Transformer fault diagnosis research method and system

Abstract

The invention discloses a transformer fault diagnosis research method and system, and relates to the technical field of transformer fault diagnosis, and the method comprises the steps: introducing an SHAP value to carry out the feature optimization of transformer fault data based on a decision tree and an integration method thereof; dividing the optimized data into a training set and a test set, and performing data cleaning and oversampling processing on the training set; and optimizing the deep mixed kernel extreme learning machine based on the improved dung beetle algorithm, constructing a transformer fault model, and performing fault diagnosis by using the transformer fault model. According to the feature system provided by the invention, the input features of the transformer fault diagnosis model can be effectively optimized, transformer fault data can be effectively utilized, and the influence of redundant features on the diagnosis model is reduced. Compared with other heuristic artificial intelligence algorithms, the improved dung beetle algorithm provided by the invention has more excellent optimization ability, and can find more excellent parameters of the deep mixed kernel extreme learning machine in the same number of iterations and range.

Description

Technical Field

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

[0002] As an important component of the power transmission and distribution process, the reliable and safe operation of power transformers is of great value to the stability and economy of the entire power system. Establishing an efficient transformer fault diagnosis model can detect and diagnose potential transformer faults in a timely and accurate manner, which helps on-site equipment operation and maintenance personnel to formulate corresponding strategies to prevent transformer faults from becoming more serious.

[0003] At present, oil-immersed transformers are commonly used in power systems. When a fault occurs, the internal transformer insulating oil will crack under high temperature and high pressure environment, generating H2, CH4, C2H6, C2H4, C2H2, CO and CO2 gases dissolved in the transformer. The method of using the above gases to diagnose transformer faults is called dissolved gas analysis (DGA). Traditional oil gas analysis methods (such as three-ratio method or four-ratio method, etc.) rely too much on expert experience and qualitative analysis, and have problems such as strong subjectivity. With the rise and development of artificial intelligence algorithms, such as support vector machines (SVM) and extreme learning machines (ELM), they are widely used in oil dissolved gas analysis. However, SVM is extremely sensitive to the selection of penalty factors and kernel parameters, and incorrect parameters often lead to poor diagnostic results. Since the input weights and biases of ELM are randomly generated, its generalization is limited.

[0004] DHKELM uses multiple extreme learning machine autoencoders (ELM-AE) to build a deep structure to extract transformer fault data layer by layer, enhances the extraction of high-level fault features, nonlinear processing and generalization capabilities, and introduces a hybrid kernel function in the output layer at the end to further divide the fault categories. Compared with the single kernel learning of traditional DKELM, the hybrid kernel function can better capture the diversity and complexity of data by weighted combination of multiple kernel functions, further improving the performance and generalization ability of the model. However, there are a large number of initial parameters in its internal multi-layer network structure and multiple kernel functions, and their selection will affect the accuracy of the model. It is necessary to use an optimization algorithm to further optimize each initial parameter. And because the fault condition of the transformer is only a low-probability event, the majority of normal samples will lead to a high imbalance of the model training data and redundant features in the fault features. These features will increase the complexity of the model and reduce the performance of the fault diagnosis model.

[0005] In view of this, this paper proposes a MIDBO-DHKELM transformer fault diagnosis model based on feature system optimization and NCL-Poly-SMOTE. Summary of the invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the problem to be solved by the present invention is that since transformer failure is only a low-probability event, the majority of normal samples will lead to a high imbalance in the existing model training data and redundant features in the fault features, which will increase the complexity of the model and reduce the performance of the fault diagnosis model. In addition, the existing methods have a large number of initial parameters in the multi-layer network structure and multiple kernel functions, and their selection will affect the accuracy of the model.

[0008] To solve the above technical problems, the present invention provides the following technical solutions: a transformer fault diagnosis research method, which includes, based on a decision tree and its integration method, and introducing a SHAP value to perform feature optimization of transformer fault data; dividing the optimized data into a training set and a test set, and performing data cleaning and oversampling processing on the training set; optimizing a deep hybrid kernel extreme learning machine based on an improved dung beetle algorithm, constructing a transformer fault model, and using the transformer fault model to perform fault diagnosis.

[0009] As a preferred solution of the transformer fault diagnosis research method described in the present invention, the integration method includes three integrated algorithm models: LightGBM, CatBoost, and XGBoost; the introduction of SHAP value for feature optimization of transformer fault data includes, based on the three integrated algorithm models, respectively calculating the |SHAP| mean of the three integrated algorithm models under each feature, and using the root mean square value of the mean of the three integrated algorithm models as a measurement indicator to rank the importance of the feature quantity under the embedded method system, and the formula is expressed as follows:

[0010]

[0011] Among them, |SHAP| represents the absolute value of SHAP, |SHAP| RMS (i), |SHAP LB |(i),|SHAP CA |(i),|SHAP XG |(i) are the root mean square value of the |SHAP| mean of the i-th feature, the |SHAP| mean of LightGBM, the |SHAP| mean of CatBoost, and the |SHAP| mean of XGBoost, respectively.

[0012] As a preferred solution of the transformer fault diagnosis research method described in the present invention, the calculation formula of the SHAP value is: i It is expressed as,

[0013]

[0014] Among them, Z is the subset that does not contain feature i, Q is the full set of features, |Z| represents the number of current features, and s(Z∪{i}) and s(Z) are the outputs under the corresponding features respectively.

[0015] As a preferred solution of the transformer fault diagnosis research method described in the present invention, the data cleaning adopts the neighborhood cleaning rule, including counting the number of samples of each category label in the statistical data set, calculating the average number of samples, and defining the minority class and the majority class. The formula is expressed as follows:

[0016]

[0017] M={l|C class [l]>C average}

[0018] n={l|C class [l]≤C average}

[0019] Among them, C class , C average is the intermediate quantity, y i For sample x i The number of labels, L is the set of all label categories, l is the number of single categories, M is the majority class sample set, and n is the minority class sample set; calculate the Euclidean distance d(x i ,x j ) is expressed as,

[0020]

[0021] in, The samples x are i and x j The value of the cth feature quantity; for each sample x i , find its nearest k neighbor samples, and the neighbor sample points constitute the neighborhood of each sample, expressed as,

[0022] N(x i )={x q ∣x q ∈x top-kneighbors}

[0023] Among them, x q Represents x i A neighboring sample oftop-kneighbors Represents the distance from x i The most recent k samples, N(x i ) represents the set of all nearest samples; for the cleaning of the majority class, the cleaning rule is sample x i Original label y i The classification label result y after voting with the neighborhood samples q If they are different, remove them. The formula is:

[0024]

[0025] For the cleaning of minority classes, the cleaning rule is sample x i The original label is removed only when the classification label result after voting by the neighborhood samples is different and the surrounding neighboring samples are the majority class. The formula is expressed as:

[0026]

[0027] in, and They are all indicator functions, taking the value of 1 when the internal conditions are met and taking the value of 0 when the internal conditions are not met.

[0028] As a preferred solution of the transformer fault diagnosis research method described in the present invention, the oversampling process includes calculating the oversampling ratio of each minority class sample as:

[0029]

[0030] Among them, R over-sampling is the oversampling ratio, N majority and N minority are the number of majority class samples and minority class samples respectively; for minority class samples, polynomial interpolation is used to synthesize new samples according to the oversampling ratio, and the formula is expressed as:

[0031]

[0032] in, is a synthetic sample, x ij For x i The jth adjacent sample of j are the polynomial coefficients, which are random numbers in the range [0,1].

[0033] As a preferred scheme of the transformer fault diagnosis research method described in the present invention, the construction method of the deep hybrid kernel extreme learning machine includes using the cosine kernel-generalized rational quadratic kernel as a hybrid kernel to improve the single kernel function of the kernel extreme learning machine, and using the extreme learning machine autoencoder to greedily train layer by layer to construct a deep structure to establish a deep hybrid kernel extreme learning machine.

[0034] As a preferred solution of the transformer fault diagnosis research method described in the present invention, the improved dung beetle algorithm includes: initializing the population using sinusoidal chaotic mapping, perturbing the position update of the rolling dancing dung beetle using the ant random walk strategy, correcting the position of the egg-laying dung beetle using an adaptive factor, and introducing nonlinear weighted Gaussian-Cauchy mixed mutation to randomly perturb the global optimal individual.

[0035] Another object of the present invention is to provide a transformer fault diagnosis research system, which can construct a transformer fault diagnosis system to diagnose transformer faults.

[0036] To solve the above technical problems, the present invention provides the following technical solutions: a system for a transformer fault diagnosis research method, comprising: a feature extraction module, a data processing module and a model construction module; the feature extraction module is based on a decision tree and its integration method, and introduces a SHAP value to perform feature optimization of transformer fault data; the data processing module divides the optimized data into a training set and a test set, and performs data cleaning and oversampling on the training set; the model construction module optimizes a deep hybrid kernel extreme learning machine based on an improved dung beetle algorithm, constructs a transformer fault model, and uses the transformer fault model to perform fault diagnosis.

[0037] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the transformer fault diagnosis research method as described above when executing the computer program.

[0038] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the transformer fault diagnosis research method as described above.

[0039] The beneficial effects of the present invention are as follows: the feature system proposed in the present invention can effectively optimize the input features of the transformer fault diagnosis model, can effectively utilize the transformer fault data and reduce the influence of redundant features on the diagnosis model.

[0040] The present invention uses NCL-Poly-SMOTE (neighborhood cleaning-polynomial oversampling processing) to clean the training model data - oversampling is beneficial to improve the imbalance of input model data and the influence of boundary samples, thereby further improving the performance of the transformer fault diagnosis model.

[0041] The MIDBO (improved dung beetle algorithm) proposed in the present invention has better optimization ability than other heuristic artificial intelligence algorithms. It can find better parameters of DHKELM (deep hybrid kernel extreme learning machine) within the same number of iterations and range, thereby improving the performance of the diagnostic model.

[0042] The cosine-generalized rational quadratic kernel proposed in the present invention has stronger robustness and adaptability than other kernel functions, and can improve the accuracy of the DHKELM transformer fault diagnosis model. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0044] Figure 1 This is a flow chart of a transformer fault diagnosis research method in Example 1.

[0045] Figure 2 This is a schematic diagram of the preferred structure of a characteristic system of a transformer fault diagnosis research method in Example 1.

[0046] Figure 3 This is a simplified schematic diagram of the DHKELM model training process of a transformer fault diagnosis research method in Example 2.

[0047] Figure 4 This is a diagram showing the ablation test results of a transformer fault diagnosis research method in Example 4.

[0048] Figure 5 This is a module structure diagram of a transformer fault diagnosis research system in Example 5. DETAILED DESCRIPTION

[0049] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0050] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0051] Example 1, reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, and which provides a transformer fault diagnosis research method including: Figure 1 As shown:

[0052] S1. Based on decision tree and its integration method, SHAP value is introduced to optimize the features of transformer fault data.

[0053] Decision trees and their ensemble methods have significant advantages in feature selection. Combined with SHAP values, they can accurately quantify the contribution of each feature to model prediction, thereby ranking the importance of features. This not only helps improve model performance, but also enhances the understanding of model behavior. Especially when it comes to high-dimensional data sets, the application of SHAP feature optimization can effectively reduce feature redundancy, simplify model structure, and improve model operability.

[0054] SHAP value φ i It is a method used to weigh the benefits of each participant in the game, and its formula is expressed as,

[0055]

[0056] Among them, Z is the subset that does not contain feature i, Q is the full set of features, |Z| represents the number of current features, and s(Z∪{i}) and s(Z) are the outputs under the corresponding features respectively.

[0057] It can be combined with common decision trees and their integration methods to traverse the tree structure to decompose and calculate the feature contribution on each node, and then the SHAP mean under the feature can be obtained, which can be used to measure the importance of the feature.

[0058] The structure of the feature optimization system based on decision tree and its ensemble method - |SHAP| is as follows Figure 2 As shown, the present invention adopts the |SHAP| mean of three integrated algorithm models, LightGBM, CatBoost, and XGBoost, and uses the root mean square value of the mean of the three integrated algorithm models as a measurement indicator to rank the importance of the feature quantities under the embedding method system. The formula is expressed as follows:

[0059]

[0060] Among them, |SHAP| represents the absolute value of SHAP, |SHAP| RMS (i), |SHAP LB |(i),|SHAP CA |(i),|SHAP XG |(i) are the root mean square value of the |SHAP| mean of the i-th feature, the |SHAP| mean of LightGBM, the |SHAP| mean of CatBoost, and the |SHAP| mean of XGBoost, respectively.

[0061] S2. Divide the optimized data into a training set and a test set, and perform data cleaning and oversampling on the training set.

[0062] The fault data is divided into training set and test set according to 8:2. Since transformer fault is a low-probability event, the majority of normal samples will lead to imbalanced training samples. At the same time, there may be instrument errors in the process of recording oil chromatographic data, resulting in a small number of errors in a small number of samples. The existence of too many boundary samples will also affect model training. For this reason, the training samples are first cleaned and oversampled before model training. In order to simulate the diagnosis of fault data in actual situations as much as possible, the above processing is not performed on the test set.

[0063] Neighborhood Cleaning Rule (NCL) is a method for dealing with multi-class imbalanced data sets with noise, including:

[0064] Count the number of samples for each category label in the statistical data set, calculate the average number of samples, and define the minority class and the majority class. The formula is expressed as:

[0065]

[0066] M={l|C class [l]>C average}

[0067] n={l|C class [l]≤C average}

[0068] Among them, C class , C average is the intermediate quantity, y i For sample x i The number of labels, L is the set of all label categories, l is the number of single categories, M is the majority class sample set, and n is the minority class sample set.

[0069] Calculate the Euclidean distance d(x) between all samples i ,x j ) is expressed as,

[0070]

[0071] in, The samples x are i and x j The value of the cth feature quantity.

[0072] For each sample x i , find its nearest k neighbor samples, and the neighbor sample points constitute the neighborhood of each sample, expressed as,

[0073] N(x i )={x q ∣x q ∈xtop-kneighbors}

[0074] Among them, x q Represents x i A neighboring sample of top-kneighbors Represents the distance from x i The most recent k samples, N(x i ) represents the set of all nearest samples.

[0075] For the majority class cleaning, the cleaning rule is sample x i Original label y i The classification label result y after voting with the neighborhood samples q If they are different, remove them. The formula is:

[0076]

[0077] For the cleaning of minority classes, the cleaning rule is sample x i The original label is removed only when the classification label result after voting by the neighborhood samples is different and most of the neighboring samples are of the majority class. The formula is:

[0078]

[0079] in, and They are all indicator functions, taking the value of 1 when the internal conditions are met and taking the value of 0 when the internal conditions are not met.

[0080] The oversampling process includes calculating the oversampling ratio of each minority class sample as,

[0081]

[0082] Among them, R over-sampling is the oversampling ratio, N majority and N minority are the number of majority class samples and minority class samples respectively;

[0083] For minority samples, polynomial interpolation is used to synthesize new samples according to the oversampling ratio. The formula is expressed as:

[0084]

[0085] in, is a synthetic sample, x ij For x i The jth adjacent sample of j are the polynomial coefficients, which are random numbers in the range [0,1].

[0086] The number of nearest neighbor samples of NCL-Poly-SMOTE is set to 5, and data cleaning-oversampling is performed on the divided training set to increase the balance of diagnostic model training samples and the applicability of samples to the model.

[0087] S3. Based on the improved dung beetle algorithm, the deep hybrid kernel extreme learning machine is optimized, a transformer fault model is constructed, and the transformer fault model is used for fault diagnosis.

[0088] Example 2, reference Figure 3 , which is the second embodiment of the present invention, and is different from the first embodiment in that: a transformer fault diagnosis research method also includes: the deep hybrid kernel extreme learning machine of the present invention is constructed by using the cosine kernel-generalized rational quadratic kernel as a hybrid kernel to improve the single kernel function of the kernel extreme learning machine, and constructing a deep structure by greedily training the extreme learning machine autoencoder layer by layer, and establishing a deep hybrid kernel extreme learning machine, such as Figure 3 shown.

[0089] First of all, the kernel extreme learning machine is an improved algorithm developed on the basis of the extreme learning machine. Compared with the original extreme learning machine using linear explicit mapping, the kernel extreme learning machine uses the kernel technique to implicitly map the input data to a high-dimensional feature space, and performs the least squares solution in this space. The output weight β in the high-dimensional space can be expressed as:

[0090]

[0091] Among them, H is the hidden layer matrix, Y is the label matrix, C is the regularization parameter, and I is the unit matrix. According to the Mercer condition, the kernel function can be defined as:

[0092] Ω=HH T

[0093] Ω i,j =h(x i )h(x j )=K(x i ,x j )

[0094] Among them, Ω represents the kernel matrix, K(x i ,x j ) is the kernel function.

[0095] For a new input sample x, the model's predicted output is,

[0096]

[0097] After sorting, the output expression can be obtained:

[0098]

[0099] Where: N is the number of training samples.

[0100] Hybrid kernel function: With the increase in the dimension and number of feature samples and the complexity of data structures, traditional single kernel functions can no longer make effective use of feature information. Hybrid kernel functions can combine the advantages of each individual kernel function by combining multiple kernel functions, better handle complex data structures, and improve the generalization of the model. The cosine kernel function can effectively measure the similarity and scale invariance of multi-dimensional feature vectors, while the generalized rational quadratic kernel function has high flexibility, multi-scale capability, smoothness control, nonlinear mapping capability and robustness compared to other traditional kernel functions. A new hybrid kernel function can be obtained by weighted combination of the above two kernel functions. Its formula is expressed as,

[0101]

[0102] K hybrid (x i ,x j )=ω·K COS +(1-ω)·K GRQ

[0103] Among them, x i and x j They represent two different groups of sample points brought into the kernel function respectively. a, b and σ, τ represent the kernel parameters of the cosine kernel function and the generalized rational quadratic kernel function respectively. ω is the weight coefficient, and its value range is [0,1].

[0104] The extreme learning machine autoencoder (ELM-AE) combines the characteristics of the extreme learning machine and the autoencoder. Since ELM-AE is a feedforward neural network model with the same input and output layer nodes, the output of the previous ELM-AE is used as the input of the next ELM-AE to build a deep structure, and then the higher-level and more abstract features of the data can be extracted layer by layer, improving the representation and generalization capabilities of the model. Its basic principles are as follows:

[0105] Using the logic of extreme learning machines, the decoding weight matrix β of the t-th layer ELM-AE is (t) It can be expressed as,

[0106]

[0107] Among them, H + is the generalized inverse matrix of the hidden layer matrix H, X (t)is the input matrix of the ELM-AE, C is the regularization parameter, and I is the unit matrix.

[0108] The next layer of ELM-AE has an input matrix X (t+1) It can be expressed as,

[0109] X (t+1) =X (t) β (t) .

[0110] Embodiment 3 is the third embodiment of the present invention, which is different from the first two embodiments in that: a transformer fault diagnosis research method also includes, the present invention improves the dung beetle algorithm specifically including:

[0111] Initialize the population using a sinusoidal chaotic map:

[0112] Initializing the population with sinusoidal chaotic mapping can effectively improve the performance of the optimization algorithm, increase the global search capability and convergence speed of the algorithm, and maintain the diversity and robustness of the population. The formula is expressed as:

[0113]

[0114] Where a is the chaotic mapping parameter, where a = 2, x k and x k+1 Represent the individual positions of the population at the kth and k+1th times respectively.

[0115] Use the ant random walk strategy to perturb the position update of the rolling ball dancing dung beetle:

[0116] The random walk strategy can help the algorithm escape from the local optimal solution, increase the chance of finding the global optimal solution, and also increase the algorithm's ability to explore in high-dimensional space. Based on this idea, if the iteration curve does not change after 5 iterations, the position of the dung beetle is updated by random walk. The formula is:

[0117]

[0118] Among them, X i (t) represents the position a of the i-th dung beetle at the t-th iteration i and b i are the random walk step set X of the i-th dimension dung beetle i The minimum and maximum values ​​in and are the minimum and maximum values ​​of the random walk variable at the tth iteration respectively.

[0119] Use the adaptive factor to correct the position of the dung beetle egg laying:

[0120] In the early stages of algorithm iteration, a larger step size can expand the search range, improve the algorithm's global search capability, and help the algorithm find a better solution faster and accelerate convergence. However, in the later stages of the algorithm, a smaller step size is beneficial to the algorithm's local search. Therefore, an adaptive factor is used here to constrain the spawning range, and the formula is expressed as,

[0121]

[0122] The revised egg-laying locations of dung beetles are as follows:

[0123] B i (t+1)=υX * +b1×(B i (t)-Lb * )+b2×(B i (t)-Ub * ).

[0124] Introduce nonlinear weighted Gaussian-Cauchy mixed mutation to randomly perturb the global optimal individual:

[0125] By introducing mutation operators into intelligent algorithms, it is possible to avoid falling into local optimality due to assimilation in the later stages of the algorithm. Based on Gauss-Cauchy mutation, this paper introduces nonlinear weights to perturb the global optimal individual, thereby further guiding the optimization of each dung beetle population after iteration. The formula is expressed as,

[0126] X new =X b ×[μ1×Cauchy+μ2×Gaussian],

[0127]

[0128] Among them, X new is an individual that has undergone Gaussian-Cauchy mutation. μ1 and μ2 are the inertia weights of Cauchy and Gaussian respectively. The purpose is to give a larger weight to the Cauchy variable in the early stage and a larger weight to the Gaussian variable in the later stage. Cauchy is a random variable that satisfies the Cauchy distribution, and Gaussian is a random variable that satisfies the Gaussian distribution.

[0129] After disturbing the optimal position, the optimal solution is formulated using the greedy strategy as follows:

[0130]

[0131] Among them, f(X) is the value of the fitness function.

[0132] Example 4, reference Figure 4, which is the third embodiment of the present invention, and is different from the first two embodiments in that: in order to further verify the stability, robustness and optimization performance of the improved algorithm of the present invention, the population numbers of the multi-strategy improved dung beetle optimization algorithm (MIDBO), dung beetle optimization algorithm (DBO), northern goshawk optimization algorithm (NGO), gray wolf optimization algorithm (GWO), and particle swarm optimization algorithm (PSO) are uniformly set to 30, the number of iterations is set to 500, and 6 common basic test functions (their expressions, dimensions, optimization ranges and theoretical optimal values ​​are shown in Table 1) are independently run 30 times, and the relevant running conditions are statistically analyzed for optimization comparison tests. The test results are shown in Table 2.

[0133] Table 1: 6 algorithm test functions

[0134]

[0135]

[0136] Table 2: Test results

[0137]

[0138] The above table compares the proposed multi-strategy improved dung beetle algorithm with other algorithms. The proposed algorithm is much better than other optimization algorithms in terms of convergence accuracy and stability in single-peak, multi-peak and composite functions. In order to further illustrate the actual improvement effect of each improvement strategy, an ablation test was conducted on test function 1 to further illustrate the improvement effect of each strategy. The results are shown in the figure below. Figure 4 As shown in the figure, MIDBO (1-4) is the algorithm model test comparison of the sequential cumulative improvement strategy and the final improvement strategy. The four strategies have improved the performance of MIDBO to a certain extent. And MIDBO4 (using all the strategies in Section 3.22) reaches the optimal value of 0 in the optimal optimization theory after 179 iterations.

[0139] Example 5, reference Figure 5 , which is the third embodiment of the present invention, and is different from the previous two embodiments in that: a system for a transformer fault diagnosis research method includes a feature extraction module 100, a data processing module 200 and a model construction module 300; the feature extraction module 100 is based on a decision tree and its integration method, and introduces a SHAP value to perform feature optimization of transformer fault data; the data processing module 200 divides the optimized data into a training set and a test set, and performs data cleaning and oversampling processing on the training set; the model construction module 300 optimizes a deep hybrid kernel extreme learning machine based on an improved dung beetle algorithm, constructs a transformer fault model, and uses the transformer fault model to perform fault diagnosis.

[0140] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0141] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0142] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0143] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0144] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A transformer fault diagnosis research method, characterized in that: include, Based on decision tree and its integration method, SHAP value is introduced to optimize the features of transformer fault data; The optimized data is divided into a training set and a test set, and the training set is cleaned and oversampled; Based on the improved dung beetle algorithm to optimize the deep hybrid kernel extreme learning machine, a transformer fault model is constructed, and the transformer fault model is used for fault diagnosis.

2. A transformer fault diagnosis research method as claimed in claim 1, characterized in that: The integration method includes three integration algorithm models: LightGBM, CatBoost, and XGBoost; The introduction of SHAP value for transformer fault data feature optimization includes calculating the |SHAP| mean values ​​of the three integrated algorithm models under each feature based on the three integrated algorithm models, and ranking the importance of the feature quantities under the embedding method system using the root mean square value of the mean values ​​of the three integrated algorithm models as a measurement indicator. The formula is expressed as follows: Among them, |SHAP| represents the absolute value of SHAP, |SHAP| RMS (i), |SHAP LB |(i),|SHAP CA |(i),|SHAP XG |(i) are the root mean square value of the |SHAP| mean of the i-th feature, the |SHAP| mean of LightGBM, the |SHAP| mean of CatBoost, and the |SHAP| mean of XGBoost, respectively.

3. A transformer fault diagnosis research method as claimed in claim 2, characterized in that: The calculation formula of the SHAP value φ i It is expressed as, Among them, Z is the subset that does not contain feature i, Q is the full set of features, |Z| represents the number of current features, and s(Z∪{i}) and s(Z) are the outputs under the corresponding features respectively.

4. A transformer fault diagnosis research method as claimed in claim 3, characterized in that: The data cleaning adopts the neighborhood cleaning rule, including counting the number of samples of each category label in the statistical data set, calculating the average number of samples, and defining the minority class and the majority class. The formula is expressed as follows: M={l∣C class [l]>C average } n={l∣C class [l]≤C average } Among them, C class , C average is the intermediate quantity, y i For sample x i The number of labels, L is the set of all label categories, l is the number of single categories, M is the majority class sample set, and n is the minority class sample set; Calculate the Euclidean distance d(x) between all samples i ,x j ) is expressed as, in, The samples x are i and x j The value of the cth characteristic quantity; For each sample x i , find its nearest k neighbor samples, and the neighbor sample points constitute the neighborhood of each sample, expressed as, N(x i )={x q ∣x q ∈x top-kneighbors } Among them, x q Represents x i A neighboring sample of top-kneighbors Represents the distance from x i The most recent k samples, N(x i ) represents the set of all nearest samples; For the majority class cleaning, the cleaning rule is sample x i Original label y i The classification label result y after voting with the neighborhood samples q If they are different, remove them. The formula is: For the cleaning of minority classes, the cleaning rule is sample x i The original label is removed only when the classification label result after voting by the neighborhood samples is different and the surrounding neighboring samples are the majority class. The formula is expressed as: in, and They are all indicator functions, taking the value of 1 when the internal conditions are met and taking the value of 0 when the internal conditions are not met.

5. A transformer fault diagnosis research method as claimed in claim 4, characterized in that: The oversampling process includes calculating the oversampling ratio of each minority class sample as: Among them, R over-sampling is the oversampling ratio, N majority and N minority are the number of majority class samples and minority class samples respectively; For minority samples, polynomial interpolation is used to synthesize new samples according to the oversampling ratio. The formula is expressed as: in, is a synthetic sample, x ij For x i The jth adjacent sample of j are the polynomial coefficients, which are random numbers in the range [0,1].

6. A transformer fault diagnosis research method as claimed in claim 5, characterized in that: The construction method of the deep hybrid kernel extreme learning machine includes using the cosine kernel-generalized rational quadratic kernel as a hybrid kernel to improve the single kernel function of the kernel extreme learning machine, and using the extreme learning machine autoencoder to greedily train layer by layer to construct a deep structure to establish a deep hybrid kernel extreme learning machine.

7. A transformer fault diagnosis research method as claimed in claim 6, characterized in that: The improved dung beetle algorithm includes: using sinusoidal chaotic mapping to initialize the population, using ant random walk strategy to perturb the position update of the rolling ball dancing dung beetle, using adaptive factors to correct the position of the egg-laying dung beetle, and introducing nonlinear weighted Gaussian-Cauchy mixed mutation to randomly perturb the global optimal individual.

8. A system using a transformer fault diagnosis research method as claimed in any one of claims 1 to 7, characterized in that: It includes a feature extraction module (100), a data processing module (200) and a model building module (300); The feature extraction module (100) is based on a decision tree and an integration method thereof, and introduces a SHAP value to perform feature optimization of transformer fault data; The data processing module (200) divides the optimized data into a training set and a test set, and performs data cleaning and oversampling processing on the training set; The model building module (300) optimizes the deep hybrid kernel extreme learning machine based on the improved dung beetle algorithm, builds a transformer fault model, and uses the transformer fault model to perform fault diagnosis.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of a transformer fault diagnosis research method described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a transformer fault diagnosis research method described in any one of claims 1 to 7 are implemented.

Images