A method, apparatus and storage medium for diagnosing transformer faults

Abstract

This invention relates to a transformer fault diagnosis method, apparatus, and storage medium. The invention employs a hybrid mutual information method to select feature attributes strongly correlated with transformer faults from the available attributes. Based on these selected feature attributes, labeled and unlabeled training sample sets are obtained for training the mapping relationship between transformer fault types and feature attributes. The dimensionality of the samples in both sets is significantly reduced, avoiding the curse of dimensionality. Semi-supervised co-training using the labeled and unlabeled training sample sets yields a two-classifier containing the mapping relationship between label values ​​and feature attributes. Data from the transformer whose fault is to be analyzed is input into the two final classifiers. The fault analysis result is obtained based on the average classification value generated by the two classifiers. This invention requires a low number of labeled training samples during training, thus achieving high model training efficiency.

Description

Technical Field

[0001] This invention relates to the field of transformer fault analysis, and in particular to a transformer fault diagnosis method, device and storage medium. Background Technology

[0002] Transformer fault diagnosis is a crucial basis for transformer operation and maintenance. Common methods include manual analysis, which requires maintenance personnel to obtain transformer protection signals and telemetry data and then analyze them based on experience to determine the fault category. Currently, research is utilizing machine learning algorithms and large amounts of historical data to train and construct fault diagnosis models that map the relationship between transformer protection signals and telemetry data and transformer fault types. This enables automatic fault type diagnosis and significantly improves the efficiency of fault analysis and processing. However, existing solutions using machine learning to build fault diagnosis models have several drawbacks: Firstly, transformer protection signals and telemetry data have numerous parameters. Using all parameters as features in the fault diagnosis model would result in an extremely high computational load, potentially leading to the curse of dimensionality. Secondly, in actual transformer operation, fault states are far fewer than normal operating states, resulting in a limited number of labeled fault samples. Obtaining a highly accurate fault diagnosis model requires an extremely large amount of data. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, or at least partially solve the above-mentioned technical problems, the present invention provides a transformer fault diagnosis method, device and storage medium.

[0004] In a first aspect, the present invention provides a transformer fault diagnosis method, comprising: constructing a transformer fault diagnosis sample library, wherein the transformer fault diagnosis sample library stores tagged samples containing attributes and tags and untagged samples containing only attributes, wherein the attributes include transformer protection signals and telemetry data;

[0005] The hybrid mutual information method is used to select feature attributes that are highly correlated with transformer faults from the attributes, and labeled training sample sets and unlabeled training sample sets are obtained based on the selected feature attributes to train the mapping relationship between transformer fault types and feature attributes.

[0006] Semi-supervised co-training using labeled and unlabeled training sample sets yields a two-classifier that includes the mapping relationship between label values ​​and feature attributes.

[0007] The data of the transformer whose fault is to be analyzed is input into the two classifiers that are finally obtained. The fault analysis result is obtained based on the average value of the classification generated by the two classifiers.

[0008] Furthermore, transformer fault operation records of a preset duration are selected. For each fault, a multidimensional attribute vector composed of various transformer protection signals and various telemetry data is combined with a label representing the fault type to form a labeled sample. Then, transformer operation records within a preset duration are selected uniformly and randomly to obtain a set number of unlabeled samples containing only multidimensional attribute vectors. The labeled and unlabeled samples constitute a transformer fault diagnosis sample library.

[0009] Furthermore, the hybrid mutual information method is used to select characteristic attributes that are highly correlated with transformer faults from the attributes, including:

[0010] The pseudo-label values ​​of unlabeled samples are estimated using the KNN nearest neighbor algorithm; the label values ​​of labeled samples and the pseudo-label values ​​of unlabeled samples are combined to form a label set Y;

[0011] Initialize an empty set as the feature attribute set X to store the feature attributes;

[0012] Iterate through each attribute △s in the attribute vector △S j Calculate attribute △s j The mutual information between the label set Y and the attribute △s corresponding to the maximum mutual information is selected. * Add it to the feature attribute set X and remove it from the attribute vector △S;

[0013] Iterate through each attribute △s in the attribute vector △S. j Calculate attribute △s j The mixed mutual information with the label set Y is used to determine the attribute △s corresponding to the maximum mixed mutual information. ** Add it to the feature attribute set X and remove it from the attribute vector △S;

[0014] The loop ends when the number of elements in the feature attribute set X reaches a preset threshold.

[0015] Furthermore, after obtaining the feature attribute set X, labeled training sample sets and unlabeled training sample sets are formed by retaining the attributes in labeled and unlabeled samples according to the elements in the feature attribute set X; so as to obtain labeled training sample sets and unlabeled training sample sets for training the mapping relationship between transformer fault types and feature attributes based on the selected feature attributes.

[0016] Furthermore, semi-supervised co-training using labeled and unlabeled training sample sets yields a two-classifier that includes the mapping relationship between label values ​​and feature attributes, including:

[0017] The labeled training sample set is divided into a first sub-labeled training sample set and a second sub-labeled training sample set. The classifier is trained using the first sub-labeled training sample set to obtain a first classifier, and the classifier is trained using the second sub-labeled training sample set to obtain a second classifier. Then, the first and second classifiers are used to estimate a subset of unlabeled training samples randomly drawn from the unlabeled training sample set. Several unlabeled training samples with high confidence and their pseudo-label values ​​are selected from the subset of unlabeled training samples of another classifier and added to the subset of labeled training samples of another classifier to train and update the classifier. The training process is repeated until the maximum number of iterations is reached or all unlabeled training samples are selected into the subset of labeled training samples.

[0018] Furthermore, the classifier employs a support vector machine.

[0019] Furthermore, the confidence level is taken as the mean squared error of the first or second classifier obtained before and after adding the unlabeled training samples to the labeled training sample set.

[0020] Furthermore, the number n of unlabeled training samples with high confidence levels selected each time... t It changes dynamically with the number of iterations, as shown below:

[0021] n t =n t-1 ρ 0.5t

[0022] Where t is the current iteration number and ρ is the dynamic screening growth rate.

[0023] In a second aspect, a transformer fault diagnosis device includes: at least one processing unit, an interface unit, a bus unit, and a storage unit, wherein the processing unit, the interface unit, and the storage unit are connected to the bus unit, the storage unit stores computer programs and data, and the processing unit reads the computer programs, processes the data, and implements the transformer fault diagnosis method.

[0024] Thirdly, a storage medium for implementing a transformer fault diagnosis method, wherein the storage medium for implementing the transformer fault diagnosis method stores at least one instruction, and a processor reads and executes the instruction to implement the transformer fault diagnosis method.

[0025] The technical solutions provided in the embodiments of the present invention have the following advantages compared with the prior art:

[0026] This invention employs a hybrid mutual information method to select feature attributes highly correlated with transformer faults from the attributes. Based on the selected feature attributes, labeled and unlabeled training sample sets are obtained for training the mapping relationship between transformer fault types and feature attributes. The dimensionality of the samples in the labeled and unlabeled training sample sets is significantly reduced, greatly reducing the computational cost of training a two-classifier that includes the mapping relationship between label values ​​and feature attributes, thus avoiding the curse of dimensionality. Considering that in actual transformer operation, fault states are far less frequent than normal operating states, resulting in a shortage of labeled samples for fault classes, this invention uses semi-supervised co-training of labeled and unlabeled training sample sets to obtain a two-classifier that includes the mapping relationship between label values ​​and feature attributes. The semi-supervised co-training randomly selects a subset of unlabeled training samples from the unlabeled training sample set, analyzes it using the classifier, and selects the unlabeled training sample with the highest confidence and its pseudo-label value from the subset of unlabeled training samples for another classifier to train and update the classifier. This reduces the need for a large number of labeled training samples, thus achieving higher model training efficiency. Attached Figure Description

[0027] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 A flowchart of a transformer fault diagnosis method provided in an embodiment of the present invention;

[0030] Figure 2 The flowchart illustrates how a hybrid mutual information method is used to select feature attributes that are strongly correlated with transformer faults from attributes, and how labeled training sample sets and unlabeled training sample sets are obtained based on the selected feature attributes to train the mapping relationship between transformer fault types and feature attributes.

[0031] Figure 3 A flowchart illustrating how a two-classifier containing the mapping relationship between label values ​​and feature attributes is obtained through semi-supervised co-training using labeled and unlabeled training sample sets, as provided in an embodiment of the present invention.

[0032] Figure 4The flowchart of the present invention provides an embodiment of the process architecture for obtaining a two-classifier containing the mapping relationship between label values ​​and feature attributes by performing semi-supervised co-training using labeled training sample sets and unlabeled training sample sets;

[0033] Figure 5 This is a schematic diagram of a transformer fault diagnosis device provided in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0036] Example 1

[0037] See Figure 1 As shown, an embodiment of the present invention provides a transformer fault diagnosis method, including:

[0038] S100, construct a transformer fault diagnosis sample library, which stores tagged samples containing attributes and labels and untagged samples containing only attributes, the attributes including transformer protection signals and telemetry data.

[0039] In the specific implementation process, transformer fault operation records of a preset duration are selected. For each fault, a 108-dimensional attribute vector ΔS, consisting of 95 types of transformer protection signals and 13 types of telemetry data, is combined with a label representing the fault type to form a labeled sample, and N is obtained. L Each sample is labeled. Then, the transformer's operation records are randomly selected uniformly over a preset time period to obtain a set number N. U N unlabeled samples containing only multidimensional attribute vectors L Each labeled sample and N U The unlabeled samples constitute a transformer fault diagnosis sample library.

[0040] The fault types include: high temperature overheating, normal, high energy discharge, medium temperature overheating, low energy discharge, and low temperature overheating; the transformer protection signals include: communication failure of the low-voltage side switch monitoring and control device, DC loss alarm of the low-voltage side switch monitoring and control device, low-voltage side switch spring not energized, low-voltage side switch circuit breaker tripping, low-voltage side switch control circuit disconnection, low-voltage side switch temperature and humidity device alarm, medium-voltage side switch monitoring and control device communication failure, medium-voltage side switch monitoring and control device DC loss alarm, low-pressure alarm of other air chambers on the medium-voltage side, voltage transformer loss of the medium-voltage side switch, alarm of the medium-voltage side switch monitoring and control device, power failure of the medium-voltage side switch energy storage power supply, low-pressure alarm of the medium-voltage side switch spring not energized, low-pressure alarm of the medium-voltage side switch air chamber, simultaneous operation of the medium-voltage side switch voltage switching equipment, and medium-voltage... Medium-voltage side switch isolation / grounding control power failure, medium-voltage side switch compartment air chamber low-pressure interlock, medium-voltage side switch control circuit disconnection, medium-voltage side switch signal power failure, medium-voltage side switch lighting / heating power failure, medium-voltage side neutral point grounding switch power loss, high-voltage side switch monitoring and control device communication failure, high-voltage side switch monitoring and control device DC loss alarm, high-voltage side switch isolation / grounding control power failure, high-voltage side switch compartment air chamber low-pressure alarm, high-voltage side other compartment air chamber low-pressure interlock, high-voltage side switch compartment air chamber low-pressure interlock, high-voltage side switch voltage transformer undervoltage, high-voltage side switch monitoring and control device DC loss, high-voltage side switch energy storage motor timeout operation, main transformer high-voltage side switch energy storage motor power loss, high-voltage side switch second group output trip, high-voltage side switch second The following are listed as possible errors / indications: Power supply disconnection, second group control circuit disconnection on the high-voltage side switch, first group output tripping on the high-voltage side switch, power supply disconnection on the first group of the high-voltage side switch, first group control circuit disconnection on the high-voltage side switch, low oil pressure alarm when the high-voltage side switch is open, low oil pressure lockout when the high-voltage side switch is open, low oil pressure alarm when the high-voltage side switch is closed, low oil pressure lockout when the high-voltage side switch is closed, loss of lighting / heating power to the high-voltage side switch, interlocking contact of the high-voltage side switch, three-phase inconsistency on the high-voltage side switch, three-phase inconsistency or non-full-phase operation of the high-voltage side switch, loss of signal power to the high-voltage side switch, reclosing alarm on the high-voltage side switch, reclosing lockout on the high-voltage side switch, loss of power to the neutral grounding switch on the high-voltage side, current transformer (CT) disconnection in the protection cabinet, current transformer (TV) disconnection in the protection cabinet, differential operation of the protection cabinet, low-voltage side alarm in the protection cabinet. The following are listed as alarm / emergency events: Low-voltage side backup activation of the protection cabinet, low-voltage side zero-sequence overvoltage of the protection cabinet, high-voltage side alarm of the protection cabinet, high-voltage side backup activation of the protection cabinet, medium-voltage side alarm of the protection cabinet, medium-voltage side backup activation of the protection cabinet, device alarm of the protection cabinet, device lockout of the protection cabinet, light gas in the main body, abnormal oil level in the main body, heavy gas in the main body, activation of the nitrogen-filled fire extinguishing device, abnormality of the nitrogen-filled fire extinguishing device (low nitrogen pressure), abnormality of the nitrogen-filled fire extinguishing device (fire alarm), abnormality of the nitrogen-filled fire extinguishing device (oil leak alarm), abnormality of the nitrogen-filled fire extinguishing device (power failure), shut-off valve closure, alarm of non-electrical quantity protection device, lockout of non-electrical quantity protection device, power failure of non-electrical quantity protection device, power failure of the air-cooled control box, power on of the air-cooled control box, single-unit operation of the air-cooled control box fan, single-unit failure of the air-cooled control box fan.The following are the possible causes of failures: dual-fan operation in the air-cooled control box, dual-fan malfunction in the air-cooled control box, power supply to the fan group in the air-cooled control box, power supply to the fan group in the air-cooled control box, complete shutdown of the fan group in the air-cooled control box, power supply failure in the air-cooled control box, winding overheating, high winding temperature, pressure release, high oil temperature, on-load tap changer tripping, on-load tap changer operation, power supply failure of the on-load switch, abnormal on-load oil level, on-load online oil filter malfunction, on-load heavy gas, and communication failure of the protection panel. The telemetry data includes: tap position, upper oil temperature of the first group, upper oil temperature of the second group, winding temperature, high-voltage bus voltage, medium-voltage bus voltage, low-voltage bus voltage, high-voltage active power, high-voltage reactive power, medium-voltage active power, medium-voltage reactive power, low-voltage active power, and low-voltage reactive power.

[0041] The protection signal has two values, 0 and 1. 1 indicates that the protection signal is activated, and 0 indicates that the protection signal is not activated.

[0042] S200 employs a hybrid mutual information method to select feature attributes highly correlated with transformer faults from the attributes. Based on the selected feature attributes, labeled and unlabeled training sample sets are obtained for training the mapping relationship between transformer fault types and feature attributes. The attribute vectors of both labeled and unlabeled samples are 108-dimensional, which is highly dimensional. If all parameters of the attribute vectors are used in the classifier training, the computational cost would be too high, leading to the curse of dimensionality. Therefore, the samples in both the labeled and unlabeled training sample sets have undergone dimensionality reduction, effectively reducing computational cost and avoiding the curse of dimensionality.

[0043] According to information theory, the statistical dependency between discrete random variables X and Y is measured by mutual information:

[0044] D x and D y Let X be the set of possible values ​​for discrete random variables X and Y, p(x) be the probability that X takes the value x, and p(y, x) be the joint probability that X takes the value x and Y takes the value y.

[0045] Given that discrete random variable X is known, the statistical dependence between discrete random variables Y and Z is measured by conditional mutual information:

[0046] D x D y and D zLet X be the set of possible values ​​for discrete random variables X, Y, and Z. Let p(y, z|x), p(y|x), and p(z|x) be the conditional probabilities of X taking the value x, and p(y, z, x) be the joint probabilities of X taking the value x, Y taking the value y, and Z taking the value z.

[0047] The probabilities involved in the calculations in this application are obtained from data in the transformer fault diagnosis sample library.

[0048] In the specific implementation process, please refer to Figure 2 As shown, step S200 includes:

[0049] S201 uses the KNN nearest neighbor algorithm to estimate the pseudo-label values ​​of unlabeled samples in the unlabeled training sample set, and combines the label values ​​of labeled samples and the pseudo-label values ​​of unlabeled samples to form a label set Y.

[0050] S202, initialize an empty set as the feature attribute set X for storing feature attributes.

[0051] S203, iterate through each attribute △s in the attribute vector △S. j Calculate attribute △s j Mutual information I(Y; Δs) between the label set Y and the label set Y j Select the attribute △s corresponding to the maximum mutual information. * Add it to the feature attribute set X and remove it from the attribute vector △S.

[0052] S204, loop through each attribute △s in the attribute vector △S. j Calculate attribute △s j Hybrid mutual information I with label set Y h (Y;Δs j |X): I h (Y;s j |X)=(1-h)*I(Y;s j )+h*I(Y;s j |X), where h is the mixing coefficient, with a value of 0.9. The attribute Δs corresponding to the maximum mixing mutual information is taken. ** Add it to the feature attribute set X and remove it from the attribute vector △S.

[0053] S205, check whether the number of elements in the feature attribute set X has reached the preset number threshold. If yes, end S204 and execute S206. Otherwise, repeat S204.

[0054] S206. After obtaining the feature attribute set X, the labeled training sample set and the unlabeled training sample set are formed by retaining the attributes of the labeled and unlabeled samples respectively according to the elements in the feature attribute set X.

[0055] S300 uses a semi-supervised co-training process with labeled and unlabeled training sample sets to obtain a two-classifier that includes the mapping relationship between label values ​​and feature attributes.

[0056] In the specific implementation process, refer to Figure 3 and Figure 4 As shown, step S300 includes:

[0057] S301, the labeled training sample set is divided into a first sub-labeled training sample set and a second sub-labeled training sample set.

[0058] S302, a first classifier is obtained by training the classifier using the first sub-labeled training sample set, and a second classifier is obtained by training the classifier using the second sub-labeled training sample set. In one feasible implementation, the classifier is a support vector machine.

[0059] The model formula for Support Vector Machine is as follows: y = h(x) = ω T φ(x), where φ(x) is a nonlinear mapping from the input space to the high-dimensional space, and ω is the weight vector.

[0060] Solve for the coefficient ω by solving the following optimization problem:

[0061]

[0062]

[0063] Where ε is the width of the insensitivity margin, ξ i , β is a "slack variable" and is a constant that controls the degree of penalty for misclassified samples.

[0064] By processing the optimization problem using Lagrange multipliers, the following dual optimization problem is obtained:

[0065]

[0066]

[0067] in, α i , for ξ i , The corresponding Lagrange multipliers. Solving the dual optimization problem yields a classifier in the following form:

[0068] Where K(x) i,x) is the kernel function, which is the inner product of feature attributes mapped to a higher dimension.

[0069] S303, using the first classifier and the second classifier to estimate a subset of unlabeled training samples randomly drawn from the unlabeled training sample set.

[0070] S304, select a classifier to analyze several unlabeled training samples with high confidence and their pseudo-label values ​​in the subset of unlabeled training samples, and add them to the subset of labeled training samples of another classifier to train and update the classifier. In specific implementation, the confidence level is taken as the mean squared error of the first or second classifier obtained before and after adding the unlabeled training samples to the subset of labeled training samples. The number n of high-confidence unlabeled training samples selected each time... t It changes dynamically with the number of iterations, as shown below:

[0071] n t =n t-1 ρ 0.5t

[0072] Where t is the current iteration number and ρ is the dynamic filtering growth rate. One feasible value for ρ is 1.2.

[0073] S305 checks whether the maximum number of iterations has been reached or whether all unlabeled training samples have been selected into the subset of labeled training samples. If yes, proceed to S306; otherwise, repeat steps S302, S303, and S304.

[0074] S306 outputs the first classifier and the second classifier.

[0075] S400 inputs the data of the transformer whose fault is to be analyzed into the two classifiers obtained in the end, and takes the average value of the classification generated by the first classifier and the second classifier as the fault analysis result.

[0076] Example 2

[0077] See Figure 5 As shown, an embodiment of the present invention provides a transformer fault diagnosis device, including: at least one processing unit, an interface unit, a bus unit, and a storage unit. The processing unit, interface unit, and storage unit are connected to the bus unit. The storage unit stores computer programs and data. The processing unit reads the computer programs, processes the data, and implements the transformer fault diagnosis method.

[0078] Example 3

[0079] This invention provides a storage medium for implementing a transformer fault diagnosis method. The storage medium stores at least one instruction, and a processor reads and executes the instruction to implement the transformer fault diagnosis method.

[0080] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0081] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0082] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0083] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A transformer fault diagnosis method characterized by, The application comprises the following steps: A transformer fault diagnosis sample library is constructed, which stores labeled samples containing attributes and labels and unlabeled samples containing only attributes, and the attributes include transformer protection signals and telemetry data; A hybrid mutual information method is used to select feature attributes with strong correlation with transformer faults from the attributes, and labeled training sample set and unlabeled training sample set for training the mapping relationship between transformer fault types and feature attributes are obtained according to the selected feature attributes; Semi-supervised collaborative training is performed on the labeled training sample set and the unlabeled training sample set to obtain two classifiers containing the mapping relationship between label values and feature attributes, including: dividing the labeled training sample set into a first sub-labeled training sample set and a second sub-labeled training sample set; training the classifier using the first sub-labeled training sample set to obtain a first classifier, and training the classifier using the second sub-labeled training sample set to obtain a second classifier; then estimating a sub-set of unlabeled training samples randomly selected from the unlabeled training sample set using the first classifier and the second classifier; selecting a number of unlabeled training samples with high confidence in the sub-set of unlabeled training samples of one classifier and their pseudo-label values, and adding them to the sub-labeled training sample set of the other classifier to train and update the classifier; continuously repeating the training process until the maximum number of iterations is reached or all unlabeled training samples are selected into the sub-labeled training sample set; The data of the transformer to be analyzed for fault are input into the two classifiers finally obtained, and the fault analysis result is obtained according to the classification mean values generated by the two classifiers.

2. The transformer fault diagnostic method according to claim 1, characterized by, The transformer fault operation records of a preset time length are selected, for each fault, a multi-dimensional attribute vector composed of multiple transformer protection signals and multiple telemetry data is combined with the label of the fault type to form a labeled sample, then the running records of the transformer in the preset time length are uniformly and randomly selected to obtain a set number of unlabeled samples containing only the multi-dimensional attribute vector, and the labeled samples and the unlabeled samples form a transformer fault diagnosis sample library.

3. The transformer fault diagnostic method according to claim 1, characterized by, The hybrid mutual information method is used to select feature attributes with strong correlation with transformer faults from the attributes, including: The pseudo-label values of the unlabeled samples are estimated by the KNN neighbor algorithm; the label values of the labeled samples and the pseudo-label values of the unlabeled samples are combined to form a label set Y; An empty set is initialized as a feature attribute set X for storing feature attributes; Traverse each attribute in the attribute vector ΔS j , calculate the mutual information between the attribute Δs j and the label set Y, select the attribute Δs corresponding to the maximum mutual information * Add to the feature attribute set X and delete from the attribute vector ΔS; Loop through each attribute in the attribute vector ΔS j , calculate the attribute Δs j The mixed mutual information with the label set Y, take the attribute Δs corresponding to the maximum mixed mutual information ** Add to the feature attribute set X, and delete from the attribute vector ΔS; The loop is ended when the number of elements in the feature attribute set X reaches a preset number threshold.

4. The transformer fault diagnostic method according to claim 3, characterized by, After obtaining the feature attribute set X, the attributes in the labeled samples and the unlabeled samples are reserved according to the elements in the feature attribute set X to form a labeled training sample set and an unlabeled training sample set, so as to obtain the labeled training sample set and the unlabeled training sample set for training the mapping relationship between transformer fault types and feature attributes according to the selected feature attributes.

5. The transformer fault diagnostic method according to claim 1, characterized by, The classifier uses a support vector machine.

6. The transformer fault diagnostic method according to claim 1, characterized by, The confidence is the mean square error of the first classifier or the second classifier before and after the unlabeled training sample is added to the sub-labeled training sample set.

7. The transformer fault diagnostic method according to claim 1, characterized by, number of high-confidence unlabeled training samples per screening dynamically changes with the number of iterations, as follows: wherein, is the current iteration number, is the dynamic sieve growth rate.

8. A transformer fault diagnostic device characterized by comprising: The application comprises the following steps: At least one processing unit, interface unit, bus unit, storage unit, the processing unit, interface unit, storage unit connected to the bus unit, the storage unit stores computer programs and data, the processing unit reads computer program processing data to realize the transformer fault diagnosis method as claimed in any one of claims 1-7.

9. A storage medium for implementing a transformer fault diagnosis method, characterized by, The storage medium for realizing the transformer fault diagnosis method stores at least one instruction, and the processor reads and executes the instruction to realize the transformer fault diagnosis method as claimed in any one of claims 1-7.

Images