Transformer fault detection method, device, equipment and medium

By combining multiple primary fusion multi-classification modules and single-layer BP neural network modules, the probability of transformer failure type is calculated, and the problem of inaccurate fault detection in the prior art is solved, and the accuracy and stability of transformer fault detection is improved.

CN118940115AActive Publication Date: 2024-11-12GUANGDONG POWER GRID CO LTD +1
View PDF 13 Cites 0 Cited by

Patent Information

Application Number
CN202410991113.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-12
Estimated Expiration
2044-07-23

Smart Images

  • Figure CN118940115A_ABST
    Figure CN118940115A_ABST
Patent Text Reader

Abstract

The invention discloses a transformer fault detection method, device and equipment and a medium. The method comprises the steps of obtaining equipment working parameters of a to-be-detected transformer; performing fault detection on the equipment working parameters based on a transformer fault detection model to obtain fault type information of the to-be-detected transformer; wherein the transformer fault detection model comprises a plurality of primary fusion multi-classification modules and a plurality of single-layer BP neural network modules; the primary fusion multi-classification module is used for determining a first probability that the to-be-detected transformer belongs to a plurality of pieces of fault type information based on the equipment working parameters; the single-layer BP neural network module is used for determining a second probability corresponding to the fault type information based on the first probabilities corresponding to the same fault type information respectively determined by the plurality of primary fusion multi-classification modules; and determining the fault type information of the to-be-detected transformer based on the maximum second probability in the second probabilities corresponding to the multiple pieces of fault type information. The accuracy and stability of transformer fault detection can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

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

[0002] Transformers are an important part of adjusting the operation mode of the power system. Timely detection of transformer faults is of great significance to the stable operation of the power system.

[0003] At present, the existing technology usually trains transformer fault classifiers based on fault sample data sets. The transformer fault classifier can automatically classify transformer faults online by analyzing the input transformer physical characteristics, which facilitates the substation maintenance personnel to respond quickly on site. However, there are many types of transformer faults, and it is difficult to accurately predict the probability of the transformer belonging to each fault type, resulting in unreliable transformer fault classification results. Summary of the invention

[0004] The present invention provides a transformer fault detection method, device, equipment and medium to solve the problem of inaccurate transformer fault detection and improve the accuracy and stability of transformer fault detection. According to one aspect of the present invention, a transformer fault detection method is provided, comprising:

[0005] Acquire equipment working data of the transformer to be detected, and acquire a transformer fault detection model corresponding to the transformer to be detected; the transformer fault detection model includes multiple primary fusion multi-classification modules and multiple single-layer BP neural network modules, the primary fusion multi-classification module includes multiple basic multi-classifiers, and at least one of the classifier type, the number of classifiers and the combination method corresponding to the multiple basic multi-classifiers included in different primary fusion multi-classification modules is different; each single-layer BP neural network module corresponds to a fault type information;

[0006] Based on the transformer fault detection model, the equipment working data is tested for faults to obtain the fault type information of the transformer to be tested. During the fault detection process:

[0007] For each primary fusion multi-classification module, multi-classification processing is performed on the equipment working data based on each basic multi-classifier in the primary fusion multi-classification module to obtain a basic multi-classification result corresponding to the basic multi-classifier, and based on the combination mode corresponding to the multiple basic multi-classifiers included in the primary fusion multi-classification module, the basic multi-classification results corresponding to the multiple basic multi-classifiers are aggregated to obtain the classification result corresponding to the primary fusion multi-classification module; the classification result includes the first probabilities corresponding to the multiple fault type information respectively belonging to the transformer to be detected;

[0008] For each fault type information, based on the single-layer BP neural network module corresponding to the fault type information, the first probabilities corresponding to the fault type information in the classification results corresponding to the multiple primary fusion multi-classification modules are calculated and processed to obtain the second probability that the transformer to be detected belongs to each fault type information;

[0009] The fault type information corresponding to the maximum second probability is determined as the fault type information of the transformer to be detected; the maximum second probability is the maximum probability among the second probabilities corresponding to the multiple fault type information respectively.

[0010] According to another aspect of the present invention, there is provided a transformer fault detection device, comprising:

[0011] The equipment working data and detection model acquisition module is used to obtain the equipment working data of the transformer to be detected, and obtain the transformer fault detection model corresponding to the transformer to be detected; the transformer fault detection model includes multiple primary fusion multi-classification modules and multiple single-layer BP neural network modules, the primary fusion multi-classification module includes multiple basic multi-classifiers, and at least one of the classifier type, number of classifiers and combination mode corresponding to the multiple basic multi-classifiers included in different primary fusion multi-classification modules is different; each single-layer BP neural network module corresponds to a fault type information;

[0012] The fault type information determination module is used to perform fault detection on the equipment working data based on the transformer fault detection model to obtain the fault type information of the transformer to be detected. In the fault detection process: for each primary fusion multi-classification module, multi-classification processing is performed on the equipment working data based on each basic multi-classifier in the primary fusion multi-classification module to obtain the basic multi-classification result corresponding to the basic multi-classifier, and based on the combination mode corresponding to the multiple basic multi-classifiers included in the primary fusion multi-classification module, the basic multi-classification results corresponding to the multiple basic multi-classifiers are aggregated to obtain the classification result corresponding to the primary fusion multi-classification module; the classification result includes the first probabilities corresponding to the multiple fault type information respectively of the transformer to be detected; for each fault type information, based on the single-layer BP neural network module corresponding to the fault type information, the first probabilities corresponding to the fault type information in the classification results corresponding to the multiple primary fusion multi-classification modules are calculated and processed to obtain the second probability that the transformer to be detected belongs to each fault type information; the fault type information corresponding to the maximum second probability is determined as the fault type information of the transformer to be detected; the maximum second probability is the maximum probability among the second probabilities corresponding to the multiple fault type information.

[0013] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0014] at least one processor; and

[0015] a memory communicatively connected to at least one processor; wherein,

[0016] The memory stores a computer program executable by at least one processor. The computer program is executed by at least one processor so that the at least one processor can execute the transformer fault detection method according to any embodiment of the present invention.

[0017] According to another aspect of the present invention, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the transformer fault detection method according to any embodiment of the present invention when executed.

[0018] The technical solution of the embodiment of the present invention, by setting multiple primary fusion multi-classification modules and multiple single-layer BP neural network modules in the transformer fault detection model, and making each single-layer BP neural network module correspond to a fault type information, instead of setting only one primary fusion multi-classification module, can calculate and process the first probabilities corresponding to the fault type information in the classification results corresponding to the multiple primary fusion multi-classification modules based on the single-layer BP neural network module, and obtain the second probability that the transformer to be detected belongs to each fault type information, so that the first probabilities corresponding to the fault type information in the classification results corresponding to different primary fusion multi-classification modules are complementary, which solves the problem of inaccuracy of the transformer fault classification results, improves the robustness of the transformer fault detection model, and helps the transformer fault detection model to extract the complex relationship in the equipment working parameters, thereby improving the accuracy and stability of transformer fault detection.

[0019] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present invention, nor is it intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of 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.

[0021] Figure 1 is a flow chart of a transformer fault detection method provided by Embodiment 1 of the present invention;

[0022] Figure 2 It is a flow chart of a transformer fault detection model training phase in a transformer fault detection method provided in Embodiment 2 of the present invention;

[0023] Figure 3 is a schematic diagram of the structure of a first generation network applicable to the second embodiment of the present invention;

[0024] Figure 4 is a schematic diagram of the structure of a first discrimination network applicable to the second embodiment of the present invention;

[0025] Figure 5 is a schematic diagram of the structure of a second discrimination network applicable to the second embodiment of the present invention;

[0026] Figure 6 is a schematic structural diagram of a transformer fault detection device provided in Embodiment 3 of the present invention;

[0027] Figure 7 It is a structural schematic diagram of an electronic device for implementing the transformer fault detection method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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 should fall within the scope of protection of the present invention.

[0029] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0030] Embodiment 1

[0031] Figure 1 This is a flowchart of a transformer fault detection method provided by the first embodiment of the present invention. This embodiment is applicable to automatic fault detection of transformers. The method can be executed by a transformer fault detection device. The transformer fault detection device can be implemented in the form of hardware and / or software. The transformer fault detection device can be configured in a computer and a server. Figure 1 As shown, the method includes:

[0032] S110. Obtain equipment working data of the transformer to be detected, and obtain a transformer fault detection model corresponding to the transformer to be detected; the transformer fault detection model includes multiple primary fusion multi-classification modules and multiple single-layer BP neural network modules, the primary fusion multi-classification module includes multiple basic multi-classifiers, and at least one of the classifier type, number of classifiers and combination method corresponding to the multiple basic multi-classifiers included in different primary fusion multi-classification modules is different; each single-layer BP neural network module corresponds to a fault type information.

[0033] In this embodiment, the transformer to be detected is a transformer that needs to be fault-detected. The equipment working parameters are parameters that characterize the state of the transformer to be detected during operation. For example, the equipment working parameters include but are not limited to vibration characteristic parameters, ordinary image characteristic parameters, infrared image characteristic parameters, and oil valve type ultra-high frequency partial discharge characteristic parameters.

[0034] Specifically, the device operating parameters of the transformer to be detected are obtained by reading the collected state parameters of the transformer to be detected in real time or calling the pre-stored state parameters of the transformer to be detected. In some embodiments, the state parameters of the transformer to be detected are collected in real time based on the parameter collection device to obtain the device operating parameters of the transformer to be detected, wherein the parameter collection device includes but is not limited to a vibration sensor, a vibration sensor collection device, an oil valve type UHF partial discharge sensor, an oil valve type ultra-high frequency partial discharge collection device, an ordinary camera, an infrared camera and an intelligent cloud platform; in some embodiments, the collected state parameters of the transformer to be detected are stored in a computer or a server, so as to subsequently call the pre-stored state parameters of the transformer to be detected, and the device operating parameters of the transformer to be detected are obtained, and this embodiment does not limit this.

[0035] The fault type information is information representing the fault type of the transformer to be detected. For example, assuming that there are V fault types, the fault type information v represents the vth fault type, where v∈[1,V]. The fault type information includes but is not limited to symbolic identifiers, text identifiers, and logical identifiers. For example, assuming that the number of fault types is 3, the fault type information corresponding to the second fault type is a one-hot code "010".

[0036] The transformer fault detection model is a classification model for fault detection of the transformer to be detected, wherein the primary fusion multi-classification module is used to integrate multiple basic multi-classifiers to perform multi-classification processing on the equipment working parameters. The basic multi-classifier is a multi-classifier based on a single classification algorithm, and the classifier type is the type of the single classification algorithm to which the basic multi-classifier belongs, wherein the classifier type corresponding to the basic multi-classifier includes but is not limited to decision tree, support vector machine, naive Bayes, K-nearest neighbor, random forest and gradient boosting tree. The number of classifiers is the number of basic multi-classifiers corresponding to each classifier type in the primary fusion multi-classification module, and the combination method is the combination method of multiple basic multi-classifiers in the primary fusion multi-classification module, wherein the combination method includes but is not limited to self-service aggregation (Bagging) method, stacking (Stacking) method, boosting (Boosting) method, simple averaging (Averaging) method, voting (voting) method and blending (blending) method.

[0037] It should be noted that the transformer fault detection model includes two or more primary fusion multi-classification modules, and any two primary fusion multi-classification modules in the transformer fault detection model are different, and one or more of the classifier type, the number of classifiers and the combination method may be different. Exemplarily, the combination methods corresponding to the primary fusion multi-classification module 1, the primary fusion multi-classification module 2 and the primary fusion multi-classification module 3 are all Stacking, and the primary fusion multi-classification module 1 and the primary fusion multi-classification module 2 both include 4 basic multi-classifiers, but the classifier types corresponding to the 4 basic multi-classifiers included in the primary fusion multi-classification module 1 are all decision trees, the classifier types corresponding to the 4 basic multi-classifiers included in the primary fusion multi-classification module 2 are all support vector machines, and the primary fusion multi-classification module 3 includes 5 basic multi-classifiers.

[0038] By setting at least one of the classifier type, number of classifiers and combination method corresponding to the multiple basic multi-classifiers included in different primary fusion multi-classification modules to be different, the diversity of the primary fusion multi-classification modules can be ensured, so that different primary fusion multi-classification modules can complement each other, which helps the transformer fault detection model to extract the complex relationship in the equipment working parameters, thereby improving the transformer fault detection performance.

[0039] A single-layer BP neural network module is used to fuse at least partial output information of multiple primary fusion multi-classification modules. The number of single-layer BP neural network modules included in the transformer fault detection model is the same as the number of fault types corresponding to the transformer, and any two single-layer BP neural network modules included in the transformer fault detection model have the same model structure but different model parameters. Exemplarily, assuming that the number of fault types corresponding to the transformer is V, the transformer fault detection model includes Q primary fusion multi-classification modules and V single-layer BP neural network modules, wherein Q and V may be equal or unequal, and this embodiment does not limit this.

[0040] S120. Perform fault detection on the equipment working data based on the transformer fault detection model to obtain the fault type information of the transformer to be detected. In the fault detection process: for each primary fusion multi-classification module, perform multi-classification processing on the equipment working data based on each basic multi-classifier in the primary fusion multi-classification module to obtain a basic multi-classification result corresponding to the basic multi-classifier, and based on the combination method corresponding to the multiple basic multi-classifiers included in the primary fusion multi-classification module, perform fusion processing on the basic multi-classification results corresponding to the multiple basic multi-classifiers to obtain the classification result corresponding to the primary fusion multi-classification module; the classification result includes the first probabilities corresponding to the multiple fault type information respectively of the transformer to be detected; for each fault type information, based on the single-layer BP neural network module corresponding to the fault type information, calculate and process the first probabilities corresponding to the fault type information in the classification results corresponding to the multiple primary fusion multi-classification modules to obtain the second probability that the transformer to be detected belongs to each fault type information; the fault type information corresponding to the maximum second probability is determined as the fault type information of the transformer to be detected; the maximum second probability is the maximum probability among the second probabilities corresponding to the multiple fault type information.

[0041] Specifically, by inputting the equipment working data into the transformer fault detection model, the transformer fault detection model performs fault detection on the transformer to be detected based on the equipment working data, and obtains the fault type information of the transformer to be detected.

[0042] Multiple primary fusion multi-classification modules in the transformer fault detection model perform multi-classification processing on the input equipment working data at the same time. It can be understood that each primary fusion multi-classification module performs multi-classification processing on the input equipment working data based on each basic multi-classifier in the primary fusion multi-classification module, and obtains the basic multi-classification results corresponding to each basic multi-classifier. It should be noted that the basic multi-classification results include the initial probabilities corresponding to the multiple fault type information respectively corresponding to the transformer to be detected. For each fault type information, each primary fusion multi-classification module fuses the initial probabilities corresponding to each fault type information in the basic multi-classification results corresponding to the multiple basic multi-classifiers included in the primary fusion multi-classification module according to the combination mode corresponding to the multiple basic multi-classifiers included in the primary fusion multi-classification module, and obtains the first probability corresponding to the fault type information corresponding to each primary fusion multi-classification module, and each primary fusion multi-classification module combines the first probabilities corresponding to the multiple fault type information to obtain the classification result.

[0043] Exemplarily, assuming that the combination method corresponding to the multiple basic multi-classifiers included in the primary fusion multi-classification module 1 is Averaging, and the primary fusion multi-classification module 1 includes three basic multi-classifiers, the initial probabilities corresponding to each fault type information in the basic multi-classification results corresponding to the three basic multi-classifiers included in the primary fusion multi-classification module 1 are averaged to obtain the first probability corresponding to each fault type information in the classification results corresponding to the primary fusion multi-classification module 1.

[0044] For each fault type information, the first probability that the transformer to be detected belongs to the fault type information in the classification results corresponding to the multiple primary fusion multi-classification modules is input into the single-layer BP neural network module corresponding to the fault type information, so that the single-layer BP neural network module calculates and processes the input multiple first probabilities to obtain the second probability that the transformer to be detected belongs to the fault type information. In some embodiments, the single-layer BP neural network module includes multiple fusion neurons, and each fusion neuron corresponds to a primary fusion multi-classification module. In this embodiment, the number of fusion neurons in each single-layer BP neural network module is the same as the number of primary fusion multi-classification modules. Exemplarily, assuming that the transformer fault detection model includes Q primary fusion multi-classification modules, each single-layer BP neural network module includes Q fusion neurons. Optionally, for each fault type information, based on the weights corresponding to the multiple fusion neurons in the single-layer BP neural network module corresponding to the fault type information, the first probabilities corresponding to the fault type information in the classification results corresponding to the multiple primary fusion multi-classification modules are weighted summed to obtain the second probability that the transformer to be detected belongs to each fault type information.

[0045] Among them, the second probability of the transformer to be detected belonging to the fault type information is expressed as: Diagnosis v =W v,1 ×C 1,v +…+W v,q ×C q,v +…+W v,Q ×C Q,v , where W v,q represents the weight corresponding to the qth fusion neuron in the vth single-layer BP neural network module, C q,v It represents the first probability corresponding to the vth fault type information in the classification result corresponding to the qth primary fusion multi-classification module; v∈[1,V], V represents the number of fault type information, q∈[1,Q], Q represents the number of primary fusion multi-classification modules.

[0046] Specifically, the vth single-layer BP neural network module corresponding to the vth fault type information converts the first probability C corresponding to the vth fault type information in the classification result corresponding to the qth primary fusion multi-classification module into q,v , input to the qth fusion neuron O of the vth single-layer BP neural network module q,v , fusion neuron O q,v The first probability C q,v Fusion neuron O q,v The weight W v,q The product processing is performed to obtain the output result corresponding to the qth fusion neuron. The second probability Diagnosis corresponding to the vth fault type information is obtained by summing the output results corresponding to the Q fusion neurons of the vth single-layer BP neural network module. v For example, assuming that there are three fault types in the transformer, the transformer fault detection model includes four primary fusion multi-classification modules, then the transformer fault detection model includes three single-layer BP neural network modules, and the single-layer BP neural network module corresponding to the fault type information 1 includes four fusion neural network modules. 1,1 , O 2,1 , O 3,1 and O 4,1 The second probability output by the single-layer BP neural network module corresponding to fault type information 1 is Diagnosis 1 =W 1,1 ×C 1,1 +W 1,2 ×C 2,1 +W 1,3 ×C 3,1 .

[0047] The technical solution of this embodiment is to correspond each fusion neuron included in the single-layer BP neural network module to a primary fusion multi-classification module, and for each fault type information, based on the weights corresponding to multiple fusion neurons in the single-layer BP neural network module corresponding to the fault type information, perform weighted fusion processing on the first probability corresponding to the fault type information in the classification results corresponding to multiple primary fusion multi-classification modules, so as to obtain the second probability that the transformer to be detected belongs to each fault type information. The second probability can be quickly calculated based on the weights corresponding to multiple fusion neurons, which helps to improve the speed of transformer fault detection.

[0048] The maximum probability among the second probabilities corresponding to the multiple fault type information is determined as the maximum second probability, and the fault type information corresponding to the maximum second probability is determined as the fault type information of the transformer to be detected. For example, assuming that the classification result corresponding to the single-layer BP neural network module is Among them, the second highest probability is Diagnosis v =max(Diagnosis 1 ,…,Diagnosis v ,…,Diagnosis V ), then the fault type information of the transformer to be detected is Diagnosis v The corresponding fault type information v.

[0049] The technical solution of this embodiment, by setting multiple primary fusion multi-classification modules and multiple single-layer BP neural network modules in the transformer fault detection model, and making each single-layer BP neural network module correspond to a fault type information, instead of setting only one primary fusion multi-classification module, can calculate and process the first probabilities corresponding to the fault type information in the classification results corresponding to the multiple primary fusion multi-classification modules based on the single-layer BP neural network module, and obtain the second probability that the transformer to be detected belongs to each fault type information, so that the first probabilities corresponding to the fault type information in the classification results corresponding to different primary fusion multi-classification modules are complementary, which solves the problem of inaccuracy of the transformer fault classification results, improves the robustness of the transformer fault detection model, and helps the transformer fault detection model to extract the complex relationship in the equipment working parameters, thereby improving the accuracy and stability of transformer fault detection.

[0050] Embodiment 2

[0051] Figure 2This is a flowchart of the transformer fault detection model training stage in a transformer fault detection method provided in the second embodiment of the present invention. This embodiment is applicable to the case of training a transformer fault detection model, and the trained transformer fault detection model can be applied to the above embodiment. The explanations of the terms that are the same or corresponding to the above embodiments are not repeated here. Figure 2 As shown, the method includes:

[0052] S210, obtaining an original sample data set corresponding to the transformer to be detected, and determining a first training set, a first validation set and a test set in the original sample data set; the original sample data set includes multiple original sample data, each of which includes fault sample data and fault classification label data corresponding to the fault sample data.

[0053] In this embodiment, the original sample data set is a data set obtained based on a transformer of the same type as the transformer to be detected, and the original sample data includes a fault sample data and fault classification label data corresponding to the fault sample data, wherein the characteristic parameter items in the fault sample data are the same as the characteristic parameter items in the equipment operating parameters, and the fault classification label data characterizes the fault type information corresponding to the fault sample data, wherein the fault classification label data includes the probability that the fault sample data belongs to multiple fault type information respectively, wherein the probability of belonging to each fault type information is 0 or 100%, and there is only one 100% among the probabilities that the fault sample data included in the fault classification label data belongs to multiple fault type information respectively.

[0054] For example, the sample data set is represented as in, Represents the i-th original sample data in the sample data set, They are Included fault sample data and fault classification label data, among which, Among them, t in Characterization The nth characteristic parameter item in the fault sample data and the equipment working parameters both include N characteristic parameter items. Among them, b iv Characterizing Fault Sample Data The probability of belonging to fault type information v, b iv is 0 or 1 (i.e., 100%), in the fault sample data In the case of fault type information v,

[0055] Specifically, the original sample data set is obtained by calling the data set of the original sample data stored locally and / or in the server. In some embodiments, the original sample data set is preprocessed, wherein the preprocessing includes normalization. A first training set, a first validation set and a test set are obtained by randomly dividing a plurality of original sample data in the original sample data set, wherein the first training set is used to adjust the parameters of the non-single-layer BP neural network module, the first validation set is used to adjust the parameters of the single-layer BP neural network module, and the test set is used to perform performance verification on the trained transformer fault detection model. Optionally, based on the preset test set proportional coefficient and the number of non-single-layer BP neural network module parameters and the number of single-layer BP neural network module parameters in the transformer fault detection model, the first training set proportional coefficient and the first validation set proportional coefficient are determined; based on the first training set proportional coefficient, the first validation set proportional coefficient and the preset test set proportional coefficient, the original sample data set is divided into the first training set, the first validation set and the test set; the preset test set proportional coefficient is 20%.

[0056] In this embodiment, the model parameters of the transformer fault detection model are composed of non-single-layer BP neural network module parameters and single-layer BP neural network module parameters, wherein the non-single-layer BP neural network module parameters are model parameters of multiple primary fusion multi-classification modules in the transformer fault detection model, and the single-layer BP neural network module parameters are model parameters of multiple single-layer BP neural network modules in the transformer fault detection model.

[0057] The preset test set ratio coefficient is the ratio coefficient of the number of original sample data in the preset test set to the number of original sample data in the original sample data set. The preset test set ratio coefficient of 20% indicates that 20% of the original sample data in the original sample data set are divided into the test set, and the sum of the number of original sample data in the first training set and the number of original sample data in the first validation set accounts for 80% of the number of original sample data in the original sample data set. The first training set ratio coefficient is used to characterize the proportion of original sample data in the first training set to the original sample data in the original sample data set, and the first validation set ratio coefficient is used to characterize the proportion of original sample data in the first validation set to the original sample data in the original sample data set.

[0058] Specifically, the difference between 100% and the preset test set proportional coefficient (i.e., 80%) is determined as the sum of the first training set proportional coefficient and the first verification set proportional coefficient, and the ratio of the number of non-single-layer BP neural network module parameters and the number of single-layer BP neural network module parameters in the transformer fault detection model is determined as the ratio between the first training set proportional coefficient and the first verification set proportional coefficient. The first training set proportional coefficient and the first verification set proportional coefficient are analyzed based on the sum and ratio between the first training set proportional coefficient and the first verification set proportional coefficient to obtain the first training set proportional coefficient and the first verification set proportional coefficient. Based on the first training set proportional coefficient, the first verification set proportional coefficient and the preset test set proportional coefficient, multiple original sample data in the original sample data set are randomly divided into three data sets to obtain the first training set, the first verification set and the test set.

[0059] For example, the preset test set ratio coefficient is α%, the number of non-single-layer BP neural network module parameters in the transformer fault detection model is β, and the number of single-layer BP neural network module parameters in the transformer fault detection model is χ. Get the first training set ratio coefficient X 1 % and the first validation set ratio coefficient X 2 %. The number of original sample data in the first training set, the first validation set, and the test set, and the proportion of the multiple original sample data in the original sample data set are respectively X 1 %, X 2 % and α%, where α=20.

[0060] The technical solution of this embodiment determines the first training set proportional coefficient and the first verification set proportional coefficient based on the preset test set proportional coefficient and the number of non-single-layer BP neural network module parameters and the number of single-layer BP neural network module parameters in the transformer fault detection model. It can ensure that the data scale of the first training set matches the number of non-single-layer BP neural network module parameters, and the data scale of the first verification set matches the number of single-layer BP neural network module parameters, ensuring that the training degrees of multiple primary fusion multi-classification modules and single-layer BP neural network modules of the transformer fault detection model are the same during the training process, which helps to ensure the stability of the training.

[0061] S220. Training a first generative adversarial network based on a first training set and a noise generator, wherein the first generative adversarial network includes a first generative network and a first discriminative network. During the training of the first generative adversarial network: adjusting model parameters of the first generative network based on a first generative loss corresponding to the first generative network to obtain an additional data generation model and an additional sample data set generated by the additional data generation model, wherein the first generation loss corresponding to the first generative network includes an additional sample adversarial loss.

[0062] In this embodiment, the additional data generation model is a model for generating an additional sample data set, the first generation network is an additional data generation model in the training process, and the first discrimination network is a model for performing authenticity discrimination on the output data of the first generation network and the original sample data. The noise generator is a model for generating input data of the first generation network. The additional sample data set is a data set generated by the additional data generation model, which is used to perform data expansion on the original sample data set.

[0063] Specifically, the input data of the first generative network is generated based on the noise generator. The output data of the first generative network is obtained by inputting the input data of the first generative network into the first generative network in the first generative adversarial network. The output data of the first generative network is subjected to authenticity discrimination based on the first discriminant network to obtain a first discrimination result. The additional sample adversarial loss corresponding to the first generative network is generated based on the first discrimination result, and the additional sample adversarial loss is used as the first generative loss. Exemplarily, the noise generator is a Gaussian white noise generator. Optionally, the first generative network is a multi-layer BP fully connected layer network; the number of neurons in the output layer of the first generative network is the same as the number of dimensions of the original sample data; the first discriminant network is a multi-layer perceptron neural network; the number of neurons in the input layer of the first discriminant network is the same as the number of dimensions of the original sample data; the number of neurons in the output layer of the first discriminant network is 1.

[0064] In this embodiment, the input layer of the first generation network is used to receive input data of the first generation network generated based on the noise generator, wherein the input data of the first generation network includes data of multiple dimensions, and each dimension of the input data of the first generation network is input into an input layer neuron of the first generation network. The dimension of the output data of the first generation network is the same as the dimension of the original sample data, and the output of each output layer neuron of the first generation network is data of one dimension in the output data of the first generation network. It should be noted that the number of input layer neurons of the first generation network may be the same as or different from the number of output layer neurons of the first generation network, and this embodiment does not limit this.

[0065] For example, Figure 3 is a schematic diagram of the structure of the first generation network applicable to the second embodiment of the present invention, such as Figure 3As shown, the first generation network includes a first fully connected layer, at least one intermediate fully connected layer and a last fully connected layer, wherein the number of neurons in the first fully connected layer is X, the number of intermediate fully connected layers and the number of neurons in each intermediate fully connected layer are set based on experience, and this embodiment does not limit this, and the number of neurons in the last fully connected layer is N+V, wherein X represents the number of dimensions of the input data of the first generation network, and N+V represents the number of dimensions of the original sample data (i.e., the number of dimensions of the output data of the first generation network).

[0066] The input layer of the first discriminant network is used to receive the output data or original sample data of the first generation network, wherein the input data of the first discriminant network includes data of multiple dimensions, and each dimension of the input data of the first discriminant network is input into an input layer neuron of the first discriminant network. The dimension of the output data of the first discriminant network is 1, which is used to characterize that the input data of the first discriminant network is the output data or original sample data of the first generation network.

[0067] For example, Figure 4 is a schematic diagram of the structure of the first discrimination network applicable to the second embodiment of the present invention, such as Figure 4 As shown, the first discriminant network includes a first fully connected layer, at least one intermediate fully connected layer and a last fully connected layer, wherein the number of neurons in the first fully connected layer is N+V, the number of intermediate fully connected layers and the number of neurons in each intermediate fully connected layer are set based on experience, and this embodiment does not limit this, and the number of neurons in the last fully connected layer is 1.

[0068] The technical solution of this embodiment, by setting the number of output layer neurons of the first generation network to be the same as the number of dimensions of the original sample data, and setting the number of input layer neurons of the first discriminant network to be the same as the number of dimensions of the original sample data, can ensure that the dimension of the output data of the first generation network is consistent with the dimension of the original sample data, which helps to ensure the data quality of the output data of the first generation network.

[0069] Specifically, in each training round of the first generative network, a generated sample data set corresponding to the training round is generated based on the first generative network of the training round, and the model parameters of the first generative network are adjusted based on the first generation loss of the training round. In the case where the current training round is the last training round of the first generative network, the first generative network of the current training round is determined as an additional data generation model, and the generated sample data set corresponding to the current training round is saved as an additional sample data set. Optionally, for the first current training batch of the first generative adversarial network, a noise data set is generated based on a noise generator, wherein the noise data set includes B*P noise data; wherein B represents a preset multiple, and P represents a preset batch size; the noise data set is input into the first generative network of the first generative adversarial network to obtain a pseudo sample data set; based on the first discriminant network of the first generative adversarial network, a first discriminant result corresponding to the pseudo sample data set and a second discriminant result corresponding to the current first training set are determined, and the current first training set is determined based on the first training set and the first current training batch; based on the first discriminant result and the second discriminant result, a first discriminant loss corresponding to the first discriminant network is determined.

[0070] In this embodiment, the noise data is the output data of the noise generator and is also the input data of the first generative network. The additional data is the output data of the additional data generation model (i.e., the trained first generative network). The first current training batch is the current training batch of the first generative adversarial network.

[0071] Specifically, in the first current training batch of the first generative adversarial network, there are a training phase of the first generative network and a training phase of the first discriminative network. In the training phase of the first discriminative network, B*P noise data are randomly generated based on the noise generator, and the B*P noise data are stored as a noise data set, and the number of noise data in the noise data set B*P is a preset multiple B of the preset batch size P, wherein the preset batch size is the number of original sample data used in each training batch during the training of the first discriminative network, and the preset multiple is a preset integer.

[0072] For example, represents the i-th noise data in the noise data set, where d ix express The data of the xth dimension in , where X represents the number of dimensions in the noise data. The noise data set is represented as

[0073] The pseudo sample data set is a data set generated by the first generation network. In the training phase of the first discriminant network of the first current training batch, the pseudo sample data set is a data set generated by the first generation network in a non-trainable state; in the training phase of the first generation network of the first current training batch, the pseudo sample data set is a data set generated by the first generation network in a trainable state.

[0074] In the training stage of the first discriminant network, each noise data in the noise data set is input into the first generative network in a non-trainable state of the first generative adversarial network, and the output result of the first generative network is determined as the pseudo sample data corresponding to the noise data, and the pseudo sample data corresponding to the multiple noise data are stored as the pseudo sample data set.

[0075] For example, the pseudo sample dataset is represented as in, Represents the i-th pseudo sample data in the pseudo sample data set, for The corresponding pseudo sample data, where The number of data items in the pseudo sample data is the same as the number of data items in the original sample data. It can be understood that the sum of the number of feature parameter items in the fault sample data and the number of data items in the fault classification label data in the original sample data is equal to the number of data items in the pseudo sample data. i1 , …, f in ,…,f iN for Multiple characteristic parameter items of the fault sample data in, f i(N+1) ,…,f i(N+v) ,…,f i(N+V) for Multiple data items of fault classification label data in .

[0076] Based on the first discriminant network of the first generative adversarial network, a first discriminant result corresponding to the pseudo sample data set and a second discriminant result corresponding to the current first training set are determined, and the current first training set is determined based on the first training set and the first current training batch; based on the first discriminant result and the second discriminant result, a first discriminant loss corresponding to the first discriminant network is determined.

[0077] In this embodiment, the current first training set is the current training batch in the first generative adversarial network, which is input into the first discriminant network in a trainable state. It can be understood that the current first training set includes a preset batch size of original sample data.

[0078] Specifically, in the training round of the first discriminant network, the updated pseudo sample data set is obtained by setting the authenticity label of each pseudo sample data in the pseudo sample data set to 0. The updated current first training set is obtained by setting the authenticity label of each original sample data in the current first training set to 1. Exemplarily, the updated pseudo sample data set is The updated current first training set is

[0079] The first discrimination result is obtained by inputting the updated pseudo sample data set into the first discriminant network in a trainable state, and the second discrimination result is obtained by inputting the updated current first training set into the first discriminant network in a trainable state. The first discrimination loss is calculated based on the first discrimination result and the second discrimination result to obtain the first discrimination loss.

[0080] Exemplarily, the first discriminant loss is expressed as:

[0081] Among them, in the current training batch of the first discriminant network, the first discriminant result corresponding to the i-th pseudo sample data in the pseudo sample data set is DA(f i1 ,…,f in ,…,f iN ,f i(N+1) ,…,f i(N+v) ,…,f i(N+V) ), the second discrimination result corresponding to the i-th original sample data in the current first training set is DA(t i1 ,…,t in ,…,t iN ,b i1 ,…,b iv ,…,b iV ).

[0082] The model parameters of the first generation network are adjusted alternately based on the first generation loss and the model parameters of the first discriminant network are adjusted based on the first discriminant loss to obtain an additional data generation model, and an additional sample data set generated based on the additional data generation model is saved.

[0083] Specifically, for the first current training batch of the first generative adversarial network, in the training stage of the first generative network, the first discriminant network is set to a non-trainable state, and the first generative network is set to a trainable state. By inputting each noise data in the noise data set into the first generative network in the trainable state of the first generative adversarial network, the output result of the first generative network obtained is determined as the pseudo sample data corresponding to the noise data, and the pseudo sample data corresponding to the multiple noise data are stored as a pseudo sample data set. By inputting the pseudo sample data set into the first discriminant network in the non-trainable state, a first discrimination result is obtained, and a first generation loss is calculated based on the first discrimination result to obtain a first generation loss, and the model parameters of the first generative network are adjusted based on the first generation loss. In the training stage of the first discriminant network, the first discriminant network is set to a trainable state, and the first generative network is set to a non-trainable state, and the model parameters of the first discriminant network are adjusted based on the determined first discriminant loss.

[0084] Exemplarily, the first generation loss is expressed as:

[0085] In the current training batch of the first discriminant network, the first discriminant result corresponding to the i-th pseudo sample data in the pseudo sample data set is DA(f i1 ,...,f in ,...,f iN ,f i(N+1) ,...,f i(n+v) ,...,f i(N+V) ).

[0086] If all the original sample data in the first training set have been input into the first generative adversarial network, the current training round is increased by one, and the current training round is compared with the preset total training rounds of the first generative adversarial network. When the current training round is equal to the preset total training rounds of the first generative adversarial network, the current training round can be considered as the last training round of the first generative adversarial network, and the first generative network of the current training round is determined as the additional data generation model. Based on the pseudo sample data sets generated by the first generative adversarial network of the current training round in multiple training batches, an additional data discrimination model corresponding to the additional data generation model is constructed, and the additional data generation model is saved.

[0087] Exemplarily, the additional sample data set is represented as in, Represents the pseudo sample dataset generated by the first generation network of the hth training batch in the last training round, H = X 1 %*I / P represents the number of training batches contained in one training round.

[0088] The technical solution of this embodiment helps to ensure the training stability of the first generative adversarial network by determining the first discriminant loss corresponding to the first discriminant network based on the first discriminant result and the second discriminant result.

[0089] S230: Determine a second training set and a second validation set in the additional sample data set.

[0090] Specifically, a second training set and a second validation set are obtained by randomly dividing a plurality of pseudo sample data in the additional sample data set. The second training set is used to adjust the parameters of the non-single-layer BP neural network module, and the second validation set is used to adjust the parameters of the single-layer BP neural network module. Optionally, based on the number of non-single-layer BP neural network module parameters and the number of single-layer BP neural network module parameters in the transformer fault detection model, the second training set proportional coefficient and the second validation set proportional coefficient are determined; based on the second training set proportional coefficient and the second validation set proportional coefficient, the additional sample data set is divided into a second training set and a second validation set.

[0091] In this embodiment, the second training set proportion coefficient is used to characterize the proportion of the second training set in the additional sample data set, and the second validation set proportion coefficient is used to characterize the proportion of the second validation set in the additional sample data set.

[0092] Specifically, 100% is determined as the sum of the second training set proportional coefficient and the second verification set proportional coefficient, and the ratio of the number of non-single-layer BP neural network module parameters and the number of single-layer BP neural network module parameters in the transformer fault detection model is determined as the ratio between the second training set proportional coefficient and the second verification set proportional coefficient. The second training set proportional coefficient and the second verification set proportional coefficient are analyzed based on the sum and ratio between the second training set proportional coefficient and the second verification set proportional coefficient to obtain the second training set proportional coefficient and the second verification set proportional coefficient. Based on the second training set proportional coefficient and the second verification set proportional coefficient, multiple pseudo sample data in the additional sample data set are randomly divided into two data sets to obtain the second training set and the second verification set.

[0093] For example, assuming that the number of non-single-layer BP neural network module parameters in the transformer fault detection model is β, and the number of single-layer BP neural network module parameters in the transformer fault detection model is χ, then by analyzing Get the second training set ratio coefficient Y 1 % and the second validation set ratio coefficient Y 2 % The number of pseudo sample data in the second training set and the second validation set, and the proportion of the multiple pseudo sample data in the additional sample data set are Y respectively. 1 % and Y 2 %.

[0094] The technical solution of this embodiment determines the second training set proportional coefficient and the second verification set proportional coefficient based on the number of non-single-layer BP neural network module parameters and the number of single-layer BP neural network module parameters in the transformer fault detection model, so as to ensure that the data scale of the second training set matches the number of non-single-layer BP neural network module parameters, and the data scale of the second verification set matches the number of single-layer BP neural network module parameters, thereby ensuring that the training degrees of multiple primary fusion multi-classification modules and single-layer BP neural network modules of the transformer fault detection model are the same during the training process, which helps to ensure the stability of the training.

[0095] S240, training each primary fusion multi-classification module based on the total training set to determine the candidate model parameters of each primary fusion multi-classification module; the total training set includes a first training set and a second training set.

[0096] In this embodiment, the candidate model parameters are model parameters of the primary fusion multi-classification module obtained through training based on the total training set.

[0097] Specifically, the first training set and the second training set are integrated to obtain a total training set. Based on the total training set, each primary fusion multi-classification module in the transformer fault detection model is trained separately to obtain a trained primary fusion multi-classification module, and the model parameters of each trained primary fusion multi-classification module are used as candidate model parameters. Exemplarily, multiple primary fusion multi-classification modules can be trained sequentially or in parallel, which is not limited in this embodiment.

[0098] S250. Training a second generative adversarial network based on the total validation set, wherein the second generative adversarial network includes a second generative network and a second discriminative network; during the training of the second generative adversarial network: each primary fusion multi-classification module in the second generative network is based on the candidate model parameter setting of the primary fusion multi-classification module; based on the second generation loss corresponding to the second generation network, the model parameters of each single-layer BP neural network module in the second generation network are adjusted to obtain a candidate transformer fault detection model, wherein the second generation loss corresponding to the second generation network includes a pseudo-label adversarial loss and a pseudo-label cross entropy loss; the total validation set includes a first validation set and a second validation set.

[0099] In this embodiment, in this embodiment, the candidate transformer fault detection model is a transformer fault detection model obtained by training based on the total training set and the total validation set. The second generation network is a candidate transformer fault detection model in the training process, and each primary fusion multi-classification module in the second generation network is set to the candidate model parameters of the primary fusion multi-classification module obtained based on the total training set. The second discriminant network is a model for authenticity discrimination of the output data of the second generation network and the fault classification label data in the original sample data. Optionally, the second discriminant network is a multilayer perceptron neural network; the number of input layer neurons of the second discriminant network is the same as the number of dimensions of the fault classification label data; the number of output layer neurons of the second discriminant network is 1.

[0100] In this embodiment, the input layer of the second discriminant network is used to receive input data of the second discriminant network, wherein the input data of the second discriminant network is the output data of the second generation network or the fault classification label data in the original sample data. The input data of the second discriminant network includes data of multiple dimensions, and each dimension of the input data of the second discriminant network is input into an input layer neuron of the second discriminant network. The dimension of the output data of the second discriminant network is 1, which is used to characterize that the input data of the second discriminant network is the output data of the second generation network or the original sample data.

[0101] Exemplarily, the second discriminant network is a multi-layer perceptron neural network model. Figure 5 is a schematic diagram of the structure of the second discrimination network applicable to the second embodiment of the present invention, such as Figure 5 As shown, the first generation network includes a first fully connected layer, at least one intermediate fully connected layer and a last fully connected layer, wherein the number of neurons in the first fully connected layer is V, the number of intermediate fully connected layers and the number of neurons in each intermediate fully connected layer are set based on experience, and this embodiment does not limit this, and the number of neurons in the last fully connected layer is 1.

[0102] Specifically, by inputting each fault sample data in the total validation set into the second generation network in the second generation adversarial network, the output data of the second generation network is obtained. The output data of the second generation network is subjected to authenticity discrimination based on the second discriminant network to obtain the first label discrimination result. Based on the first label discrimination results corresponding to the plurality of fault sample data, the pseudo-label adversarial loss corresponding to the second generation network is generated, and based on the output data of the second generation network corresponding to the plurality of fault sample data and the fault classification label data, the pseudo-label cross entropy loss corresponding to the second generation network is generated, and the pseudo-label adversarial loss and the pseudo-label cross entropy loss are combined into the second generation loss. It should be noted that the model parameters of each primary fusion multi-classification module in the second generation network are non-adjustable candidate model parameters, and only the model parameters of the single-layer BP neural network module are in an adjustable state. Based on the second generation loss of the current training round, the model parameters of the single-layer BP neural network module in the second generation network are adjusted. In the case where the current training round is the last training round of the second generation network, the second generation network of the current training round is determined as a candidate transformer fault detection model.

[0103] The technical solution of this embodiment, by setting the number of input layer neurons of the second discriminant network to be the same as the number of dimensions of the fault classification label data, can ensure that the dimension of the output data of the second generation network is consistent with the dimension of the fault classification label data in the original sample data, which helps to ensure the data quality of the output data of the second generation network.

[0104] The candidate transformer fault detection model is a transformer fault detection model obtained by training based on the total training set and the total validation set. The second generation network is a candidate transformer fault detection model in the training process, and each primary fusion multi-classification module in the second generation network is set to the candidate model parameters of the primary fusion multi-classification module obtained based on the total training set. The second discriminant network is a model for authenticity discrimination of the output data of the second generation network and the fault classification label data in the original sample data.

[0105] Specifically, by inputting each fault sample data in the total validation set into the second generation network in the second generation adversarial network, the output data of the second generation network is obtained. The output data of the second generation network is subjected to authenticity discrimination based on the second discriminant network to obtain the first label discrimination result. Based on the first label discrimination results corresponding to the plurality of fault sample data, the pseudo-label adversarial loss corresponding to the second generation network is generated, and based on the output data of the second generation network corresponding to the plurality of fault sample data and the fault classification label data, the pseudo-label cross entropy loss corresponding to the second generation network is generated, and the pseudo-label adversarial loss and the pseudo-label cross entropy loss are combined into the second generation loss. It should be noted that the model parameters of each primary fusion multi-classification module in the second generation network are non-adjustable candidate model parameters, and only the model parameters of the single-layer BP neural network module are in an adjustable state. Based on the second generation loss of the current training round, the model parameters of the single-layer BP neural network module in the second generation network are adjusted. In the case where the current training round is the last training round of the second generation network, the second generation network of the current training round is determined as a candidate transformer fault detection model. Optionally, for the second current training batch of the second generative adversarial network, multiple fault sample data in the current total verification set are input into the second generative network of the second generative adversarial network to obtain a pseudo-fault classification label data set corresponding to the current total verification set; the current total verification set is determined based on the total verification set and the second current training batch; the current total verification set includes P random fault classification label data in the total verification set; based on the second discriminant network of the second generative adversarial network, a first label discrimination result corresponding to the pseudo-fault classification label data set and a second label discrimination result corresponding to the current total verification set are determined; based on the first label discrimination result and the second label discrimination result, a second discriminant loss corresponding to the second discriminant network is determined.

[0106] In this embodiment, the second current training batch is a current training batch of the second generative adversarial network.

[0107] Specifically, in the second current training batch, multiple fault sample data of the current training batch in the total verification set are combined into the current total verification set, and the current total verification set is input into each primary fusion multi-classification module in the second generation network, so that each primary fusion multi-classification module performs multi-classification processing on the multiple fault sample data in the current total verification set respectively, and obtains the classification prediction result corresponding to each primary fusion multi-classification module, wherein the classification prediction result includes the first prediction probability corresponding to each fault sample data belonging to multiple fault type information respectively.

[0108] Exemplarily, in each training batch, P fault sample data in the current total verification set are simultaneously input into the qth primary fusion multi-classification module, and the classification prediction result of the qth primary fusion multi-classification module is expressed as:

[0109] in, is the classification prediction result corresponding to the i-th fault sample data in the current training batch, C i,q,v represents the first predicted probability that the i-th fault sample data in the current training batch belongs to the fault type information v based on the q-th primary fusion multi-classification module, C i,q,v A probability value between 0 and 100%.

[0110] In the second current training batch of the second generative adversarial network, a training phase of the second generative network and a training phase of the second discriminative network are included.

[0111] In the training stage of the second discriminant network, based on the untrainable single-layer BP neural network module in the second generation network, for the classification prediction results of each fault sample data, the first prediction probability corresponding to the same fault type information in the classification prediction results corresponding to the multiple primary fusion multi-classification modules is fused to obtain the first prediction probability corresponding to each fault type information of the fault sample data. Based on the first prediction probability corresponding to the multiple fault type information respectively corresponding to the fault sample data, the pseudo-fault classification label data corresponding to the fault sample data is formed, and it should be noted that the first prediction probability is 0 or 100%. The pseudo-fault classification label data set is obtained by saving the pseudo-fault classification label data corresponding to each fault type information of the multiple fault sample data in the current total verification set.

[0112] Exemplarily, the first prediction probability corresponding to the fault type information v of the i-th fault sample data in the current training batch is expressed as:

[0113] Diagnosis i,v =W v,1 ×C i,1,v +…+W v,q ×C i,q,v +…+W v,Q ×C i,Q,v , then the pseudo fault classification label dataset corresponding to the current total verification set is expressed as:

[0114] in, Represents the pseudo fault classification label data corresponding to the i-th fault sample data in the current training batch.

[0115] By setting the authenticity label of each pseudo fault classification label data in the pseudo fault classification label data set to 0, an updated pseudo fault classification label data set is obtained. By setting the authenticity label of each fault classification label data in the current total verification set to 1, an updated current total verification set is obtained. Exemplarily, the updated pseudo fault classification label data set is The updated total validation set is

[0116] The first label discrimination result is obtained by inputting the updated pseudo-fault classification label data set into the second discriminant network in a trainable state, and the second label discrimination result is obtained by inputting the updated current total verification set into the second discriminant network in a trainable state. The second discrimination loss is calculated based on the first label discrimination result and the second label discrimination result to obtain the second discrimination loss. The model parameters of the second discriminant network are adjusted based on the second discrimination loss.

[0117] Exemplarily, the second discriminative loss is expressed as:

[0118] in, Among them, in the current training batch of the second discriminant network, the first label discrimination result corresponding to the ith pseudo fault classification label data in the pseudo fault classification label data set is DB (Diagnosis i,1 ,…,Diagnosis i,v ,…,Diagnosis i,V ), the second label discrimination result corresponding to the i-th fault classification label data in the current total verification set is DB(b i1 ,…,b iv ,…,b iV ).

[0119] The model parameters of the single-layer BP neural network module in the second generation network are adjusted alternately based on the second generation loss and the model parameters of the second discriminant network are adjusted based on the second discriminant loss to obtain a candidate transformer fault detection model.

[0120] Specifically, for the second current training batch of the second generative adversarial network, in the training stage of the first generative network, the second discriminant network is set to a non-trainable state, and the model parameters of the single-layer BP neural network module in the second generative network are set to a trainable state. By inputting the current total verification set into the second generative network whose model parameters are in a trainable state, a pseudo-fault classification label data set corresponding to the current total verification set is obtained, and by inputting the pseudo-fault classification label data set into the second discriminant network in a non-trainable state, a first label discrimination result corresponding to each pseudo-fault classification label data in the pseudo-fault classification label data set is obtained. Based on the first label discrimination result corresponding to each pseudo-fault classification label data in the pseudo-fault classification label data set, a pseudo-label adversarial loss corresponding to the second generative network is generated, and based on each pseudo-fault classification label data in the pseudo-fault classification label data set and the fault classification label data corresponding to the pseudo-fault classification label data, a pseudo-label cross entropy loss corresponding to the second generative network is generated, and the pseudo-label adversarial loss and the pseudo-label cross entropy loss are combined into a second generative loss. The model parameters of the single-layer BP neural network module in the second generative network are adjusted based on the second generative loss.

[0121] Exemplarily, the second generation loss is expressed as:

[0122] in,

[0123]

[0124] If all fault sample data in the total verification set have been input into the second generative adversarial network, the current training round is increased by one, and the current training round is compared with the preset total training rounds of the second generative adversarial network. When the current training round is equal to the preset total training rounds of the second generative adversarial network, the current training round can be considered as the last training round of the second generative adversarial network, and the second generative network of the current training round is determined as a candidate transformer fault.

[0125] The technical solution of this embodiment helps to ensure the training stability of the second generative adversarial network by determining the second discriminant loss corresponding to the second discriminant network based on the first label discrimination result and the second label discrimination result.

[0126] S260. Verify the candidate transformer fault detection model based on the test set. If the verification fails, retrain the first generative adversarial network, multiple primary fusion multi-classification modules and the second generative adversarial network until a trained transformer fault detection model is obtained; the trained transformer fault detection model is the verified candidate transformer fault detection model.

[0127] Specifically, for each fault sample data in the test set, the fault sample data is input into the candidate transformer fault detection model to obtain the predicted fault type information corresponding to the fault sample data, and the detection accuracy of the candidate transformer fault detection model is determined based on the predicted fault type information and fault classification label data corresponding to multiple fault sample data. By comparing the detection accuracy with the preset accuracy threshold, if the detection accuracy is greater than or equal to the preset accuracy threshold, the detection accuracy of the candidate transformer fault detection model can be considered accurate, and the candidate transformer fault detection model is determined as a trained transformer fault detection model; if the detection accuracy is less than the preset accuracy threshold, the detection accuracy of the candidate transformer fault detection model can be considered inaccurate, and the first generative adversarial network, multiple primary fusion multi-classification modules and the second generative adversarial network are retrained until the detection accuracy of the re-obtained candidate transformer fault detection model is greater than or equal to the preset accuracy threshold. The preset accuracy threshold is a threshold of the detection accuracy set in advance, which is used to determine whether the verification is passed. Exemplarily, the preset accuracy threshold is 98%.

[0128] The technical solution of this embodiment can ensure that the trained transformer fault detection model corresponds to the transformer to be detected by obtaining the original sample data set corresponding to the transformer to be detected. By saving the additional sample data set generated by the additional data generation model, the data set used to train the transformer fault detection model can be expanded to improve the generalization of the transformer fault detection model. By determining the candidate model parameters of each primary fusion multi-classification module based on the total training set, and setting each primary fusion multi-classification module in the second generation network based on the candidate model parameters of the primary fusion multi-classification module, the amount of calculation in the training process of the second generation adversarial network can be reduced, and the training speed of the second generation adversarial network can be improved. By verifying the candidate transformer fault detection model based on the test set, the accuracy of the transformer fault detection model can be further improved.

[0129] Embodiment 3

[0130] Figure 6 Schematic diagram of the structure of a transformer fault detection device provided by the third embodiment of the present invention. Figure 6 As shown, the device comprises:

[0131] The equipment working data and detection model acquisition module 310 is used to obtain the equipment working data of the transformer to be detected, and obtain the transformer fault detection model corresponding to the transformer to be detected; the transformer fault detection model includes multiple primary fusion multi-classification modules and multiple single-layer BP neural network modules, the primary fusion multi-classification module includes multiple basic multi-classifiers, and at least one of the classifier type, number of classifiers and combination mode corresponding to the multiple basic multi-classifiers included in different primary fusion multi-classification modules is different; each single-layer BP neural network module corresponds to a fault type information;

[0132] The fault type information determination module 320 is used to perform fault detection on the equipment working data based on the transformer fault detection model to obtain the fault type information of the transformer to be detected. In the fault detection process: for each primary fusion multi-classification module, multi-classification processing is performed on the equipment working data based on each basic multi-classifier in the primary fusion multi-classification module to obtain the basic multi-classification result corresponding to the basic multi-classifier, and based on the combination method corresponding to the multiple basic multi-classifiers included in the primary fusion multi-classification module, the basic multi-classification results corresponding to the multiple basic multi-classifiers are aggregated to obtain the classification result corresponding to the primary fusion multi-classification module; the classification result includes the first probabilities corresponding to the multiple fault type information respectively of the transformer to be detected; for each fault type information, based on the single-layer BP neural network module corresponding to the fault type information, the first probabilities corresponding to the fault type information in the classification results corresponding to the multiple primary fusion multi-classification modules are calculated and processed to obtain the second probability that the transformer to be detected belongs to each fault type information; the fault type information corresponding to the maximum second probability is determined as the fault type information of the transformer to be detected; the maximum second probability is the maximum probability among the second probabilities corresponding to the multiple fault type information.

[0133] The technical solution of this embodiment, by setting multiple primary fusion multi-classification modules and multiple single-layer BP neural network modules in the transformer fault detection model, and making each single-layer BP neural network module correspond to a fault type information, instead of setting only one primary fusion multi-classification module, can calculate and process the first probabilities corresponding to the fault type information in the classification results corresponding to the multiple primary fusion multi-classification modules based on the single-layer BP neural network module, and obtain the second probability that the transformer to be detected belongs to each fault type information, so that the first probabilities corresponding to the fault type information in the classification results corresponding to different primary fusion multi-classification modules are complementary, which solves the problem of inaccuracy of the transformer fault classification results, improves the robustness of the transformer fault detection model, and helps the transformer fault detection model to extract the complex relationship in the equipment working parameters, thereby improving the accuracy and stability of transformer fault detection.

[0134] On the basis of the above embodiment, optionally, the single-layer BP neural network module includes multiple fusion neurons, each fusion neuron corresponds to a primary fusion multi-classification module; accordingly, the fault type information determination module 320 is specifically used to: for each fault type information, based on the weights corresponding to the multiple fusion neurons in the single-layer BP neural network module corresponding to the fault type information, perform weighted summation processing on the first probabilities corresponding to the fault type information in the classification results corresponding to the multiple primary fusion multi-classification modules, to obtain the second probability that the transformer to be detected belongs to each fault type information; wherein the second probability that the transformer to be detected belongs to the fault type information is expressed as: Diagnosis v =W v,1 ×C 1,v +…+W v,q ×C q,v +…+W v,Q ×C Q,v , where W v,q represents the weight corresponding to the qth fusion neuron in the vth single-layer BP neural network module, C q,v It represents the first probability corresponding to the vth fault type information in the classification result corresponding to the qth primary fusion multi-classification module; v∈[1,V], V represents the number of fault type information, q∈[1,Q], Q represents the number of primary fusion multi-classification modules.

[0135] On the basis of the above embodiment, optionally, the transformer fault detection device also includes: a transformer fault detection model training module; wherein the transformer fault detection model training module is used to: obtain the original sample data set corresponding to the transformer to be detected, and determine the first training set, the first verification set and the test set in the original sample data set; the original sample data set includes multiple original sample data, each original sample data includes fault sample data and fault classification label data corresponding to the fault sample data; based on the first training set and the noise generator, the first generative adversarial network is trained, wherein the first generative adversarial network includes a first generative network and a first discriminant network; based on the first generation loss corresponding to the first generation network, the model parameters of the first generation network are adjusted to obtain an additional data generation model and an additional sample data set generated by the additional data generation model, wherein the first generation loss corresponding to the first generation network includes an additional sample adversarial loss; determine the second training set and the second verification set in the additional sample data set; based on the total training set, each primary fusion multi-classification module is trained to determine The candidate model parameters of each primary fusion multi-classification module; the total training set includes a first training set and a second training set; the second generative adversarial network is trained based on the total verification set, wherein the second generative adversarial network includes a second generative network and a second discriminant network, and each primary fusion multi-classification module in the second generative network is set based on the candidate model parameters of the primary fusion multi-classification module; the model parameters of each single-layer BP neural network module in the second generative network are adjusted based on the second generation loss corresponding to the second generation network to obtain a candidate transformer fault detection model, wherein the second generation loss corresponding to the second generation network includes a pseudo-label adversarial loss and a pseudo-label cross entropy loss; the total verification set includes a first verification set and a second verification set; the candidate transformer fault detection model is verified based on the test set, and if the verification fails, the first generation adversarial network, multiple primary fusion multi-classification modules and the second generation adversarial network are retrained until a trained transformer fault detection model is obtained; the trained transformer fault detection model is a candidate transformer fault detection model that has passed the verification.

[0136] On the basis of the above embodiment, optionally, the transformer fault detection model training module is specifically used to: determine the first training set proportional coefficient and the first verification set proportional coefficient based on the preset test set proportional coefficient and the number of non-single-layer BP neural network module parameters and the number of single-layer BP neural network module parameters in the transformer fault detection model; divide the original sample data set into a first training set, a first verification set and a test set based on the first training set proportional coefficient, the first verification set proportional coefficient and the preset test set proportional coefficient; the preset test set proportional coefficient is 20%; accordingly, based on the number of non-single-layer BP neural network module parameters and the number of single-layer BP neural network module parameters in the transformer fault detection model, determine the second training set proportional coefficient and the second verification set proportional coefficient; based on the second training set proportional coefficient and the second verification set proportional coefficient, divide the additional sample data set into a second training set and a second verification set.

[0137] On the basis of the above embodiments, optionally, a transformer fault detection model training module is specifically used to: generate a noise data set based on a noise generator for the first current training batch of the first generative adversarial network, wherein the noise data set includes B*P noise data; wherein B represents a preset multiple, and P represents a preset batch size; input the noise data set into the first generative network of the first generative adversarial network to obtain a pseudo sample data set; determine a first discrimination result corresponding to the pseudo sample data set and a second discrimination result corresponding to the current first training set based on the first discriminant network of the first generative adversarial network, and the current first training set is determined based on the first training set and the first current training batch; determine a first discrimination loss corresponding to the first discriminant network based on the first discrimination result and the second discrimination result; accordingly, alternately adjust the model parameters of the first generative network based on the first generation loss and adjust the model parameters of the first discriminant network based on the first discrimination loss to obtain an additional data generation model, and save the additional sample data set generated based on the additional data generation model.

[0138] On the basis of the above embodiments, optionally, a transformer fault detection model training module is specifically used to: for the second current training batch of the second generative adversarial network, input multiple fault sample data in the current total verification set into the second generative network of the second generative adversarial network to obtain a pseudo-fault classification label data set corresponding to the current total verification set; the current total verification set is determined based on the total verification set and the second current training batch; the current total verification set includes P random fault classification label data in the total verification set; based on the second discriminant network of the second generative adversarial network, determine the first label discrimination result corresponding to the pseudo-fault classification label data set and the second label discrimination result corresponding to the current total verification set; determine the second discriminant loss corresponding to the second discriminant network based on the first label discrimination result and the second label discrimination result; accordingly, alternately adjust the model parameters of the single-layer BP neural network module in the second generative network based on the second generation loss and adjust the model parameters of the second discriminant network based on the second discrimination loss to obtain a candidate transformer fault detection model.

[0139] Based on the above embodiments, optionally, the first generation network is a multi-layer BP fully connected layer network; the number of output layer neurons of the first generation network is the same as the dimension of the original sample data; the first discriminant network is a multi-layer perceptron neural network; the number of input layer neurons of the first discriminant network is the same as the number of dimensions of the original sample data; the number of output layer neurons of the first discriminant network is 1; the second discriminant network is a multi-layer perceptron neural network; the number of input layer neurons of the second discriminant network is the same as the number of dimensions of the fault classification label data; the number of output layer neurons of the second discriminant network is 1.

[0140] The transformer fault detection device provided in the embodiment of the present invention can execute the transformer fault detection method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0141] Embodiment 4

[0142] Figure 7 1 is a schematic diagram of the structure of an electronic device for implementing the transformer fault detection method of an embodiment of the present invention. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices (such as helmets, glasses, watches, etc.) and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or required herein.

[0143] like Figure 7As shown, the electronic device 10 includes at least one processor 11, and a memory connected to the at least one processor 11, such as a read-only memory (ROM) 12, a random access memory (RAM) 13, etc., wherein the memory stores a computer program that can be executed by at least one processor, and the processor 11 can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 12 or the computer program loaded from the storage unit 18 to the random access memory (RAM) 13. In the RAM 13, various programs and data required for the operation of the electronic device 10 can also be stored. The processor 11, the ROM 12, and the RAM 13 are connected to each other through a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0144] A number of components in the electronic device 10 are connected to the I / O interface 15, including: an input unit 16, such as a keyboard, a mouse, etc.; an output unit 17, such as various types of displays, speakers, etc.; a storage unit 18, such as a disk, an optical disk, etc.; and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 19 allows the electronic device 10 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0145] The processor 11 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The processor 11 executes the various methods and processes described above, such as a transformer fault detection method.

[0146] In some embodiments, the transformer fault detection method may be implemented as a computer program, which is tangibly contained in a computer-readable storage medium, such as a storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 10 via the ROM 12 and / or the communication unit 19. When the computer program is loaded into the RAM 13 and executed by the processor 11, one or more steps of the transformer fault detection method described above may be performed. Alternatively, in other embodiments, the processor 11 may be configured to perform the transformer fault detection method in any other appropriate manner (e.g., by means of firmware).

[0147] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0148] The computer program for implementing the transformer fault detection method of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that when the computer program is executed by the processor, the functions / operations specified in the flow chart and / or block diagram are implemented. The computer program can be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0149] Embodiment 5

[0150] Embodiment 5 of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to execute a transformer fault detection method, the method comprising:

[0151] Acquire the equipment working data of the transformer to be detected, and acquire the transformer fault detection model corresponding to the transformer to be detected; the transformer fault detection model includes multiple primary fusion multi-classification modules and multiple single-layer BP neural network modules, the primary fusion multi-classification module includes multiple basic multi-classifiers, and at least one of the classifier type, the number of classifiers and the combination method corresponding to the multiple basic multi-classifiers included in different primary fusion multi-classification modules is different; each of the single-layer BP neural network modules corresponds to a fault type information; based on the transformer fault detection model, perform fault detection on the equipment working data to obtain the fault type information of the transformer to be detected. In the fault detection process: for each of the primary fusion multi-classification modules, perform multi-classification processing on the equipment working data based on each of the basic multi-classifiers in the primary fusion multi-classification module to obtain the basic multi-classifier corresponding to the basic multi-classifier The primary fusion multi-classification module comprises a plurality of basic multi-classifiers, and based on the combination mode corresponding to the plurality of basic multi-classifiers included in the primary fusion multi-classification module, the basic multi-classification results corresponding to the plurality of basic multi-classifiers are fused to obtain the classification result corresponding to the primary fusion multi-classification module; the classification result comprises the first probabilities corresponding to the plurality of fault type information respectively corresponding to the transformer to be detected; for each of the fault type information, based on the single-layer BP neural network module corresponding to the fault type information, the first probabilities corresponding to the fault type information in the classification results corresponding to the plurality of primary fusion multi-classification modules are calculated to obtain the second probability that the transformer to be detected belongs to each of the fault type information; the fault type information corresponding to the maximum second probability is determined as the fault type information of the transformer to be detected; the maximum second probability is the maximum probability among the second probabilities corresponding to the plurality of fault type information respectively.

[0152] In the context of the present invention, a computer-readable storage medium may be a tangible medium that may contain or store a computer program for use by or in combination with an instruction execution system, device or equipment. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0153] To provide interaction with a user, the systems and techniques described herein may be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or trackball) through which the user can provide input to the electronic device. Other types of devices may also be used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form (including acoustic input, voice input, or tactile input).

[0154] The systems and techniques described herein may be implemented in a computing system that includes a backend (e.g., as a data server), or a computing system that includes middleware (e.g., an application server), or a computing system that includes a frontend (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backends, middleware, or frontends. The systems may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0155] A computing system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The client and server relationship is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services.

[0156] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of the present invention can be achieved, and this document does not limit this.

[0157] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A transformer fault detection method, characterized in that: include: Acquire equipment working data of the transformer to be detected, and acquire a transformer fault detection model corresponding to the transformer to be detected; The transformer fault detection model includes a plurality of primary fusion multi-classification modules and a plurality of single-layer BP neural network modules, wherein the primary fusion multi-classification module includes a plurality of basic multi-classifiers, and at least one of the classifier type, the number of classifiers and the combination mode corresponding to the plurality of basic multi-classifiers included in different primary fusion multi-classification modules is different; Each of the single-layer BP neural network modules corresponds to a piece of fault type information; Based on the transformer fault detection model, fault detection is performed on the equipment working data to obtain the fault type information of the transformer to be detected. In the fault detection process: For each of the primary fusion multi-classification modules, multi-classification processing is performed on the equipment working data based on each of the basic multi-classifiers in the primary fusion multi-classification module to obtain a basic multi-classification result corresponding to the basic multi-classifier, and based on the combination mode corresponding to the multiple basic multi-classifiers included in the primary fusion multi-classification module, the basic multi-classification results corresponding to the multiple basic multi-classifiers are fused to obtain the classification result corresponding to the primary fusion multi-classification module; the classification result includes the first probabilities corresponding to the multiple fault type information respectively belonging to the transformer to be detected; For each of the fault type information, based on the single-layer BP neural network module corresponding to the fault type information, the first probabilities corresponding to the fault type information in the classification results respectively corresponding to the multiple primary fusion multi-classification modules are calculated and processed to obtain the second probability that the transformer to be detected belongs to each of the fault type information; The fault type information corresponding to the maximum second probability is determined as the fault type information of the transformer to be detected; the maximum second probability is the maximum probability among the second probabilities corresponding to the multiple fault type information.

2. The method according to claim 1, characterized in that The single-layer BP neural network module includes a plurality of fusion neurons, each fusion neuron corresponds to one of the primary fusion multi-classification modules; accordingly, For each of the fault type information, based on the single-layer BP neural network module corresponding to the fault type information, the first probabilities corresponding to the fault type information in the classification results respectively corresponding to the multiple primary fusion multi-classification modules are calculated and processed to obtain the second probability that the transformer to be detected belongs to the fault type information, including: For each of the fault type information, based on the weights corresponding to the multiple fusion neurons in the single-layer BP neural network module corresponding to the fault type information, weighted sum processing is performed on the first probabilities corresponding to the fault type information in the classification results corresponding to the multiple primary fusion multi-classification modules, so as to obtain the second probability that the transformer to be detected belongs to each of the fault type information; The second probability that the transformer to be detected belongs to the fault type information is expressed as: Diagnosis v =W v,1 ×C 1,v +…+W v,q ×C q,v +…+W v,Q ×C Q,v , where W v,q represents the weight corresponding to the qth fusion neuron in the vth single-layer BP neural network module, C q,v Represents the first probability corresponding to the vth fault type information in the classification result corresponding to the qth primary fusion multi-classification module; v∈[1,V], V represents the number of the fault type information, q∈[1,Q], Q represents the number of the primary fusion multi-classification modules.

3. The method according to claim 1, characterized in that The training method of the transformer fault detection model comprises: Obtain an original sample data set corresponding to the transformer to be detected, and determine a first training set, a first validation set, and a test set in the original sample data set; the original sample data set includes a plurality of original sample data, each of which includes fault sample data and fault classification label data corresponding to the fault sample data; Based on the first training set and the noise generator, a first generative adversarial network is trained, wherein the first generative adversarial network includes a first generative network and a first discriminative network; during the training of the first generative adversarial network: based on a first generative loss corresponding to the first generative network, a model parameter of the first generative network is adjusted to obtain an additional data generation model and an additional sample data set generated by the additional data generation model, wherein the first generation loss corresponding to the first generative network includes an additional sample adversarial loss; Determine a second training set and a second validation set in the additional sample data set; Training each of the primary fusion multi-classification modules based on a total training set to determine candidate model parameters of each of the primary fusion multi-classification modules; the total training set includes the first training set and the second training set; The second generative adversarial network is trained based on the total validation set, wherein the second generative adversarial network includes a second generative network and a second discriminative network; during the training process of the second generative adversarial network: each of the primary fusion multi-classification modules in the second generative network is set based on the candidate model parameters of the primary fusion multi-classification module; the model parameters of each of the single-layer BP neural network modules in the second generative network are adjusted based on the second generation loss corresponding to the second generation network to obtain a candidate transformer fault detection model, wherein the second generation loss corresponding to the second generation network includes a pseudo-label adversarial loss and a pseudo-label cross entropy loss; the total validation set includes the first validation set and the second validation set; The candidate transformer fault detection model is verified based on the test set. If the verification fails, the first generative adversarial network, the multiple primary fusion multi-classification modules and the second generative adversarial network are retrained until the trained transformer fault detection model is obtained; the trained transformer fault detection model is the verified candidate transformer fault detection model.

4. The method according to claim 3, characterized in that The determining of the first training set, the first validation set, and the test set in the original sample data set comprises: Determine a first training set proportionality coefficient and a first validation set proportionality coefficient based on a preset test set proportionality coefficient and the number of non-single-layer BP neural network module parameters and the number of single-layer BP neural network module parameters in the transformer fault detection model; Based on the first training set ratio coefficient, the first validation set ratio coefficient and the preset test set ratio coefficient, the original sample data set is divided into the first training set, the first validation set and the test set; the preset test set ratio coefficient is 20%; The determining of the second training set and the second validation set in the additional sample data set comprises: Determine a second training set proportionality coefficient and a second validation set proportionality coefficient based on the number of the non-single-layer BP neural network module parameters and the number of the single-layer BP neural network module parameters in the transformer fault detection model; Based on the second training set ratio coefficient and the second validation set ratio coefficient, the additional sample data set is divided into the second training set and the second validation set.

5. The method according to claim 3, characterized in that: The step of training a first generative adversarial network based on the first training set and the noise generator includes: For a first current training batch of the first generative adversarial network, generate a noise data set based on the noise generator, wherein the noise data set includes B*P noise data; wherein B represents a preset multiple, and P represents a preset batch size; Inputting the noise data set into the first generative network of the first generative adversarial network to obtain a pseudo sample data set; Determine, based on a first discriminant network of the first generative adversarial network, a first discriminant result corresponding to the pseudo sample data set and a second discriminant result corresponding to a current first training set, wherein the current first training set is determined based on the first training set and the first current training batch; Determine a first discriminant loss corresponding to the first discriminant network based on the first discriminant result and the second discriminant result; Accordingly, the adjusting the model parameters of the first generation network based on the first generation loss corresponding to the first generation network to obtain the additional data generation model and the additional sample data set generated by the additional data generation model includes: Alternately adjusting the model parameters of the first generation network based on the first generation loss and adjusting the model parameters of the first discriminant network based on the first discriminant loss to obtain an additional data generation model, and saving an additional sample data set generated based on the additional data generation model.

6. The method according to claim 3, characterized in that The training of the second generative adversarial network based on the total verification set includes: For the second current training batch of the second generative adversarial network, a plurality of fault sample data in the current total verification set are input into the second generative network of the second generative adversarial network to obtain a pseudo fault classification label data set corresponding to the current total verification set; the current total verification set is determined based on the total verification set and the second current training batch; the current total verification set includes P random fault classification label data in the total verification set; Determine, based on a second discriminant network of the second generative adversarial network, a first label discrimination result corresponding to the pseudo-fault classification label data set and a second label discrimination result corresponding to the current total verification set; Determine a second discriminant loss corresponding to the second discriminant network based on the first label discriminant result and the second label discriminant result; Correspondingly, the model parameters of the single-layer BP neural network module in the second generation network are adjusted based on the second generation loss corresponding to the second generation network to obtain a candidate transformer fault detection model, including: The candidate transformer fault detection model is obtained by alternately adjusting the model parameters of the single-layer BP neural network module in the second generation network based on the second generation loss and adjusting the model parameters of the second discrimination network based on the second discrimination loss.

7. The method according to claim 3, characterized in that The first generation network is a multi-layer BP fully connected layer network; the number of neurons in the output layer of the first generation network is the same as the dimension of the original sample data; The first discriminant network is a multilayer perceptron neural network; the number of neurons in the input layer of the first discriminant network is the same as the number of dimensions of the original sample data; the number of neurons in the output layer of the first discriminant network is 1; The second discriminant network is a multilayer perceptron neural network; the number of neurons in the input layer of the second discriminant network is the same as the number of dimensions of the fault classification label data; and the number of neurons in the output layer of the second discriminant network is 1.

8. A transformer fault detection device, characterized in that: include: The equipment working data and detection model acquisition module is used to obtain the equipment working data of the transformer to be detected, and obtain the transformer fault detection model corresponding to the transformer to be detected; the transformer fault detection model includes multiple primary fusion multi-classification modules and multiple single-layer BP neural network modules, the primary fusion multi-classification module includes multiple basic multi-classifiers, and at least one of the classifier type, the number of classifiers and the combination method corresponding to the multiple basic multi-classifiers included in different primary fusion multi-classification modules is different; Each of the single-layer BP neural network modules corresponds to a piece of fault type information; A fault type information determination module is used to perform fault detection on the equipment working data based on the transformer fault detection model to obtain the fault type information of the transformer to be detected. In the fault detection process: for each of the primary fusion multi-classification modules, multi-classification processing is performed on the equipment working data based on each of the basic multi-classifiers in the primary fusion multi-classification module to obtain the basic multi-classification results corresponding to the basic multi-classifiers, and based on the combination mode corresponding to the multiple basic multi-classifiers included in the primary fusion multi-classification module, the basic multi-classification results corresponding to the multiple basic multi-classifiers are aggregated to obtain the basic multi-classification results of the primary fusion multi-classification module. The corresponding classification results; the classification results include the first probabilities that the transformer to be detected belongs to the multiple fault type information respectively; for each of the fault type information, based on the single-layer BP neural network module corresponding to the fault type information, the first probabilities corresponding to the fault type information in the classification results corresponding to the multiple primary fusion multi-classification modules are calculated and processed to obtain the second probability that the transformer to be detected belongs to each of the fault type information; the fault type information corresponding to the maximum second probability is determined as the fault type information of the transformer to be detected; the maximum second probability is the maximum probability among the second probabilities corresponding to the multiple fault type information.

9. An electronic device, characterized in that: The electronic device comprises: at least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the transformer fault detection method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the transformer fault detection method according to any one of claims 1 to 7 when executed.

Citation Information

Patent Citations

  • Multi-sensor signal fusion technology-based fault diagnosis method for wind turbine blades

    CN102944416A

  • Oil-immersed transformer fault diagnosis method based on fault probability distribution model

    CN103745119A

  • Transformer fault diagnosis method based on improved multi-class AdaBoost

    CN108229581A

  • A high-resolution remote sensing image multi-classifier joint classification method based on OpenStreetMap

    CN109034233A

  • Transformer fault diagnosis method based on GoogleNet model

    CN109765333A