Hybrid semi-supervised transformer fault diagnosis method

Abstract

The invention provides a hybrid semi-supervised transformer fault diagnosis method. Comprising the steps of establishing a semi-supervised iterative transformer fault diagnosis model combined with filtering type cross validation feature optimization of an XYZ strategy, establishing a transformer fault diagnosis model based on wrapping type feature optimization, and establishing a transformer fault diagnosis model based on hybrid semi-supervised feature optimization. According to the method, the feature complexity is reduced by combining the advantages of filtering type and package type feature optimization, and compared with a single feature selection method, the distinguishable capability of the features is enhanced; a small amount of DGA marked data is used as guidance, a large amount of DGA unmarked data is used for improving learning performance, and the problem of transformer fault diagnosis under the condition that fault marked samples are scarce is solved.

Description

Technical Field The invention relates to the technical field of transformer fault diagnosis, and in particular to a hybrid semi-supervisory transformer fault diagnosis method. Background Art Dissolved Gas Analysis (DGA) is the most commonly used method for on-site power transformer fault diagnosis. In recent years, various DGA-based transformer fault diagnosis criteria proposed at home and abroad are based on on-site experience. They are often misjudged or missed due to incomplete coding and overly absolute boundaries, which brings many problems to maintenance work. In recent years, with the extensive development of artificial intelligence technology in various fields, artificial intelligence technologies including support vector machines, decision trees, neural networks, clustering, fuzzy logic reasoning, etc. have been widely used in transformer fault diagnosis and have achieved good results. However, a large amount of DGA data with labeled fault types is required to build models. Models built with too little labeled data often lead to poor generalization performance of learners and reduced fault diagnosis rates.

[0003] In fact, the reliable labeled fault samples available on site are very limited. Nevertheless, unlabeled DGA data on site is easy to obtain. The semi-supervised classification method can use a small amount of DGA labeled data as a guide and utilize a large amount of DGA unlabeled data to improve learning performance and classification accuracy. It is a powerful mathematical model to solve the transformer fault diagnosis problem when DGA fault samples are scarce.

[0004] Most existing semi-supervised classification techniques for transformer fault diagnosis use features such as specific types of gas concentrations, various gas ratios or relative percentages, which results in highly complex features that seriously affect the performance of the initial semi-supervised learning model, causing errors caused by feature redundancy to accumulate during the semi-supervised learning iteration process, making it difficult to improve the classifier performance. Very few semi-supervised transformer fault diagnosis techniques use a single metric or the same type of feature selection method to evaluate the distinguishability of each feature, but this easily leads to suboptimal features. Summary of the invention The present invention provides a hybrid semi-supervised transformer fault diagnosis method to overcome the defect that the existing semi-supervised feature selection method for transformer fault diagnosis cannot fairly and impartially utilize all information to evaluate the discrimination ability of features.

[0006] This specification discloses a hybrid semi-supervised transformer fault diagnosis method, comprising: S1. Establish a semi-supervised iterative transformer fault diagnosis model based on filtered cross-validation feature optimization, as follows: Get DGA data and divide it into training set S T and validation set S V , training set ST Including the initial labeled sample set S I and the unlabeled sample set S U ; Using the initial labeled sample set S I Train the initial classifier and establish a transformer fault diagnosis multi-classification model based on support vector machine; The transformer fault diagnosis multi-classification model based on support vector machine is used to predict the unlabeled sample set S U and validation set S V Based on the one-to-many calculation of the transformer fault diagnosis multi-classification model based on support vector machine in the unlabeled sample set S U Multiple posterior probabilities on each sample, and then average the posterior probabilities of each sample; S2. Establish a semi-supervised iterative model for filtering cross-validation feature selection combined with the XYZ strategy, as follows: In the m In the iteration, the unlabeled sample set S is selected using the X strategy U The samples with high confidence are combined with the labeled sample set S I Integrate and reconstruct the labeled training set S L ( m ); Using the labeled training set S L ( m ) Retrain the multi-classifier of the support vector machine and re-predict the unlabeled sample set S U and validation set S V The fault category; Recalculate the unlabeled sample set S U The posterior probability of each sample in ; Find the m The iteration and m -1 iteration for the unlabeled sample set S U The predicted labels are calculated, and the number of different labels before and after is calculated. If the number of different labels is less than 1, the iteration stops; if the number is greater than 1, the iteration is repeated. m +1 iteration until the m The upper limit value of After the iteration stops, the validation set S is obtained V The final classification rate; According to the new training set S L ( m )’s feature space and the corresponding final label, sort the feature space using a filtering method, and select the optimal feature using the Y strategy; go through r * K Fold cross validation, calculate the validation set S VThe average classification rate on the graph, and the Z strategy is used to select the optimal features; S3. Establish a transformer fault diagnosis model based on wrap-around feature optimization, as follows: Establish a feature optimization model combining genetic algorithm with support vector machine, and simultaneously optimize the parameters of support vector machine and the input DGA features; Set genetic algorithm parameters, including population, number of iterations, and chromosome length; Encode the support vector machine parameters and input DGA features, perform random initial population, chromosome decoding, individual fitness ranking, genetic operations, generate offspring individuals, offspring chromosome decoding, and offspring chromosome individual fitness calculation; When the stopping condition is reached, the optimal feature space, optimal support vector machine parameters and cross-validation rate are output; S4. Establish a hybrid semi-supervised feature optimization transformer fault diagnosis model, as follows: Construction of feature space: Select specific types of gas concentrations, various gas ratios and relative percentages recommended by IEC or IEEE as the initial feature space; Reconstruct the original DGA data, based on the data set formed by the initial feature space, and use the synthetic minority class oversampling technology to change the imbalance of the data; Dataset preprocessing: Perform arctangent transformation and normalization on the feature space of the dataset with balanced data distribution; A semi-supervised iterative model based on filtered cross-validation feature optimization is established, and the final classification rate is used as the value of individual fitness in the transformer fault diagnosis model based on parcel feature optimization. The chromosomes are re-encoded using the obtained optimal feature space. Among them, the X strategy is: if the sample set S is not labeled U The posterior probability of the sample prediction classification in is greater than that of the unlabeled sample set S U The average posterior probability of all sample prediction classifications in the training set S is considered to be credible, and the credible samples are added to the training set S T Reconstruct the labeled training set S L ( m ); The Y strategy is: select the feature space of each proportion according to the sorting, based on S L ( m ) for training and on the validation set S V Verify on the above, select the ratio corresponding to the maximum classification rate, and replace the encoding with the new feature space; The Z strategy is: r * K The optimal feature space matrix obtained by cross-validation is selected, and the feature space corresponding to the smallest difference with the average classification rate is selected as the optimal feature.

[0007] In summary, the present invention has at least the following beneficial effects: The present invention combines the advantages of filtering and wrapping feature selection to reduce feature complexity and enhance the distinguishability of features compared with a single feature selection method. It uses a small amount of DGA labeled data as a guide and a large amount of DGA unlabeled data to improve learning performance, solving the problem of transformer fault diagnosis when fault labeled samples are scarce. BRIEF DESCRIPTION OF THE DRAWINGS In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0009] Figure 1 It is a flow chart of the hybrid semi-supervised transformer fault diagnosis method involved in the present invention. DETAILED DESCRIPTION In the following, only some exemplary embodiments are briefly described. As those skilled in the art will appreciate, the described embodiments may be modified in various ways without departing from the spirit or scope of the embodiments of the present invention. Therefore, the drawings and descriptions are considered to be exemplary and non-restrictive in nature.

[0011] The disclosure below provides many different embodiments or examples to implement different structures of the embodiments of the present invention. In order to simplify the disclosure of the embodiments of the present invention, the parts and settings of specific examples are described below. Of course, they are only examples, and the purpose is not to limit the embodiments of the present invention. In addition, the embodiments of the present invention can repeat reference numbers and / or reference letters in different examples, and this repetition is for the purpose of simplification and clarity, which itself does not indicate the relationship between the various embodiments and / or settings discussed.

[0012] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0013] like Figure 1 As shown, this embodiment provides a hybrid semi-supervisory transformer fault diagnosis method, including: S1. Establish a semi-supervised iterative transformer fault diagnosis model based on filtered cross-validation feature optimization, as follows: Get DGA data and divide it into training set S T and validation set S V , training set S T Including the initial labeled sample set S I and the unlabeled sample set SU ; Using the initial labeled sample set S I Train the initial classifier and establish a transformer fault diagnosis multi-classification model based on support vector machine; The transformer fault diagnosis multi-classification model based on support vector machine is used to predict the unlabeled sample set S U and validation set S V Based on the one-to-many calculation of the transformer fault diagnosis multi-classification model based on support vector machine in the unlabeled sample set S U Multiple posterior probabilities on each sample, and then average the posterior probabilities of each sample; S2. Establish a semi-supervised iterative model for filtering cross-validation feature selection combined with the XYZ strategy, as follows: In the m In the iteration, the unlabeled sample set S is selected using the X strategy U The samples with high confidence are combined with the labeled sample set S I Integrate and reconstruct the labeled training set S L ( m ); Using the labeled training set S L ( m ) Retrain the multi-classifier of the support vector machine and re-predict the unlabeled sample set S U and validation set S V The fault category; Recalculate the unlabeled sample set S U The posterior probability of each sample in ; Find the m The iteration and m -1 iteration for the unlabeled sample set S U The predicted labels are calculated, and the number of different labels before and after is calculated. If the number of different labels is less than 1, the iteration stops; if the number is greater than 1, the iteration is repeated. m +1 iteration until the m The upper limit value of After the iteration stops, the validation set S is obtained V The final classification rate; According to the new training set S L ( m )’s feature space and the corresponding final label, sort the feature space using a filtering method, and select the optimal feature using the Y strategy; go through r * K Fold cross validation, calculate the validation set S V The average classification rate on the graph, and the Z strategy is used to select the optimal features; S3. Establish a transformer fault diagnosis model based on wrap-around feature optimization, as follows: Establish a feature optimization model combining genetic algorithm with support vector machine, and simultaneously optimize the parameters of support vector machine and the input DGA features; Set genetic algorithm parameters, including population, number of iterations, and chromosome length; Encode the support vector machine parameters and input DGA features, perform random initial population, chromosome decoding, individual fitness ranking, genetic operations, generate offspring individuals, offspring chromosome decoding, and offspring chromosome individual fitness calculation; When the stopping condition is reached, the optimal feature space, optimal support vector machine parameters and cross-validation rate are output; S4. Establish a hybrid semi-supervised feature optimization transformer fault diagnosis model, as follows: Construction of feature space: Select specific types of gas concentrations, various gas ratios and relative percentages recommended by IEC or IEEE as the initial feature space; Reconstruct the original DGA data, based on the data set formed by the initial feature space, and use the synthetic minority class oversampling technology to change the imbalance of the data; Dataset preprocessing: Perform arctangent transformation and normalization on the feature space of the dataset with balanced data distribution; A semi-supervised iterative model based on filtered cross-validation feature optimization is established, and the final classification rate is used as the value of individual fitness in the transformer fault diagnosis model based on parcel feature optimization. The chromosomes are re-encoded using the obtained optimal feature space. Among them, the X strategy is: if the sample set S is not labeled U The posterior probability of the sample prediction classification in is greater than that of the unlabeled sample set S U The average posterior probability of all sample prediction classifications in the training set S is considered to be credible, and the credible samples are added to the training set S T Reconstruct the labeled training set S L ( m ); The Y strategy is: select the feature space of each proportion according to the sorting, based on S L ( m ) for training and on the validation set S V Verify on the above, select the ratio corresponding to the maximum classification rate, and replace the encoding with the new feature space; The Z strategy is: r * K The optimal feature space matrix obtained by cross-validation is selected, and the feature space corresponding to the smallest difference with the average classification rate is selected as the optimal feature.

[0014] The technical concept of the present invention is as follows: 1) Establish a semi-supervised iterative transformer fault diagnosis model (Filter Cross-validation Method based Semi-supervised, FCM-SS) based on XYZ strategy Initialization: (Steps 1-3 describe the initialization process) Step 1: Divide the data set into training set S T and validation set S V , the training set is divided into the initial labeled sample set S I and the unlabeled sample set S U , the number of labeled sample sets is much smaller than that of unlabeled sample sets.

[0015] Step 2: Use the initial labeled sample set to train the initial classifier and establish a transformer fault diagnosis multi-classification model based on support vector machine (SVM).

[0016] Step 3. Use the SVM-based transformer fault diagnosis multi-classification model to predict the unlabeled sample set S U and validation set S V Based on the fault category of one-against-all (OAA), multi-classification SVM is calculated on the unlabeled sample set S U Multiple posterior probabilities on each sample, and then average the posterior probabilities of each sample.

[0017] Iterative process (steps 4-10 describe the mth iteration (m=1,2,...m0)) Step 4 In m In the iteration, use the X strategy to select S U The samples with high confidence are combined with the labeled sample set S I Integrate and reconstruct the labeled training set S L ( m ).

[0018] Step 5: Use the labeled training set S L ( m ) Retrain the SVM multi-classifier and re-predict the unlabeled sample set S U and validation set S V The fault category.

[0019] Step 6 Recalculate S U The posterior probability of each sample in .

[0020] Step 7 Find the m The iteration and m-1 iteration predicts the label of the unlabeled sample SU, and calculates the number of different labels before and after. If the number of different labels is less than 1, the iteration stops. If the number is greater than 1, return to step 4 for the m +1 iteration until the m The upper limit value m0.

[0021] Step 8 After the iteration stops, obtain the validation set S V The final classification rate.

[0022] Step 9: Based on the new training set S L ( m )’s feature space and the corresponding final labels, sort the feature space using a filtering method, and select the optimal features using the Y strategy.

[0023] Step 10 r * K The average classification rate on SV was calculated using the Z strategy to select the optimal features.

[0024] 2) Establish a transformer fault diagnosis model based on package feature optimization Step 11: Establish a feature optimization model combining a genetic algorithm with a support vector machine, and simultaneously optimize the parameters of the support vector machine and the input DGA features.

[0025] Step 12: Genetic algorithm parameter settings, including population, number of iterations, chromosome length, etc.

[0026] Step 13: Encode the SVM parameters and input DGA features, randomize the initial population, decode the chromosomes, sort the individual fitness, perform genetic operations, generate offspring individuals, decode the offspring chromosomes, and calculate the fitness of the offspring chromosome individuals.

[0027] 3) Establish a hybrid semi-supervised feature selection transformer fault diagnosis model (Hybrid Feature Selection based Semi-supervised, HFS-SS) Step 14: Construction of feature space: Select the concentration of a specific type of gas, various gas ratios, and relative percentages recommended by IEC or IEEE (see Table 2) as the initial feature space.

[0028] Step 15: Reconstruct the original DGA data. Based on the data set formed by the initial feature space, use the synthetic minority oversampling technique (SMOTE) to change the imbalance of the data.

[0029] Step 16: Data set preprocessing: perform arctangent transformation (AT) and normalization on the reconstructed DGA data.

[0030] Step 17, establish a semi-supervised iterative model based on filtered cross-validation feature optimization, and use the final classification rate obtained as the value of individual fitness in each iteration of the transformer fault diagnosis method based on parcel feature optimization. Use the feature space obtained in each iteration of the genetic algorithm to re-encode the chromosome.

[0031] Step 18: When the preset number of iterations of the genetic algorithm is reached, the optimal feature space, optimal SVM parameters and cross-validation rate are output.

[0032] In a specific embodiment, the present invention uses the IEC TC10 public data set of transformers to establish a hybrid semi-supervised feature-optimized transformer fault diagnosis model (HFS-SS) and train the algorithm to select the optimal feature subset. The IECTC10 public data set contains a total of 167 samples, including six types of faults: low energy discharge (LED), high energy discharge (HED), medium and low temperature overheating (LMT), high temperature overheating (HT), partial discharge (PD) and normal condition (NC). The types of dissolved gases in the oil of each sample are measured to be H2, CH4, C2H2, C2H4, C2H6, CO and CO2. Due to the small amount of partial discharge data, it is not used in model establishment and algorithm training. The final number of samples and fault types are shown in Table 1.

[0033] Table 1. Transformer DGA sample quantity and category

[0034] Feature space construction. Looking at domestic and foreign research, related work focuses on the existing DGA ratios. There are obvious differences in the DGA ratios used in diagnostic models in different literatures, and a unified standard for gas ratios used in diagnostic models has not yet been formed. The present invention analyzes the relationship between DGA-related features and faults, and mines the original gas combination ratio features that are more relevant to transformer DGA fault diagnosis. It is found from the literature that the ratio features related to CO and CO2 have a weak correlation in transformer fault diagnosis. Therefore, the original features constructed by the present invention exclude the ratio features related to CO and CO2, and select features with strong correlation with transformer fault diagnosis, with a total of 34 dimensions, as shown in Table 2. Despite this, there are still redundant features among these features, which need to be optimized using the feature selection algorithm of the present invention, as shown in Table 2.

[0035] Table 2. Characteristics of various ratios

[0036] The 34-dimensional DGA dataset is preprocessed, including using the synthetic minority oversampling technology SMOTE to change the imbalance of the data. The number of balanced samples is shown in Table 1. The inverse tangent transformation AT is used to process the case where the denominator or numerator of the feature ratio corresponding to the balanced data is 0, and normalization is performed. The balanced data includes 249 samples.

[0037] ; ; in x ij and x' ij Respectively represent i The samples before and after normalization j eigenvalues, x j max and x j min Respectively represent all samples j The maximum and minimum values ​​of the features.

[0038] Transformer fault diagnosis using a hybrid semi-supervised feature optimization algorithm 1) First, the genetic algorithm parameters are set. The population size and maximum number of iterations are set to 50 and 200 respectively, the gene length is set to 34, 18, 18, the range of the support vector machine parameters C and γ is [0, 200], and the number of iterations is 200. Then the chromosome is encoded. Encoding process: The chromosome of the genetic algorithm is generated in binary encoding. The chromosome is divided into three segments of genes. The first segment represents the feature space of DGA. The length is consistent with the initial feature dimension. 1 indicates that the feature is selected and 0 indicates that the feature is not selected. The second and third segments represent two parameters of SVM, the penalty factor C and kernel parameters c , which consists of a binary surface and a decimal value C and c The value of .

[0039] 2) Calculation of chromosome individual fitness: using a semi-supervised iterative model with filtered cross-validation feature optimization to calculate individual fitness.

[0040] The preprocessed data set is divided into a training set S consisting of 199 samples according to a 4*5-fold cross validation. T and a validation set S of 50 samples V , the training set consists of the initial labeled sample set S I and the unlabeled sample set S U The present invention constructs a fault diagnosis model based on different numbers of initial labeled sample sets. (10*5-fold cross validation divides the sample set into 5 equal parts, 4 of which are used as training sets and 1 as verification set. The order of the sample set is randomly shuffled 10 times to form 50 different combinations of training sets and verification sets).

[0041] A transformer fault diagnosis model based on multi-classification SVM is established. The initial labeled sample set is used to train the initial classifier. SVM is a method suitable for solving small sample classification problems and is suitable for semi-supervised algorithms. OAA is used to calculate multiple posterior probabilities on each sample of the multi-classification SVM unlabeled sample set SU. f ( x ), and predict the fault class for the unlabeled sample set and the validation set. Here is how to get a single posterior probability: ; ; ; ; in is the Lagrange multiplier; Cis a penalty factor used to control the degree of penalty for misclassified samples. The kernel function projects the original feature space into a higher-dimensional space. The present invention uses a radial basis kernel function with better classification effect. The kernel parameter c Included in the kernel function, b It is the domain value of the classification, and the function is used to predict the category of the sample.

[0042] A semi-supervised iterative model for filtering cross-validation feature optimization is established. m In the iteration, use the X strategy to select S U The samples with high confidence are integrated with the labeled sample set SI to reconstruct the labeled training set S L ( m ).

[0043] X strategy: If S U The posterior probability of the fault category predicted by the sample in is greater than S U The average posterior probability of all sample prediction classifications in the training set is considered to be credible, and the credible samples are added to the training set to reconstruct the labeled training set S. L ( m ).

[0044] ; Retrain the SVM multi-classifier using the labeled training set SL(m) and re-predict the unlabeled sample set S U and validation set S V Recalculate S U The posterior probability of each sample in is m The iteration and m -1 iteration for unlabeled samples S U The predicted label is calculated, and the number of different labels before and after is calculated. If the number of different labels is less than 1, the iteration stops. If the number is greater than 1, the iteration is returned to the next step. m +1 iterations until m =10. After the iteration stops, the validation set S is obtained V The final classification rate. According to the new training set S L ( m ) feature space and the corresponding final label, and sort the feature space using ReliefF in the filtering method. The Y strategy is used to select the optimal feature.

[0045] Y strategy: According to the sorting, select the feature space of each proportion, based on S L ( m ) for training, and then in S V Verify on the above, select the ratio corresponding to the maximum classification rate, and construct a new feature space to replace the encoding.

[0046] go throughr * K fold cross validation, calculate D V The average classification rate on , and the Z strategy is used to select the optimal features.

[0047] ; Z strategy: r * K The optimal feature space matrix obtained by folding is selected, and the feature space corresponding to the smallest difference with the average classification rate is selected as the optimal feature.

[0048] ; 3) The final classification rate is used as the value of individual fitness in the transformer fault diagnosis method based on package feature optimization. The chromosome is re-encoded using the optimal feature space, genetic operations are performed to generate offspring individuals, offspring chromosomes are decoded, and the fitness of offspring chromosome individuals is calculated by the semi-supervised cross-validation transformer fault diagnosis method based on the filtering method.

[0049] 4) Meet the required number of genetic algorithm iterations and output the optimal feature space, optimal SVM parameters and cross-validation rate on the validation set.

[0050] Based on different initial labeled training sets, according to the DGA data set of IEC TC 10, the present invention compares the average correct rates of transformer fault diagnosis on the validation set after 10*5-fold cross validation for hybrid semi-supervised feature optimization, hybrid standard feature optimization, wrapped semi-supervised feature optimization, wrapped standard feature optimization, filtered semi-supervised feature optimization, and filtered standard feature optimization, and the results are shown in Table 3. Among them, the filtering method does not use the genetic algorithm because it is not combined with the wrapped method, and the wrapped method does not use the XYZ strategy because it is not combined with the filtering method. As shown in Table 3, based on different initial labeled training sets, for the hybrid, wrapped and filtered methods, the semi-supervised method can obtain a higher average correct rate than the standard method (since the standard method does not use unlabeled samples, the X strategy is not used), which shows that the semi-supervised algorithm that uses unlabeled samples to expand the labeled sample set can optimize the model of the standard algorithm and improve the fault diagnosis rate. Based on different initial labeled training sets, regardless of the semi-supervised algorithm or the standard algorithm, the algorithm using the hybrid feature optimization can select more optimized features than the wrapped and filtered methods, thereby improving the fault diagnosis accuracy. Moreover, as the number of initial labeled training sets increases, the accuracy of all algorithms improves in most cases.

[0051] In order to prove the advantages of the XYZ strategy, since the hybrid semi-supervised feature selection must include the Z strategy, the present invention uses the hybrid semi-supervised feature selection of the XYZ strategy, the hybrid semi-supervised feature selection of the YZ strategy, the hybrid semi-supervised feature selection of the XZ strategy, and the hybrid semi-supervised feature selection without the XYZ strategy (i.e., the wrapped semi-supervised feature selection) method for comparison based on 90 initial labeled training sets, as shown in Table 4. It is found that the hybrid semi-supervised feature selection using the XYZ strategy obtains the best accuracy, while the hybrid semi-supervised feature selection of the XZ strategy is better than the hybrid semi-supervised feature selection of the YZ strategy, which illustrates the importance of the X strategy in selecting samples with high confidence in unlabeled samples to improve the model.

[0052] Table 3. Average accuracy and optimal feature subsets of transformer fault diagnosis obtained by different algorithms

[0053] Table 4. Effect of XYZ strategy on average accuracy of transformer fault diagnosis

[0054] The DGA dataset (312 samples in total) provided by the local power company is used to test the robustness of the optimal feature subset selected by the proposed method. Table 5 shows the feature subsets selected by different traditional methods, hybrid semi-supervised feature selection, wrapped semi-supervised feature selection and filtered semi-supervised feature selection, and the average fault diagnosis rate obtained after 50 trainings on the DGA dataset provided by the power company through semi-supervised 10*5-fold cross validation (where the initial training set is 80 samples). Obviously, the average accuracy obtained by hybrid semi-supervised feature selection, wrapped semi-supervised feature selection and filtered semi-supervised feature selection is much higher than that of other traditional methods, and the hybrid semi-supervised feature selection has the most obvious improvement in accuracy, which shows the strong robustness of the optimal feature subset selected by the proposed algorithm.

[0055] Table 5 Features selected by different methods and corresponding semi-supervised average accuracy

[0056] The above-described embodiments are used to illustrate the present invention, not to limit the present invention, so changes in the exemplified values ​​or replacement of equivalent elements should still fall within the scope of the present invention.

[0057] From the above detailed description, it can be understood by those skilled in the art that the present invention can indeed achieve the aforementioned objectives and is in compliance with the provisions of the Patent Law.

[0058] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic creative concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. It should be pointed out that any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0059] It should be noted that the above description of the relevant process is only for example and explanation, and does not limit the scope of application of this specification. For those skilled in the art, various modifications and changes can be made to the process under the guidance of this specification. However, these modifications and changes are still within the scope of this specification.

[0060] The basic concepts have been described above. Obviously, for those of ordinary skill in the art who have read this application, the above invention disclosure is only for example and does not constitute a limitation of this application. Although not explicitly stated here, those of ordinary skill in the art may make various modifications, improvements and amendments to this application. Such modifications, improvements and amendments are suggested in this application, so such modifications, improvements and amendments still belong to the spirit and scope of the exemplary embodiments of this application.

[0061] At the same time, the present application uses specific words to describe the embodiments of the present application. For example, "one embodiment", "an embodiment", and / or "some embodiments" refer to a certain feature, structure or characteristic related to at least one embodiment of the present application. Therefore, it should be emphasized and noted that "one embodiment" or "an embodiment" or "an alternative embodiment" mentioned twice or more in different positions in this specification does not necessarily refer to the same embodiment. In addition, some features, structures or characteristics in one or more embodiments of the present application can be appropriately combined.

[0062] In addition, it will be understood by those skilled in the art that various aspects of the present application may be illustrated and described by a number of patentable categories or situations, including any new and useful process, machine, product or combination of substances, or any new and useful improvement thereof. Therefore, various aspects of the present application may be implemented entirely by hardware, entirely by software (including firmware, resident software, microcode, etc.), or by a combination of hardware and software. The above hardware or software may be referred to as a "unit", "module" or "system". In addition, various aspects of the present application may take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0063] The computer program code required for the operation of each part of the present application can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, Python, etc., conventional procedural programming languages ​​such as C programming language, VisualBasic, Fortran2103, Perl, COBOL2102, PHP, ABAP, dynamic programming languages ​​such as Python, Ruby and Groovy, or other programming languages. The program code can be run entirely on the user's computer, or run on the user's computer as an independent software package, or run partially on the user's computer and partially on the remote computer, or run entirely on the remote computer or server. In the latter case, the remote computer can be connected to the user's computer through any network form, such as a local area network (LAN) or a wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service such as software as a service (SaaS).

[0064] In addition, unless explicitly stated in the claims, the order of the processing elements and sequences described in this application, the use of alphanumeric characters, or the use of other names are not intended to limit the order of the processes and methods of this application. Although the above disclosure discusses some invention embodiments that are currently considered useful through various examples, it should be understood that such details are only for illustrative purposes, and the attached claims are not limited to the disclosed embodiments. On the contrary, the claims are intended to cover all modifications and equivalent combinations that are consistent with the essence and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a pure software solution, for example, installation on an existing server or mobile device.

[0065] Similarly, it should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the above description of the embodiments of the application, multiple features are sometimes combined into one embodiment, figure or description thereof. However, this method of the application should not be interpreted as reflecting the intention that the claimed object requires more features than those explicitly stated in each claim. On the contrary, the subject of the invention should have fewer features than the above single embodiment.

Claims

1. A hybrid semi-supervised transformer fault diagnosis method, characterized in that: include: S1. Establish a semi-supervised iterative transformer fault diagnosis model based on filtered cross-validation feature optimization, as follows: Get DGA data and divide it into training set S T and validation set S V , training set S T Including the initial labeled sample set S I and the unlabeled sample set S U ; Using the initial labeled sample set S I Train the initial classifier and establish a transformer fault diagnosis multi-classification model based on support vector machine; The transformer fault diagnosis multi-classification model based on support vector machine is used to predict the unlabeled sample set S U and validation set S V Based on the one-to-many calculation of the transformer fault diagnosis multi-classification model based on support vector machine in the unlabeled sample set S U Multiple posterior probabilities on each sample, and then average the posterior probabilities of each sample; S2. Establish a semi-supervised iterative model for filtering cross-validation feature selection combined with the XYZ strategy, as follows: In the m In the iteration, the unlabeled sample set S is selected using the X strategy U The samples with high confidence are combined with the labeled sample set S I Integrate and reconstruct the labeled training set S L ( m ); Using the labeled training set S L ( m ) Retrain the multi-classifier of the support vector machine and re-predict the unlabeled sample set S U and validation set S V The fault category; Recalculate the unlabeled sample set S U The posterior probability of each sample in ; Find the m The iteration and m -1 iteration for the unlabeled sample set S U For the predicted labels, calculate the number of different labels before and after. If the number of different labels is less than 1, the iteration stops. If the number is greater than 1, proceed to m +1 iteration until the m The upper limit value of After the iteration stops, the validation set S is obtained V The final classification rate; According to the new training set S L ( m )’s feature space and the corresponding final label, sort the feature space using a filtering method, and select the optimal feature using the Y strategy; go through r * K Fold cross validation, calculate the validation set S V The average classification rate on the graph, and the Z strategy is used to select the optimal features; S3. Establish a transformer fault diagnosis model based on wrap-around feature optimization, as follows: Establish a feature optimization model combining genetic algorithm with support vector machine, and simultaneously optimize the parameters of support vector machine and the input DGA features; Set genetic algorithm parameters, including population, number of iterations, and chromosome length; Encode the support vector machine parameters and input DGA features, perform random initial population, chromosome decoding, individual fitness ranking, genetic operations, generate offspring individuals, offspring chromosome decoding, and offspring chromosome individual fitness calculation; When the stopping condition is reached, the optimal feature space, optimal support vector machine parameters and cross-validation rate are output; S4. Establish a hybrid semi-supervised feature optimization transformer fault diagnosis model, as follows: Construction of feature space: Select specific types of gas concentrations, various gas ratios and relative percentages recommended by IEC or IEEE as the initial feature space; Reconstruct the original DGA data, based on the data set formed by the initial feature space, and use the synthetic minority class oversampling technology to change the imbalance of the data; Dataset preprocessing: Perform arctangent transformation and normalization on the feature space of the dataset with balanced data distribution; A semi-supervised iterative model based on filtered cross-validation feature optimization is established, and the final classification rate is used as the value of individual fitness in the transformer fault diagnosis model based on parcel feature optimization. The chromosomes are re-encoded using the obtained optimal feature space. Among them, the X strategy is: if the sample set S is not labeled U The posterior probability of the sample prediction classification in is greater than that of the unlabeled sample set S U The average posterior probability of all sample prediction classifications in the training set S is considered to be credible, and the credible samples are added to the training set S T Reconstruct the labeled training set S L ( m ); The Y strategy is: select the feature space of each proportion according to the sorting, based on S L ( m ) for training and on the validation set S V Verify on the above, select the ratio corresponding to the maximum classification rate, and replace the encoding with the new feature space; The Z strategy is: r * K The optimal feature space matrix obtained by cross-validation is selected, and the feature space corresponding to the smallest difference with the average classification rate is selected as the optimal feature.

Images