Transformer fault diagnosis method based on Bayesian network

Abstract

The invention discloses a transformer fault diagnosis method based on a Bayesian network, and the method comprises the steps: firstly carrying out the structure learning of the Bayesian network based on an MWST-MTE-K2 algorithm, namely, firstly building a maximum support tree MWST, taking the maximum node degree in the maximum support tree MWST as the maximum number of father nodes, obtaining the arrangement order of the nodes according to the MTE algorithm, and carrying out the structure learning of the Bayesian network; inputting the arrangement sequence of the nodes and the maximum number of father nodes into a K2 algorithm for Bayesian network structure learning, and selecting an optimal Bayesian network structure by using a heuristic algorithm; and finally, constructing a transformer fault diagnosis model based on the optimal Bayesian network structure, and carrying out transformer fault type diagnosis by adopting the transformer fault diagnosis model. According to the method, the accuracy of structure learning is improved, the logicality and coherence of a network structure are better maintained, and thus the accuracy of transformer fault diagnosis is improved.

Description

Technical Field

[0001] The invention relates to the technical field of power fault diagnosis, and in particular to a transformer fault diagnosis method based on a Bayesian network. Background Art

[0002] In modern society, electricity is a key energy source to support economic development and social operation. With the continuous growth of China's economy, the total electricity consumption of the whole society has continued to rise, which has made the requirements for the reliability of power supply in the power system more and more stringent. As the core equipment of the power transformation link of the power system, the operating status of the transformer is directly related to the stability of the power system. Transformer failure may cause partial or even overall paralysis of the power system, causing huge economic losses. The traditional transformer maintenance mode has changed from post-maintenance to regular maintenance. Although it has solved the maintenance efficiency problem to a certain extent, it still has the disadvantages of waste of resources and excessive maintenance. With the development of technologies such as the Internet of Things and artificial intelligence, the condition maintenance mode has emerged. It can detect potential faults in advance and effectively avoid accidents by real-time monitoring of the transformer status. At present, oil-immersed transformers are widely used in power systems due to their many advantages. When a transformer fails, the hydrocarbons in the insulating oil will react to generate characteristic gases. Based on the dissolved gas analysis (DGA) in oil, the type of transformer fault can be diagnosed by analyzing the content and composition of these characteristic gases. However, the causes of transformer failures are complex and the process of generating characteristic gases is full of uncertainty, which makes it difficult to improve the accuracy of fault diagnosis. This has also become a hot issue in current research.

[0003] In recent years, artificial intelligence technology has developed rapidly and has been widely used in the field of transformer fault diagnosis, such as fuzzy theory, artificial neural network, support vector machine, extreme learning machine and Bayesian network. However, these algorithms have problems such as insufficient mining of gas in oil data and poor recognition ability when fault evidence information is incomplete. Although Bayesian network has unique advantages in dealing with uncertainty problems, it is difficult to determine the causal relationship between its nodes in transformer fault diagnosis. Therefore, it is of great practical significance to study how to optimize Bayesian network and improve its accuracy in transformer fault diagnosis.

[0004] Bayesian Network (BN) not only involves the knowledge of probability theory but also the knowledge of graph theory, and combines the two in the form of numerical and graphical combination to solve problems from the perspective of quantity and graphics. As a powerful uncertain causal reasoning tool, it has achieved great success and wide application in various fields, such as data mining, clinical medicine, natural language processing, fault diagnosis, etc. It is one of the best tools for solving uncertainty problems.

[0005] The learning of Bayesian networks mainly consists of structural learning and parameter learning, among which structural learning is the basis and core of Bayesian network construction and application. However, since the search space grows exponentially with the increase of the problem scale, it has been proved that automatically learning the BN structure from limited observation data is an NP-hard problem. To address this problem, researchers have proposed three main learning methods: constraint-based methods, scoring function-based methods, and hybrid learning methods. In recent years, it has been proposed to use a naturally inspired swarm intelligence optimization algorithm to solve this problem, but this type of algorithm is often accompanied by problems such as easy to fall into local optimality and high computational cost. In recent years, it has been proposed to use the K2 algorithm to construct the Bayesian network structure. The K2 algorithm can obtain a better network structure under the correct node order, but this type of algorithm is heavily dependent on the node order in the node space, which makes the structure of the learned Bayesian network deviate from the real structure, so the inference result will also change accordingly. Therefore, how to obtain the correct node order is still a difficult problem. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a transformer fault diagnosis method based on a Bayesian network, and to obtain the optimal network structure by performing structural learning of the transformer fault diagnosis model based on the Bayesian network structure learning algorithm of the MWST-MTE-K2 algorithm, thereby improving the accuracy of structural learning and better maintaining the logic and coherence of the network structure, thereby improving the accuracy of transformer fault diagnosis.

[0007] The technical solution of the present invention is: A transformer fault diagnosis method based on Bayesian network specifically includes the following steps: (1) Bayesian network structure learning is performed based on the MWST-MTE-K2 algorithm, that is, first, the corresponding variable set is constructed according to the input data set, and the maximum spanning tree MWST is established according to the mutual information between the nodes in the variable set, and the maximum node degree in the maximum spanning tree MWST is used as the maximum number of parent nodes; then the node arrangement order is obtained according to the MTE algorithm, that is, the connection information of each node is calculated by transfer entropy, and after the initial network is generated, the initial network is decomposed into multiple strongly connected units and topological sorting and full permutation are performed to obtain the node arrangement order; finally, the node arrangement order and the maximum number of parent nodes are input into the K2 algorithm for Bayesian network structure learning, and a series of Bayesian network structures are obtained, and then the heuristic algorithm is used to select the optimal Bayesian network structure; (2) A transformer fault diagnosis model is constructed based on the optimal Bayesian network structure, and the transformer fault diagnosis model is used to diagnose the transformer fault type.

[0008] The specific steps of constructing a corresponding variable set according to the input data set, establishing a maximum spanning tree MWST according to the mutual information between nodes in the variable set, and taking the maximum node degree in the maximum spanning tree MWST as the maximum number of parent nodes are as follows: S11. First, build the corresponding variable set based on the sample data in the data set D. Each variable in the variable set is used as a node in the Bayesian network structure. The mutual information between nodes is calculated according to the following formula (1). The calculated mutual information between nodes is arranged from large to small. The edges with the largest mutual information are continuously selected to join the maximum spanning tree MWST, and all nodes are traversed to complete the construction of the maximum spanning tree MWST. (1); In formula (1), Represents a variable set Midpoint and nodes The mutual information of Represents a variable set Midpoint and nodes The joint probability of Represents a variable set Midpoint The probability of Represents a variable set Midpoint The probability of Represents a variable set The total number of nodes; S12. Find the node with the maximum node degree from the maximum spanning tree MWST, and use the maximum node degree corresponding to the node as the maximum number of parent nodes K.

[0009] The specific steps of obtaining the arrangement order of nodes according to the MTE algorithm are: S13, initialize the network structure, first according to the variable set Create a collection ,gather The maximum capacity is , ; S14, variable set calculated by transfer entropy The connection information of each node in the , and then use the following formula (2) to find the maximum transfer entropy Node ,node The corresponding transfer entropy is , so we get the node and nodes The edge of the composition ; (2); The transfer entropy in formula (2) It is equivalent to the formula (3) , It is calculated by the following formula (3): (3); In formula (3), Indicates the change in the uncertainty of node X caused by the information of node Y, that is, the amount of information transmitted by node Y to node X; represents the value of node X at time t, Indicates the value of node X at time t+1; represents the value of node Y at time t; , , represents the order of node X, Represents the order of node Y; represent , and The joint probability distribution of Represents the conditional probability distribution, given that and Under the conditions, The probability of taking a certain value; Represents the conditional probability distribution, given that conditions, The probability of taking a certain value; The initial network consists of the variable set The nodes in the Connection, in the generated initial network In the network, each node has directionality, that is, the initial network consists of a set of directed edges. Composed of set E; S15, the initial network Decompose into m strongly connected units, ,in, , , ; S16. Use the TP function to topologically sort each strongly connected unit, and keep the relative position of the strongly connected unit in the topological sorting unchanged, and each node only appears in one strongly connected unit; S17. Use the Permu function to fully arrange each strongly connected unit after topological sorting, so as to obtain the arrangement order O of all nodes in the Bayesian network structure.

[0010] The heuristic algorithm uses a greedy algorithm to score and search a series of Bayesian network structures, and finds the Bayesian network structure with the largest score as the optimal Bayesian network structure.

[0011] The transformer fault diagnosis model is constructed based on the optimal Bayesian network structure, specifically using the content of characteristic gas and the mutual ratio between the characteristic gas contents to judge the type of fault occurring in the transformer, that is, the content of characteristic gas and the mutual ratio between the characteristic gas contents are used as the characteristic nodes in the variable set, and the fault type is used as the target node, and the transformer fault diagnosis model is trained by using the Bayesian network structure to select the optimal Bayesian network structure as the transformer fault diagnosis model.

[0012] The characteristic gases include H2, CH4, C2H2, C2H4 and C2H6; the ratios between the characteristic gas contents include the following nine ratios: CH4 / H2, C2H2 / C2H4, C2H4 / C2H6, C2H2 / (C1+C2), H2 / (H2+C1+C2), C2H4 / (C1+C2), CH4 / (C1+C2), C2H6 / (C1+C2), (CH4+ C2H4) / (C1+C2), wherein C1 represents the content of CH4, and C2 represents the sum of the contents of C2H2, C2H4 and C2H6; the five characteristic gases and the nine ratios are encoded as fourteen characteristic nodes in the variable set.

[0013] The fault types include seven operating states of the transformer, namely normal, low temperature overheating, medium temperature overheating, high temperature overheating, partial discharge, spark discharge and arc discharge.

[0014] The data of the characteristic nodes in the variable set are nonlinear. The data of each characteristic node is discretized and then normalized before being input into the transformer fault diagnosis model for model training, testing or transformer fault type diagnosis.

[0015] Advantages of the present invention: (1) The K2 algorithm is a greedy algorithm that uses node order and maximum number of parent nodes as input and learns network structure from data. The existence of node order effectively avoids the problem of inter-node ambiguity caused by likelihood equivalence and greatly reduces the algorithm's search space. Therefore, the K2 algorithm has long been considered a very efficient structure learning algorithm. In the existing applications of the K2 algorithm, determining a suitable maximum number of parent nodes usually relies on a large number of attempts and empirical judgments, or is based on some inaccurate prior knowledge. This traditional method often has subjectivity and uncertainty when selecting appropriate parameters. The method based on the maximum spanning tree (MWST) and maximum node degree of the present invention can determine the maximum number of parent nodes according to the characteristics of the data and the network structure itself, thereby providing a relatively objective and data-driven determination method for the K2 algorithm. Since the determination of the maximum number of parent nodes is based on the analysis of the data set and the MWST structure, it can better reflect the true characteristics of the data, thereby avoiding redundant searches and calculations caused by setting the upper limit of parent nodes too high or too low during the execution of the K2 algorithm, effectively simplifying the selection process of the maximum number of parent nodes, and reducing the risk of poor algorithm performance due to improper parameter selection.

[0016] (2) The present invention calculates the connection information of each node through transfer entropy to generate an initial network. The node connection information based on transfer entropy can reflect the dependency relationship between variables (nodes), providing a data-driven basis for the subsequent determination of the arrangement order of nodes, which helps to improve the accuracy of structural learning.

[0017] (3) The present invention uses strong connectivity analysis and network structure decomposition, combined with the TP function to perform topological sorting of strongly connected units, which can maintain the relative order between each strongly connected unit and the nodes, thereby better maintaining the logic and coherence of the network structure, so that the arrangement order of the nodes finally determined is more consistent with the original architecture characteristics of the network; and uses the Permu function to fully arrange the interior of the strongly connected unit to generate the arrangement order of all nodes in the Bayesian network structure, providing the K2 algorithm with a variety of high-quality arrangement options. This diversity of arrangement helps to explore a better network structure and improve the accuracy of the algorithm.

[0018] (4) The present invention constructs a transformer fault diagnosis model based on the optimal Bayesian network structure, and effectively handles data noise and uncertainty through probabilistic reasoning. It is suitable for incomplete or ambiguous diagnosis scenarios in transformer faults (uncertainty of characteristic gas analysis data in oil). The model is highly interpretable, and the network structure intuitively displays the causal relationship between variables (the association between gas concentration and fault type), which is better than the black box model and facilitates understanding of the diagnostic logic and verification of the results. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1It is a flow chart of the present invention.

[0020] Figure 2 It is a Bayesian network structure diagram of the transformer fault diagnosis model of the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] See Figure 1 , a transformer fault diagnosis method based on Bayesian network, specifically includes the following steps: (1) Bayesian network structure learning based on the MWST-MTE-K2 algorithm includes the following steps: S11. First, build the corresponding variable set based on the sample data in the data set D. Each variable in the variable set is used as a node in the Bayesian network structure. The mutual information (MI) between nodes is calculated according to the following formula (1). The calculated mutual information between nodes is arranged from large to small. The edges with the largest mutual information are continuously selected to join the maximum spanning tree MWST, and all nodes are traversed to complete the construction of the maximum spanning tree MWST. (1); In formula (1), Represents a variable set Midpoint and nodes The mutual information of Represents a variable set Midpoint and nodes The joint probability of Represents a variable set Midpoint The probability of Represents a variable set Midpoint The probability of Represents a variable set The total number of nodes; when , here we discuss two cases: a) when When and nodes There is a dependency between them, but it is not certain whether rely ,still rely , so it is impossible to judge and Who is the cause and who is the effect, so use an undirected edge to connect and , The closer the value is to 1, the and The stronger the dependence between them; b) when When and nodes Independent of each other, and There is no connection between them; S12, find the node with the maximum node degree from the maximum spanning tree MWST, and use the maximum node degree corresponding to the node as the maximum number of parent nodes K; S13, initialize the network structure, first according to the variable set Create a collection ,gather The maximum capacity is , ; S14. Since the mutual information between variables cannot measure the information flow and transmission direction between random variables and has no directionality, the transfer entropy probability is introduced. Transfer entropy is a kind of information entropy that can solve the problem of no direction in mutual information theory. The connection information of each node in the , and then use the following formula (2) to find the maximum transfer entropy Node ,node The corresponding transfer entropy is , so we get the node and nodes The edge of the composition ; (2); The transfer entropy in formula (2) It is equivalent to the formula (3) , i.e. node Equal to formula (3) node X, node Equal to the node Y of equation (3), It is calculated by the following formula (3): (3); In formula (3), Indicates the change in the uncertainty of node X caused by the information of node Y, that is, the amount of information transmitted by node Y to node X; represents the value of node X at time t, Indicates the value of node X at time t+1; represents the value of node Y at time t; , , represents the order of node X, Represents the order of node Y; represent , and The joint probability distribution of Represents the conditional probability distribution, given that and Under the conditions, The probability of taking a certain value; Represents the conditional probability distribution, given that conditions, The probability of taking a certain value; The Bayesian network is presented as a directed acyclic graph (DAG) in terms of graph structure. The nodes in the network clearly reveal the relationship of mutual influence and dependence between the nodes with the help of connection edges with specific directions. At the same time, in the transfer entropy theory, information is carried out in a directed manner between two random variables with random variables as carriers. Information entropy has asymmetric characteristics in the process of information transmission, which enables transfer entropy to be integrated with the construction of Bayesian networks in a unique way. Given that the node connections in the Bayesian network structure are directional (i.e. asymmetric), transfer entropy can play a role in promoting and supervising the generation of the Bayesian network structure. Transfer entropy is also based on causal learning, and the network structure is constructed according to the relevant principles of information theory and probability theory. Therefore, after the network structure is constructed using the transfer entropy method, the subsequent learning process of the network structure can be carried out based on this causal structure. After getting all nodes and nodes in the network After the transfer entropy is calculated, the formula (2) can be used to calculate the maximum transfer entropy A node of , thus obtaining a set of edges , maximizing the transfer entropy is used as a method to determine the optimal Bayesian network structure, thereby generating the initial network of the Bayesian network. The initial network consists of the variable set The nodes in the Connection, in the generated initial network In the network, each node has directionality, that is, the initial network consists of a set of directed edges. Composed of set E; S15, the initial network Decompose into m strongly connected units, ,in, , , ; S16. Use TP function for each strongly connected unit Perform topological sorting, and keep the relative positions of strongly connected units in the topological sorting unchanged, and each node only appears in one strongly connected unit; S17, using the Permu function to fully arrange each strongly connected unit after topological sorting, thereby obtaining the arrangement order O of all nodes in the Bayesian network structure; S18, inputting the node arrangement order O and the maximum number of parent nodes K into the K2 algorithm to perform Bayesian network structure learning, obtaining a series of Bayesian network structures, and then using a greedy algorithm to score and search the series of Bayesian network structures, and finding the Bayesian network structure with the largest score as the optimal Bayesian network structure; (2) At present, oil-immersed transformers are the main type of transformers in the power system. When a fault occurs inside an oil-immersed transformer, a variety of characteristic gases will be generated. The type of transformer fault is determined based on the content and composition of the characteristic gas dissolved in the oil based on the dissolved gas analysis method in oil. A transformer fault diagnosis model is constructed based on the optimal Bayesian network structure, which specifically includes the following steps: S21, encoding the content of characteristic gases, the mutual ratio between the characteristic gas contents and the fault type; Characteristic gases include H2, CH4, C2H2, C2H4 and C2H6; The mutual ratios between characteristic gas contents include the following nine ratios: CH4 / H2, C2H2 / C2H4, C2H4 / C2H6, C2H2 / (C1+C2), H2 / (H2+C1+C2), C2H4 / (C1+C2), CH4 / (C1+C2), C2H6 / (C1+C2), (CH4+ C2H4) / (C1+C2), where C1 represents the content of CH4 and C2 represents the sum of the contents of C2H2, C2H4 and C2H6. These ratios can further explore the potential relationship between characteristic gases and provide richer information for the judgment of transformer fault type and severity. For example, the ratio of C2H2 / C2H4 can be used to distinguish between discharge faults and overheating faults. When the ratio is higher, it is more inclined to discharge faults. The variable set includes fourteen characteristic nodes consisting of five characteristic gases and nine ratios. The encoding of the characteristic nodes is shown in Table 1 below; Table 1

[0023] The fault types include seven operating states of the transformer, namely normal, low temperature overheating, medium temperature overheating, high temperature overheating, partial discharge, spark discharge and arc discharge. The codes of the fault types are shown in Table 2 below; Table 2

[0024] S22, the sample data is divided into a test set and a training set. Since the data of the fourteen feature nodes in the variable set are all nonlinear, the data of each feature node is discretized with H2 (encoding X 10 ) as an example, after discretization of the data, the data volume of H2 is divided into three intervals, namely 0~0.325, 0.325~0.714 and above 0.714. These three intervals are used as the three state groups of the feature node H2, namely 1, 2 and 3; Because the content of dissolved gas in the oil is highly dispersed when the transformer has different faults, in order to reduce the error, the following formula (4) is used to normalize the data; (4); In formula (4), is the normalized feature node data, is the component of the jth dimension of the i-th sample; represents the component of the nth dimension in the i-th sample, where n represents all feature dimensions (from 1 to 5) of sample i; for example, when n=1, is the first gas content of the ith sample, and so on;

[0025] In order to verify that the constructed transformer fault diagnosis model can accurately diagnose transformer faults, the DGA data from the IECTC 10 database and the collected literature are used as sample data of transformer faults, which contains a total of 1857 fault information; after classifying the fault types, the sample data are classified according to the ratio of 8:2 to realize the division of training set and test set, among which, there are 1485 samples in the training set and 372 samples in the test set. The distribution of sample data is shown in Table 3 below;

[0026] Table 3

[0027] S23, the training set samples after data processing (discretization and normalization) are used as input, wherein the content of characteristic gas and the mutual ratio between characteristic gas contents are used as characteristic nodes in the variable set, and the fault type is used as the target node (see Figure 2), the transformer fault diagnosis model is trained by using the Bayesian network structure learning, and then the test set is input into the transformer fault diagnosis model to test the transformer fault diagnosis. The fault diagnosis result is compared with the actual fault category to obtain the transformer fault diagnosis accuracy; finally, the optimal Bayesian network structure is selected as the transformer fault diagnosis model, and the transformer fault diagnosis model is used to diagnose the transformer fault type.

[0028] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A transformer fault diagnosis method based on Bayesian network, characterized in that: The specific steps include: (1) Bayesian network structure learning is performed based on the MWST-MTE-K2 algorithm, that is, first, the corresponding variable set is constructed according to the input data set, and the maximum spanning tree MWST is established according to the mutual information between the nodes in the variable set, and the maximum node degree in the maximum spanning tree MWST is used as the maximum number of parent nodes; then the node arrangement order is obtained according to the MTE algorithm, that is, the connection information of each node is calculated by transfer entropy, and after the initial network is generated, the initial network is decomposed into multiple strongly connected units and topological sorting and full permutation are performed to obtain the node arrangement order; Finally, the node arrangement order and the maximum number of parent nodes are input into the K2 algorithm for Bayesian network structure learning, and a series of Bayesian network structures are obtained. Then, the heuristic algorithm is used to select the optimal Bayesian network structure. (2) A transformer fault diagnosis model is constructed based on the optimal Bayesian network structure, and the transformer fault diagnosis model is used to diagnose the transformer fault type.

2. A transformer fault diagnosis method based on Bayesian network according to claim 1, characterized in that: The specific steps of constructing a corresponding variable set according to the input data set, establishing a maximum spanning tree MWST according to the mutual information between nodes in the variable set, and taking the maximum node degree in the maximum spanning tree MWST as the maximum number of parent nodes are as follows: S11. First, build the corresponding variable set based on the sample data in the data set D. Each variable in the variable set is used as a node in the Bayesian network structure. The mutual information between nodes is calculated according to the following formula (1). The calculated mutual information between nodes is arranged from large to small. The edges with the largest mutual information are continuously selected to join the maximum spanning tree MWST, and all nodes are traversed to complete the construction of the maximum spanning tree MWST. （1）； In formula (1), Represents a variable set Midpoint and nodes The mutual information of Represents a variable set Midpoint and nodes The joint probability of Represents a variable set Midpoint The probability of Represents a variable set Midpoint The probability of Represents a variable set The total number of nodes; S12. Find the node with the maximum node degree from the maximum spanning tree MWST, and use the maximum node degree corresponding to the node as the maximum number of parent nodes K.

3. A transformer fault diagnosis method based on Bayesian network according to claim 2, characterized in that: The specific steps of obtaining the arrangement order of nodes according to the MTE algorithm are: S13, initialize the network structure, first according to the variable set Create a collection ,gather The maximum capacity is , ; S14, variable set calculated by transfer entropy The connection information of each node in the , and then use the following formula (2) to find the maximum transfer entropy Node ,node The corresponding transfer entropy is , so we get the node and nodes The edge of the composition ; （2）； The transfer entropy in formula (2) It is equivalent to the formula (3) , It is calculated by the following formula (3): （3）； In formula (3), Indicates the change in the uncertainty of node X caused by the information of node Y, that is, the amount of information transmitted by node Y to node X; represents the value of node X at time t, Indicates the value of node X at time t+1; represents the value of node Y at time t; , , represents the order of node X, Represents the order of node Y; represent , and The joint probability distribution of Represents the conditional probability distribution, given that and Under the conditions, The probability of taking a certain value; Represents the conditional probability distribution, given that conditions, The probability of taking a certain value; The initial network consists of the variable set The nodes in the Connection, in the generated initial network In the network, each node has directionality, that is, the initial network consists of a set of directed edges. Composed of set E; S15, the initial network Decompose into m strongly connected units, ,in, , , ; S16. Use the TP function to topologically sort each strongly connected unit, and keep the relative position of the strongly connected unit in the topological sorting unchanged, and each node only appears in one strongly connected unit; S17. Use the Permu function to fully arrange each strongly connected unit after topological sorting, so as to obtain the arrangement order O of all nodes in the Bayesian network structure.

4. The transformer fault diagnosis method based on Bayesian network according to claim 1 is characterized in that: The heuristic algorithm uses a greedy algorithm to score and search a series of Bayesian network structures, and finds the Bayesian network structure with the largest score as the optimal Bayesian network structure.

5. The transformer fault diagnosis method based on Bayesian network according to claim 1 is characterized in that: The transformer fault diagnosis model is constructed based on the optimal Bayesian network structure, specifically using the content of characteristic gas and the mutual ratio between the characteristic gas contents to judge the type of fault occurring in the transformer, that is, the content of characteristic gas and the mutual ratio between the characteristic gas contents are used as the characteristic nodes in the variable set, and the fault type is used as the target node, and the transformer fault diagnosis model is trained by using the Bayesian network structure to select the optimal Bayesian network structure as the transformer fault diagnosis model.

6. A transformer fault diagnosis method based on Bayesian network according to claim 5, characterized in that: The characteristic gases include H2, CH4, C2H2, C2H4 and C2H6; the ratios between the characteristic gas contents include the following nine ratios: CH4 / H2, C2H2 / C2H4, C2H4 / C2H6, C2H2 / (C1+C2), H2 / (H2+C1+C2), C2H4 / (C1+C2), CH4 / (C1+C2), C2H6 / (C1+C2), (CH4+ C2H4) / (C1+C2), wherein C1 represents the content of CH4, and C2 represents the sum of the contents of C2H2, C2H4 and C2H6; the five characteristic gases and the nine ratios are encoded as fourteen characteristic nodes in the variable set.

7. The transformer fault diagnosis method based on Bayesian network according to claim 5 is characterized in that: The fault types include seven operating states of the transformer, namely normal, low temperature overheating, medium temperature overheating, high temperature overheating, partial discharge, spark discharge and arc discharge.

8. The transformer fault diagnosis method based on Bayesian network according to claim 6 is characterized in that: The data of the characteristic nodes in the variable set are nonlinear. The data of each characteristic node is discretized and then normalized before being input into the transformer fault diagnosis model for model training, testing or transformer fault type diagnosis.

Images