to extract features from the sample data, and then, based on the extreme learning machine (ELM), uses the kernel extreme learning machine (KELM) for optimization, improving the generalization performance and stability of the ELM, and uses the sparrow search algorithm (ISSA) to optimize two important parameters that affect the model performance in the KELM model, namely the kernel parameter γ and the regularization coefficient C.

[0057] The transformer fault diagnosis method based on CNN-ISSA-KELM proposed by the present invention is to extract features from the fault samples and then input them into the KELM model for training. At the same time, the improved sparrow search algorithm is used to optimize two important parameters of the KELM model, thereby improving the accuracy of the model.

[0058] Reference Figure 6 , the kernel extreme learning machine model structure includes an input layer, a hidden layer, and an output layer. Among them, p (p1, p 2… p m ) usually represents input parameters or features. x (x1, x 2… x m ) represents the nodes of the input layer, which receive the input feature p and convert it into a form suitable for network processing. In the figure, x1, x2,..., x m represent the nodes of the input layer. h represents the nodes of the hidden layer. In KELM, these nodes are connected to the input layer through a kernel function to achieve non-linear mapping. t represents the nodes of the output layer, which generate the final prediction result according to the output of the hidden layer. In the figure, t1, t2, t3 represent the nodes of the output layer, and each node corresponds to an output characteristic, such as high-energy discharge, low-energy discharge, and partial discharge.

[0059] Figure 1 The following is a flowchart of a transformer fault diagnosis method provided by an embodiment of the present invention. The following will be combined with Figure 1 to introduce in detail a transformer fault diagnosis method of the present invention, which specifically includes the following steps:

[0060] S1: Collect transformer insulating oil chromatogram fault data. From the IECTC10 database and the fault cases of the Chinese electric power industry standard DL / T722-2014, analyze and sort out the transformer insulating oil chromatogram fault data to form an experimental data set. The experimental data simultaneously includes four DGA data of H2, CH4, C2H6, and C2H4. DGA data refers to the data set obtained by analyzing the insulating oil of a power transformer to obtain its gas components and contents. The data set is divided into a training set and a test set.

[0061] S2: Perform data preprocessing on the transformer insulating oil chromatogram fault data. The data preprocessing method is linear interpolation to achieve unified time series.

[0062] In practice, the chromatographic data of transformer insulating oil are often recorded by different data acquisition systems, which may lead to differences in their sampling time intervals, resulting in deviations in the time series of the data. To solve this problem, the present invention adopts a time series unification algorithm for aligning the time series data of insulating oil. The main purpose of this algorithm is to unify the time axes of each time series for more accurate and efficient comparison and analysis. Its core method is to map the time stamps of each data series to a common time axis to achieve temporal consistency. The linear interpolation method is used to calculate the points to be interpolated, and the missing time points and corresponding oil and gas values are estimated through the existing time series data points, so that the data has a consistent time interval and sampling rate, realizing time series unification.

[0063] S3: Extract features from the preprocessed fault data.

[0064] The Convolutional Neural Network (CNN) model draws on the mechanism of biological neural signal processing, where artificial neurons can transmit signals through connections with surrounding nerve units. Feature extraction is achieved through convolutional operations, which can be divided into different forms such as one-dimensional convolution and two-dimensional convolution, and the specific application depends on the dimension of the data. In the convolutional operation, by adjusting the parameters of the convolutional kernel, features can be extracted from the original data. By changing the weights of the convolutional kernel, various features at different positions can be extracted in the feature map, forming a set of feature matrices. The set of these feature matrices constitutes the output of the convolutional layer. The CNN model mainly consists of three parts: the convolutional layer, the pooling layer, and the fully connected layer. The specific structure is as Figure 2 shown.

[0065] In the Convolutional Neural Network (CNN) model, the convolutional layer is one of the key components, mainly used for feature extraction from the original data. Through the operations of the convolutional layer, the key features in the data can be extracted. The convolutional operation is implemented through the convolutional kernel. The convolutional kernel performs operations based on the set parameters. This process effectively extracts the information in the local area while preserving the spatial relationship. Taking the two-dimensional convolutional network as an example, this convolutional process is as Figure 3 shown. This enables the CNN to capture the spatial and local features in the input data, providing important basic information for the subsequent pooling and fully connected layers.

[0066] When dealing with time series data, the Convolutional Neural Network (CNN) usually adopts one-dimensional convolution, whose main advantage is that it can increase or decrease the dimension by increasing the number of channels of the model without changing the shape of the feature matrix. In addition, one-dimensional convolution can increase the depth of the network, thereby improving the prediction performance. The calculation formula of one-dimensional convolution is as follows:

[0067]

[0068] In the formula, w k is the convolution kernel, and x t-k+1 is the time series.

[0069] The pooling layer screens the features output by the convolutional layer to extract more critical information, and at the same time realizes dimensionality reduction, thereby improving the computational efficiency and reducing the number of model parameters. According to the operation method, the pooling layer is mainly divided into max pooling (MaxPooling) and average pooling (MeanPooling). Max pooling selects the maximum value in the pooling area as the output, while average pooling calculates the average value of the numerical values in the pooling area as the output.

[0070] Figure 4 The structure diagram of Dropout is shown. The fully connected layer is located at the last part of the network structure, used to integrate the features after convolution and pooling, and add an appropriate activation function according to the requirements of the prediction task. Dropout can be introduced in the fully connected layer, and its role is to randomly mask some neurons during the training process, thereby enhancing the non-linear expression ability of the model, improving the robustness, and at the same time optimizing the computational efficiency and accelerating the training speed.

[0071] S4: Build a fault diagnosis model.

[0072] The sparrow search algorithm is inspired by the behavior of sparrow groups in nature. There are different divisions of labor in the group foraging process: discoverers, predators, and perceivers.

[0073] In the sparrow search algorithm, the discoverers are the guides of the entire population. They will explore unknown areas and update their positions. Therefore, when performing iterative calculations on the sparrow population, the position update formula for the discoverers is:

[0074]

[0075] In the formula: is the j-th element value in the population , is the position of sparrow i at the t-th iteration, iter max is the number of iterations, α1 and α2 are uniformly distributed random numbers between (0, 1), R2 (R2 ∈ [0, 1], the random value of a single sparrow individual) and ST (ST ∈ [0.5, 1]) are the range thresholds of danger and safety respectively, and c follows a normal distribution.

[0076] The position update formula for the predators is:

[0077]

[0078] In the formula, and are the optimal and worst positions of the discoverer, Q follows a normal distribution, both A and L are 1×d matrices, the elements of A are all 1 or -1, and the elements of L are all 1; I is the population size, and N is the number of sparrow populations.

[0079] The position update formula for the perceiver is:

[0080]

[0081] where: β is the step size control parameter that follows a normal distribution, K is a random number between -1 and 1, f i , f b and f w are the fitness values, the best fitness value, and the worst fitness value of the i-th sparrow, and ε is a constant infinitely close to 0.

[0082] Its processing process includes:

[0083] Parameter initialization, including the sparrow population size, the maximum number of iterations iter max , the proportion PD of discoverers, the proportion ST of predators, and the proportion SD of perceivers.

[0084] Calculate the fitness values of the sparrows, and find the positions X best and X worst corresponding to the best fitness and the worst fitness respectively.

[0085] Update the positions of the sparrows according to the sparrow search algorithm.

[0086] Recalculate the fitness value after updating the position. If it is higher than the fitness value in the previous iteration, the new position value is used as the best fitness value; otherwise, it remains unchanged.

[0087] Referring to Figure 5 , to make the performance of the sparrow algorithm better, for the population initialization, the optimal individual position, and the follower position of the original sparrow algorithm, the following improvement strategies are added for improvement:

[0088] First, improve the population initialization method. The original sparrow algorithm uses a random initialization method, which may lead to a gradual decrease in population diversity. To overcome this problem, the present invention introduces the Chebyshev chaotic map, which is uniformly distributed and has a wider range on the interval [-1, 1] and has strong stability. Therefore, the Chebyshev chaotic map is used to enrich the population diversity and improve the effectiveness and flexibility of initialization at the same time. Its expression is as follows:

[0089] x k+1 = cos(kcos -1 (x k ))), xk ∈[0,1] (5);

[0090] Secondly, improve the position of the optimal individual. The original sparrow algorithm quickly approaches the global optimal solution at the initial stage of iteration, which easily leads to the population falling into the local optimum. Draw on the behavior of the whale algorithm to simulate whales preying on prey. Among them, the bubble net method approaches and attacks prey through a spiral mechanism. By appropriately adjusting the spiral mechanism in the bubble net method, it is used to perturb the current optimal solution, thereby enhancing the ability of the algorithm to jump out of the local optimum. The specific improvement formula is as follows:

[0091]

[0092] In the formula, b determines the shape of the whale's spiral forward, and setting b = 1 represents an ordinary spiral line, a random number of l ∈ [0,1].

[0093] Finally, improve the position of the predator. The golden sine algorithm optimizes through the sine function and introduces the golden section coefficient, thereby enhancing the local search performance of the algorithm and more effectively balancing the ability of global development and local exploration. Combining the characteristics of the sparrow algorithm, this improvement makes its update of the predator position more accurate. The specific formula is as follows:

[0094]

[0095] In the formula, x1 = a(1 - t)+bt, x2 = at + b(1 - t), generally take a = -π, b = π, x1 and x2 are obtained by the golden section coefficient t; r1 ∈ [0,2π] is the manifestation of the individual's moving distance in the next iteration, r2 ∈ [0,π] is the direction of the next movement, and both r1 and r2 are random numbers; among them is the individual's current position, is the individual's optimal position. If i > n / 2, then update the position according to the original predator formula (3), otherwise update the predator position as follows:

[0096]

[0097] KELM is an improved algorithm proposed by combining the extreme learning machine (ELM) and the kernel function. Among them, ELM is a new type of algorithm based on the feedforward neural network, with fewer parameters to be adjusted and no need for multiple iterations. And KELM not only retains the advantages of ELM, but also can optimize the various performances of ELM, such as generalization and robustness. In dealing with large-scale data sets, KELM also shows strong performance.

[0098] Its formula is:

[0099]

[0100] where: x i , x j are input samples.

[0101] K(x i , x j ) is selected as the RBF kernel parameter:

[0102]

[0103] where γ is the kernel parameter.

[0104] Therefore, the output of the KELM model is:

[0105]

[0106] Among them, the kernel parameter γ and the regularization coefficient C in the KELM model will affect the performance of the model.

[0107] The present invention uses the ISSA algorithm to optimize the two parameters of KELM, specifically including the following steps:

[0108] (1) Take the combination of the kernel parameter and the regularization coefficient as the sparrow individuals of the improved sparrow search algorithm, and initialize the relevant parameters of the improved sparrow algorithm, including the sparrow population size, the maximum number of iterations, the proportion of discoverers, the proportion of predators, and the proportion of perceivers.

[0109] (2) Initialize the sparrow population through chaotic mapping.

[0110] (3) Obtain the fitness value of each sparrow after population initialization, and find the position corresponding to the current best fitness and the position corresponding to the worst fitness. The fitness value reflects the performance of the model. The selected performance index fitness is defined as: 2 - accuracy - accuracy1, where accuracy is the correct rate of the training set and accuracy1 is the correct rate of the test set. The fitness value reflects the performance of the model, and the smaller the better. Ideally, when the correct rates of the training set and the test set are both close to 1, the fitness value is close to 0. In this way, the fitness function can be used to optimize the model parameters and find the parameter combination that maximizes the correct rates of the training and test sets.

[0111] (4) Update the optimal individual positions of the discoverers, predators, and perceivers in the sparrow population through the whale algorithm; recalculate the fitness value of each sparrow after updating the position, and compare it with the fitness value in the previous iteration. If it is higher than the original fitness value, then take the new position value as the best fitness value; otherwise, keep the original fitness value unchanged;

[0112] (5) Determine whether the maximum number of iterations or the solution accuracy is reached. If so, stop the iterative process and return the sparrow position information with the best fitness, which is the optimal (C, γ) combination. Otherwise, update the position of the discoverer according to formula (2), continue the iteration, and check whether the current iteration number is greater than half of the total number of individuals. If so, update the position of the predator using formula (3); otherwise, determine whether the current iteration number exceeds one-fifth of the maximum number of iterations. If so, update the position of the predator using formula (3); otherwise, perform the golden sine algorithm to update the predator;

[0113] (6) Update the position of the aware according to formula (4). If the position of an individual exceeds the boundary, adjust it according to the boundary control steps of the algorithm, and finally output the optimized optimal solution, that is, the optimal combination of kernel parameters and regularization coefficients.

[0114] The present invention uses a convolutional neural network to extract features from fault samples, introduces a kernel extreme learning machine model on the basis of the traditional extreme learning machine, and introduces an improved sparrow search algorithm for parameter optimization, thereby improving the accuracy and precision of fault diagnosis.

[0115] The present invention proposes an improved sparrow search algorithm to optimize the kernel extreme learning machine model. By using the Chebyshev chaotic map to initialize the population, the premature convergence of the algorithm is avoided. Secondly, the whale algorithm is used to perturb and mutate the individual optimal solutions to improve the convergence accuracy of the algorithm and enhance the ability of the algorithm to avoid local optimal problems. Finally, the golden sine algorithm is introduced to update the position of the followers to accelerate the convergence speed of the algorithm.

[0116] II. Evidence of the relevant effects of the embodiments. Some positive effects have been achieved during the research and development or use of the embodiments of the present invention, and there are indeed great advantages compared with the prior art. The following content is described in combination with the data and charts in the test process.

[0117] The present invention conducts an example verification to verify the effectiveness of the proposed fault diagnosis method. The accuracy of the fault diagnosis model based on CNN-ISSA-KELM is illustrated by a transformer fault example. In order to verify the fault diagnosis performance of the proposed method, from the IECTC10 database and the fault cases of the Chinese electric power industry standard DL / T722-2014, the fault data of the transformer and the corresponding normal state and three fault types are analyzed and sorted out: normal state, high-energy discharge, low-energy discharge, and partial discharge, to form an experimental data set. The experimental data simultaneously includes four DGA data of H2, CH4, C2H6, and C2H4. First, the original data is subjected to feature extraction and input into the CNN model to obtain feature data. For the convenience of analysis, as shown in Table 1, the feature data is numbered, and 480 groups of feature data are divided into a training set and a test set, where the training set has 320 groups and the test set has 160 groups.

[0118] Table 1 Distribution of Different Fault Types

[0119] Fault type Category label Sample quantity Normal state 1 120 High-energy discharge 2 120 Low-energy discharge 3 120 Partial discharge 4 120

[0120] The number of iterations of the algorithm is 100, the size of the sparrow population is 20, the proportion of discoverers is 0.7, and the safety threshold of ISSA is set to 0.6.

[0121] For comparative analysis, this test intends to use another extreme learning machine (ELM) as a reference model to compare with the model proposed in the present invention. The feature data is divided into a training set and a test set, and then input into the KELM and ISSA-KELM diagnostic models respectively. The fault diagnosis results of the two models are shown in the figure.

[0122] Figure 7 It is the fitness change curve graph for improving the sparrow search algorithm. From Figure 7 it can be seen that the best fitness value is reached when iterating to the 8th time, and in the ISSA-KELM model, after the optimization of the improved sparrow search algorithm, the regularization coefficient is 0.7597 and the kernel parameter is 44.8236; Figure 8 It shows the training results of inputting the training set of the fault data of the transformer oil chromatogram after feature extraction into the diagnostic model. Figure 9 It shows the results of inputting the test set into the trained model for fault diagnosis, as shown in the figure, where the accuracy rate of the test set is 0.96875; Figure 10 It shows the results of inputting the test set into the trained KELM model for fault diagnosis, and the accuracy rate of the test set is 0.8625.

[0123] The fault diagnosis results of the two models show that the diagnostic accuracy of the KELM fault diagnosis model is relatively low, the effect of fault diagnosis of the transformer is not ideal, and it is easy to be misdiagnosed as other types of faults; compared with the KELM model, the ISSA-KELM fault diagnosis model optimized by the improved sparrow search algorithm has higher accuracy and better stability.

[0124] Based on the same concept, the present invention also provides a transformer fault diagnosis device, including an acquisition module, an optimization module, a training module, and a diagnosis module.

[0125] The acquisition module is used to collect and obtain the historical insulation oil chromatogram fault data of the transformer, and perform feature extraction on the historical insulation oil chromatogram fault data to obtain multiple groups of feature data corresponding to different fault types.

[0126] The optimization module is used to optimize the kernel parameters and regularization coefficients in the kernel extreme learning machine (KELM) model through the improved sparrow search algorithm (ISSA) to obtain the ISSA-KELM model. Among them, the chaotic mapping is used to replace the random population initialization strategy in the original sparrow search algorithm, the optimal individual position after population initialization is updated by the whale algorithm, and the position of the predator in the iterative process is updated by the golden sine algorithm to obtain ISSA.

[0127] The training module is used to train the ISSA-KELM model with multiple groups of feature data as input and the corresponding fault types as output to obtain a fault diagnosis model.

[0128] The diagnosis module is used to extract features from the real-time collected chromatographic data of transformer insulating oil, input the extracted multiple groups of feature data into the fault diagnosis model, and obtain the corresponding fault types.

[0129] The present invention also provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the above-mentioned transformer fault diagnosis method is implemented.

[0130] The present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the above-mentioned transformer fault diagnosis method is implemented.

[0131] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concepts. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications falling within the scope of the present invention.

[0132] Obviously, those skilled in the art can make various changes and deformations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and deformations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention also intends to include these changes and deformations.

Claims

1. A transformer fault diagnosis method, characterized in that: The following steps are involved: Acquire historical insulating oil chromatogram fault data of the transformer, and perform feature extraction on the historical insulating oil chromatogram fault data to obtain multiple groups of feature data corresponding to different fault types; The kernel parameters and regularization coefficients in the kernel extreme learning machine KELM model are optimized by improving the sparrow search algorithm ISSA, and the ISSA-KELM model is obtained. The random population initialization strategy in the original sparrow search algorithm is replaced by chaotic mapping, the optimal individual position after population initialization is updated by the whale algorithm, and the predator position in the iterative process is updated by the golden sine algorithm to obtain the ISSA. With multiple sets of feature data as input and corresponding fault types as output, the ISSA-KELM model is trained to obtain a fault diagnosis model; The transformer insulating oil chromatographic data collected in real time is subjected to feature extraction, and the extracted multiple sets of feature data are input into the fault diagnosis model to obtain the corresponding fault type.

2. A transformer fault diagnosis method according to claim 1, characterized in that: The method of optimizing the kernel parameters and regularization coefficients in the kernel extreme learning machine KELM model by improving the sparrow search algorithm ISSA comprises the following steps: The combination of kernel parameters and regularization coefficients is used as the sparrow individual of ISSA, and multiple iterations are performed on the sparrow individual; Initialize the relevant parameters of ISSA, including the size of the sparrow population, the maximum number of iterations, the proportion of discoverers, the proportion of predators, and the proportion of perceivers; Initialize the sparrow population through chaotic mapping; The optimal individual positions of the discoverers, predators and perceivers in the initialized sparrow population are updated through the whale algorithm; the new fitness value of each sparrow after the updated position is recalculated and compared with the original fitness value; if the new fitness value of the current sparrow is greater than the original fitness value, the new position value is used as the optimal fitness value; Determine whether the maximum number of iterations or solution accuracy has been reached. If so, return the sparrow position information with the best fitness value, which is the optimal combination of kernel parameters and regularization coefficients; otherwise, update the discoverer position. If the current population size is less than half of the total number of individuals and the number of iterations is less than one-fifth of the maximum number of iterations, update the predator position using the golden sine algorithm. Update the perceiver's position. If the individual's position is outside the boundary, adjust it according to the boundary control. Output the optimized optimal solution, that is, the optimal combination of kernel parameters and regularization coefficients.

3. A transformer fault diagnosis method according to claim 2, characterized in that: If the current population size is greater than half of the total number of individuals or the current population size is less than half of the total number of individuals and the number of iterations is greater than one-fifth of the maximum number of iterations, the predator update formula in the original sparrow search algorithm is used to update the predator position.

4. A transformer fault diagnosis method according to claim 1, characterized in that: The random population initialization strategy in the original sparrow search algorithm is replaced by chaotic mapping, as shown below: x k+1 =cos(kcos -1 (x k )),x k ∈[0,1]; In the formula, x k+1 is the recursive value of the k+1th dimension, x k is the k-th dimension variable value.

5. A transformer fault diagnosis method according to claim 1, characterized in that: The optimal individual position after population initialization is updated by the whale algorithm, as shown below: In the formula, is the position of sparrow i in the j dimension at the t+1th iteration, is the position of sparrow i in the jth dimension at the tth iteration, is the current global optimal position, e is a natural constant, b is a shape parameter, and l is a random number in [0,1].

6. A transformer fault diagnosis method according to claim 1, characterized in that: The predator position is updated in the iterative process by the golden sine algorithm as follows: In the formula, is the position of sparrow i at the t+1th iteration, is the position of sparrow i at the tth iteration, is the optimal position of the individual, x1 and x2 are obtained by the golden section coefficient t, and r1 and r2 are random numbers.

7. A transformer fault diagnosis method according to claim 1, characterized in that: The multiple groups of characteristic data include gas component concentration, gas component ratio and time series characteristics. Before extracting the characteristics of the historical insulating oil chromatogram fault data, the insulating oil chromatogram fault data at different time points need to be mapped to the same time axis, and the mapped insulating oil chromatogram fault data need to be linearly interpolated.

8. A transformer fault diagnosis device, characterized in that: include: An acquisition module is used to collect and acquire historical insulating oil chromatogram fault data of the transformer, and perform feature extraction on the historical insulating oil chromatogram fault data to obtain multiple groups of feature data corresponding to different fault types; The optimization module is used to optimize the kernel parameters and regularization coefficients in the kernel extreme learning machine KELM model by using the improved sparrow search algorithm ISSA to obtain the ISSA-KELM model; wherein, the random population initialization strategy in the original sparrow search algorithm is replaced by the chaotic mapping, the optimal individual position after the population initialization is updated by the whale algorithm, and the predator position in the iteration process is updated by the golden sine algorithm to obtain the ISSA; A training module is used to train the ISSA-KELM model with multiple sets of feature data as input and corresponding fault types as output to obtain a fault diagnosis model; The diagnosis module is used to extract features from the transformer insulating oil chromatographic data collected in real time, and input the extracted multiple sets of feature data into the fault diagnosis model to obtain the corresponding fault type.

9. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the transformer fault diagnosis method described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the transformer fault diagnosis method according to any one of claims 1 to 7 is implemented.

Images