Transformer winding fault diagnosis method and system based on structure entropy weight method

Abstract

The invention provides a transformer winding fault diagnosis method and system based on a structure entropy weight method, and relates to the technical field of power equipment state monitoring. The method specifically comprises the following steps: collecting importance ranking of monitoring indexes of a transformer by experts by utilizing a Delhu method, and determining a weight value of each monitoring index by adopting a structure entropy weight method based on a ranking result; performing weighting processing on the normalized monitoring indexes of the transformer in the real-time operation state by using the weight values to obtain weighted monitoring indexes; and constructing and training a transformer winding fault intelligent diagnosis model, inputting the weighted monitoring indexes into the trained transformer winding fault intelligent diagnosis model for fault identification, and taking the output transformer winding state as a fault diagnosis result. The invention aims to solve the problems of single signal diagnosis, low diagnosis result accuracy and lack of objectivity of the traditional weight determination method in the current transformer winding diagnosis method.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, and in particular to a method and system for diagnosing transformer winding faults based on the structural entropy weight method. Background Technology

[0002] Transformers are the most important core equipment in a power system, and their operating status directly affects the system's stability and power supply reliability. With the construction and development of smart grids, not only is reliable transformer operation required, but also higher demands are placed on the accuracy of transformer fault diagnosis. Comprehensive and scientific fault diagnosis of transformers provides the correct basis for power operation departments to formulate maintenance strategies. As the core component of a transformer, the deterioration of the mechanical and insulation conditions of the transformer windings is a major cause of major faults.

[0003] Current methods for assessing the condition of transformer windings have significant limitations: First, traditional weight determination methods (such as AHP and expert scoring) rely excessively on subjective experience, leading to a lack of objectivity in the allocation of indicator weights. This is especially problematic when dealing with heterogeneous data from multiple sources, such as vibration, chemical, and electrical data, where expert bias amplifies systemic errors. Second, existing technologies fail to effectively integrate key information such as vibration spectra, dissolved gases, and electrical parameters. Independent analysis of each data source results in fragmented condition assessments, failing to capture the coupling relationship between winding mechanical deformation and insulation degradation. Finally, mainstream diagnostic models (such as BP neural networks and SVM) have deficiencies in fault classification, such as failing to establish a correlation between weights and fault characteristics and neglecting the control of missed fault risks, leading to insufficient accuracy in fault identification. These problems severely restrict the development of intelligent operation and maintenance of power equipment, urgently requiring an innovative method that integrates subjective and objective weight calculations with collaborative diagnosis based on multi-source data. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention optimizes the weights of multiple monitoring indicators using the structural entropy weighting method and utilizes an intelligent diagnostic model for transformer winding faults based on a Relevance Vector Machine (RVM) classifier to identify the state of the weighted multiple monitoring indicators. This proposes a transformer winding fault diagnosis method and system based on the structural entropy weighting method, aiming to solve the problems of single-signal diagnosis, low accuracy of diagnostic results, and the over-reliance on subjective experience in traditional weight determination methods (such as AHP and expert scoring), leading to a lack of objectivity in indicator weight allocation.

[0005] On the one hand, this invention proposes a transformer winding fault diagnosis method based on the structural entropy weight method, which includes the following process:

[0006] Data acquisition transformer in real-time operating status The monitoring indicators were normalized to obtain the normalized monitoring indicators;

[0007] Experts used the Delphi method to collect data. The importance of each monitoring indicator is ranked, and based on the ranking results, the weight value of each monitoring indicator is determined using the structural entropy weight method.

[0008] The normalized monitoring indicators are weighted using the aforementioned weight values ​​to obtain the weighted monitoring indicators.

[0009] A smart diagnostic model for transformer winding faults is constructed and trained to obtain a well-trained smart diagnostic model for transformer winding faults.

[0010] The weighted monitoring indicators are input into the trained intelligent diagnostic model for transformer winding faults for fault identification, and the output transformer winding status is used as the fault diagnosis result.

[0011] Furthermore, the monitoring indicators include, but are not limited to: fundamental frequency energy ratio, low-frequency harmonic energy sum, axial-radial energy ratio, C2H2 / H2 ratio, micro-water growth rate, furfural content, short-circuit impedance deviation, dielectric loss angle increment, and high-frequency current harmonics.

[0012] Furthermore, the expert analysis using the Delphi method... The importance of each monitoring indicator is ranked, and based on the ranking results, the weight value of each monitoring indicator is determined using the structural entropy weighting method. The specific details are as follows:

[0013] Based on the Delphi method, invitation Each expert or expert group fills out a ranking table, and... The monitoring indicators were ranked by importance, and based on... The ranking list of individual experts or expert groups yields the ranking matrix of the experts.

[0014] Based on the structural entropy weighting method, the membership function is used to convert the expert ranking matrix into a membership degree matrix;

[0015] For any monitoring indicator Calculate based on the membership matrix Each expert or expert group on monitoring indicators average awareness And based on average awareness calculate Each expert or expert group on monitoring indicators Cognitive blindness ;

[0016] Using the calculated average recognition and awareness of blindness ,calculate Each expert or expert group on each monitoring indicator Overall awareness ;

[0017] based on The overall understanding of all monitoring indicators by individual experts or expert groups is obtained by normalizing the overall understanding of each monitoring indicator. The weight values ​​of each monitoring indicator.

[0018] Furthermore, the transformer winding condition includes: normal condition and fault condition; wherein the fault condition includes: axial loosening, radial deformation, inter-turn short circuit and insulation aging.

[0019] Furthermore, the specific content of constructing and training the intelligent diagnostic model for transformer winding faults, resulting in the trained intelligent diagnostic model for transformer winding faults, is as follows:

[0020] Obtain several sets of transformers under different transformer winding conditions. The system acquires several monitoring indicators, and then weights the acquired monitoring indicators according to the weight value of each monitoring indicator. A set of weighted monitoring indicators is then used as a sample to construct a sample set.

[0021] The sample set is divided into a training set and a validation set according to a preset ratio;

[0022] Construct an intelligent diagnostic model for transformer winding faults;

[0023] The intelligent diagnostic model for transformer winding faults was trained using the training set to obtain the trained intelligent diagnostic model for transformer winding faults, and then validated using the validation set.

[0024] Furthermore, the intelligent diagnostic model for transformer winding faults consists of five binary correlation vector machine classifiers connected in series, denoted as RVM1, RVM2, RVM3, RVM4, and RVM5. RVM1 is used to identify whether the transformer winding is in a normal state; RVM2 is used to identify whether the transformer winding has axial looseness; RVM3 is used to identify whether the transformer winding has radial deformation; RVM4 is used to identify whether the transformer winding has inter-turn short circuit; and RVM5 is used to identify whether the transformer winding has insulation aging.

[0025] Furthermore, the specific content of training the intelligent diagnostic model for transformer winding faults using the training set is as follows:

[0026] For any binary correlation vector machine classifier in the intelligent diagnosis model for transformer winding faults, based on the binary classification task performed by the classifier, samples belonging to the two categories in the binary classification task are extracted from the training set to construct the training set for the classifier. ;in for A 3D feature vector is used to represent a sample; For the sample The label is used to define the fault category in this binary classification task, and ;

[0027] Determine the kernel function of the binary relevance vector machine classifier;

[0028] Initialize the weight vector of the binary correlation vector machine classifier. and for the weight vector Set a zero-mean Gaussian prior for the weight vector. Each weight parameter in Set an initial hyperparameter. To obtain the hyperparameter vector ;

[0029] The binary correlation vector machine classifier is iteratively optimized, and the following process is performed in each iteration:

[0030] With the current hyperparameters fixed, the weight vector is solved by Bayesian posterior inference. Maximum a posteriori estimation ;

[0031] In maximum a posteriori estimation At this point, calculate the negative Hessian matrix of the log-posterior probability and invert it to obtain the posterior covariance matrix. ;

[0032] Based on maximum a posteriori estimation and posterior covariance matrix Update hyperparameter vector ;

[0033] If the current hyperparameter vector If the iteration fails to converge or the preset maximum number of iterations is not reached, then the next iteration begins; if the current hyperparameter vector... If convergence is achieved or the preset maximum number of iterations is reached, then based on the current hyperparameter vector... and maximum a posteriori estimation Construct a trained binary correlation vector machine classifier;

[0034] All trained binary correlation vector machine classifiers are concatenated in the order of task execution to obtain a trained intelligent diagnostic model for transformer winding faults.

[0035] On the other hand, this invention proposes a transformer winding fault diagnosis system based on the structural entropy weight method, the system comprising:

[0036] The data acquisition and processing module is used to collect data on the transformer's real-time operating status. The monitoring indicators were normalized to obtain the normalized monitoring indicators;

[0037] The indicator weight allocation module uses the Delphi method to collect expert data. The importance of each monitoring indicator is ranked, and based on the ranking results, the weight value of each monitoring indicator is determined using the structural entropy weight method.

[0038] The indicator weighting module uses the weight values ​​to perform weighting processing on the normalized monitoring indicators to obtain the weighted monitoring indicators.

[0039] The diagnostic model training module is used to build and train an intelligent diagnostic model for transformer winding faults, resulting in a trained intelligent diagnostic model for transformer winding faults.

[0040] The intelligent fault diagnosis module is used to input the weighted monitoring indicators into the trained intelligent fault diagnosis model of transformer windings for fault identification, and to use the output transformer winding status as the fault diagnosis result.

[0041] Thirdly, the present invention proposes an electronic device comprising: one or more processors, and a memory for storing instructions, which, when executed by the one or more processors, cause the one or more processors to execute the transformer winding fault diagnosis method based on the structural entropy weight method.

[0042] Fourthly, the present invention proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the transformer winding fault diagnosis method based on the structural entropy weight method.

[0043] The beneficial effects of adopting the above technical solution are as follows:

[0044] 1. This invention abandons the traditional approach of single-signal analysis and uses multi-information signals for analysis and diagnosis, which effectively reduces the risk of misdiagnosis and makes the diagnostic results more accurate and reliable.

[0045] 2. The structural entropy weighting method solves the traditional weight allocation problem by using Delphi expert surveys to generate typical rankings, eliminating noise through membership function transformation and blindness analysis, quantifying the importance of indicators, and dynamically adjusting weights.

[0046] 3. By using the structural entropy weighting method to optimize the weights of multi-source monitoring indicators in a unified subjective and objective manner, the diagnostic bias caused by improper feature importance allocation in traditional RVM is solved, and the physical interpretability and noise resistance of sparse models are significantly improved. Attached Figure Description

[0047] Figure 1 This is a flowchart of the transformer winding fault diagnosis method based on the structural entropy weighting method in this embodiment;

[0048] Figure 2 This is a flowchart of the fault diagnosis process for the transformer windings in this embodiment;

[0049] Figure 3 This is a structural diagram of the transformer winding fault diagnosis system based on the structural entropy weighting method in this embodiment. Detailed Implementation

[0050] To facilitate understanding of this application, specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. The following embodiments are illustrative of the invention but are not intended to limit its scope. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application.

[0051] Example 1:

[0052] This embodiment presents a transformer winding fault diagnosis method based on the structural entropy weighting method, such as... Figure 1 As shown, the method includes the following steps:

[0053] Data acquisition transformer in real-time operating status The monitoring indicators were normalized to obtain the normalized monitoring indicators.

[0054] In this embodiment, multiple data signals, i.e., monitoring indicators, of the transformer in real-time operation are collected, including: fundamental frequency energy ratio: the proportion of vibration energy at 50 / 100Hz; low-frequency harmonic energy sum: the integral of vibration energy in the 0-500Hz frequency band; axial-radial energy ratio: the ratio of vibration energy between the Z-axis and X-axis; C2H2 / H2 ratio: the ratio of acetylene to hydrogen concentration; micro-water growth rate: the rate of change in water content in the oil; furfural content: the concentration of aging products in the insulating paper; short-circuit impedance deviation: the relative error between the measured impedance and the reference value; dielectric loss angle increment; high-frequency current harmonics: the current distortion rate at 1-5kHz. These signals cover mechanical vibration signals, chemical signals, and electrical signals.

[0055] To eliminate the impact of differences in units and value ranges among data features, and to ensure that different features have the same scale unit, this embodiment normalizes the collected monitoring indicators, enabling accurate fusion of multi-source heterogeneous data within a unified framework. The normalized monitoring indicators are represented in vector form, i.e. ;in This represents the normalized monitoring indicators.

[0056] Experts used the Delphi method to collect data. The importance of each monitoring indicator is ranked, and based on the ranking results, the weight value of each monitoring indicator is determined using the structural entropy weight method.

[0057] The experts who used the Delphi method to collect data... The importance of each monitoring indicator is ranked, and based on the ranking results, the weight value of each monitoring indicator is determined using the structural entropy weighting method. The specific details are as follows:

[0058] Based on the Delphi method, invitation Each expert or expert group fills out a ranking table, and... The monitoring indicators were ranked by importance, and based on... The ranking list of experts or expert groups yields the ranking matrix of experts.

[0059] (1)

[0060] in Represents a sorting matrix; For the first in the sorting matrix Line 1 The elements of the column represent experts. For monitoring indicators The sorting value; The index value is for the expert or expert panel, and ; For the index value of the monitoring indicator, and .

[0061] In this embodiment, multiple monitoring indicators are used as evaluation indicators to establish a transformer evaluation indicator set. Then, based on the transformer evaluation indicator set, a corresponding weight set is determined. The Delphi method is used to collect expert opinions, and a qualitative ranking "Expert Survey Form for Indicator System Weights" is designed. Following the standard procedures and requirements of the Delphi method, questionnaires are distributed to multiple qualified experts. Expert selection must meet the principles of representativeness, authority, and impartiality, with priority given to professionals familiar with the evaluated object and related indicator systems. Experts complete the consultation form anonymously, independently providing qualitative ranking opinions on the importance of the indicator set based on their professional knowledge and practical experience, specifically by marking with "√". After multiple rounds of opinion solicitation and feedback, consensus is gradually reached, and the final expert ranking results can be defined as the "typical ranking" of the indicators. Table 1 below shows an example ranking, where the table is divided into three expert groups, each with three experts.

[0062] Table 1 Typical Ranking Results of Data Indicators

[0063]

[0064] Table 1 is transformed into a ranking matrix. Table 1 contains 9 monitoring indicators and 3 expert groups. The matrix transformation rule is as follows: if an expert group considers a monitoring indicator as their "first choice," the value in the matrix is ​​1; if they consider it their "second choice," the value is 2, and so on for other monitoring indicators. The ranking matrix for each expert group is as follows.

[0065] (2)

[0066] The elements in the matrix are the ranking values ​​of the monitoring indicators by experts, and the importance of the ranking values ​​decreases sequentially from 1 to 9.

[0067] Based on the structural entropy weighting method, the membership function is used to convert the sorting matrix into a membership degree matrix.

[0068] In this embodiment, each expert's qualitative ranking of each monitoring indicator is converted into a quantitative membership degree. The membership function used in this conversion process is:

[0069] (3)

[0070] in Indicates sorting value The corresponding membership degree; These are normalization coefficients used to ensure that the range of the final calculated membership values ​​is within a certain range. Within the range; Sort by value The probability of importance.

[0071] make ,Pick Substituting into equation (3) above, we get:

[0072] (4)

[0073] in For the conversion parameters, ; It is the largest sort number.

[0074] Simplifying equation (4), we get:

[0075] (5)

[0076] Divide both sides of the above equation (5) by the same value. ,get:

[0077] (6)

[0078] And let , For definition in Variables in range for The corresponding membership function values ​​are obtained as follows:

[0079] (7)

[0080] in Sort by value The corresponding membership function values; thus, the sorting matrix is ​​obtained. The corresponding membership matrix .

[0081] Assuming the sorting participants Each expert or expert group on monitoring indicators With equal say, calculation Each expert or expert group on monitoring indicators The "consensus" is called the average level of awareness. The specific calculation process is as follows:

[0082] (8)

[0083] Each expert or expert group will monitor the indicators. The uncertainty arising from cognition is called "cognitive blindness," denoted as _____. The expression for its calculation is as follows:

[0084] (9)

[0085] in .

[0086] For each monitoring indicator to be evaluated ,calculate Each expert or expert group on each monitoring indicator Overall awareness The calculation expression is as follows:

[0087] (10)

[0088] From equation (10), we can obtain Evaluation vector of all monitoring indicators by an expert or expert group .

[0089] based on The overall understanding of all monitoring indicators by individual experts or expert groups is obtained by normalizing the overall understanding of each monitoring indicator. The weight values ​​of each monitoring indicator.

[0090] To obtain the weights of the evaluation indicators, equation (10) is normalized, and its specific expression is as follows:

[0091] (11)

[0092] in For monitoring indicators The normalized overall awareness, i.e., the monitoring indicators The final weight satisfies and That is, the overall consensus of m experts on the importance of the indicator; This is called the index set. The weight vector.

[0093] The normalized monitoring indicators are weighted using the weight values ​​to obtain the weighted monitoring indicators.

[0094] To better diagnose transformer winding faults, this embodiment employs a correlation-based method. First, the normalized monitoring indicators are weighted:

[0095] (12)

[0096] in The weights of the monitoring indicators; These are the normalized monitoring indicators.

[0097] A smart diagnostic model for transformer winding faults was constructed and trained to obtain a well-trained smart diagnostic model for transformer winding faults.

[0098] The intelligent diagnostic model for transformer winding faults is composed of five binary correlation vector machine classifiers connected in series, denoted as RVM1, RVM2, RVM3, RVM4, and RVM5. RVM1 is used to identify whether the transformer winding is in a normal state; RVM2 is used to identify whether the transformer winding has axial looseness; RVM3 is used to identify whether the transformer winding has radial deformation; RVM4 is used to identify whether the transformer winding has inter-turn short circuit; and RVM5 is used to identify whether the transformer winding has insulation aging.

[0099] The specific content of constructing and training a smart diagnostic model for transformer winding faults, and obtaining a trained smart diagnostic model for transformer winding faults, is as follows:

[0100] Obtain several sets of transformers under different transformer winding conditions. Each monitoring indicator is selected, and the obtained monitoring indicators are weighted according to the weight value of each monitoring indicator. A set of weighted monitoring indicators is then used as a sample to construct a sample set.

[0101] The sample set is divided into a training set and a validation set according to a preset ratio.

[0102] The transformer winding condition includes: normal condition and fault condition; wherein the fault condition includes: axial loosening, radial deformation, inter-turn short circuit and insulation aging.

[0103] In this embodiment, data from five different transformer winding states were selected from the experimental data source, including: normal state and four other fault modes (axial loosening, radial deformation, inter-turn short circuit, and insulation aging). 100 samples from each state together constitute the data foundation of this study. To ensure the effectiveness of the model evaluation, these 500 samples were divided, with 80% used for model training and the remaining 20% ​​reserved for subsequent diagnostic verification.

[0104] Construct an intelligent diagnostic model for transformer winding faults.

[0105] The intelligent diagnostic model for transformer winding faults consists of five binary correlation vector machine classifiers connected in series, denoted as RVM1, RVM2, RVM3, RVM4, and RVM5. RVM1 is used to identify whether the transformer winding is in a normal state; RVM2 is used to identify whether the transformer winding has axial looseness; RVM3 is used to identify whether the transformer winding has radial deformation; RVM4 is used to identify whether the transformer winding has inter-turn short circuit; and RVM5 is used to identify whether the transformer winding has insulation aging.

[0106] In this embodiment, the operating states of the transformer are classified into five typical types: normal state, axial mechanical loosening, radial deformation, inter-turn short circuit, and insulation material aging. The diagnostic model consists of five binary correlation vector machine classifiers connected in series, denoted as RVM1, RVM2, RVM3, RVM4, and RVM5, each responsible for identifying one type of fault state. This structured design enables accurate identification and effective isolation of multiple fault types.

[0107] The intelligent diagnostic model for transformer winding faults was trained using the training set to obtain the trained intelligent diagnostic model for transformer winding faults, and then validated using the validation set.

[0108] The specific content of training the intelligent diagnostic model for transformer winding faults using the training set is as follows:

[0109] For any binary correlation vector machine classifier in the intelligent diagnosis model for transformer winding faults, based on the binary classification task performed by the classifier, samples belonging to the two categories in the binary classification task are extracted from the training set to construct the training set for the classifier. ;in for A 3D feature vector is used to represent a sample; For the sample The label is used to define the fault category in this binary classification task, and ; This represents the number of samples.

[0110] In this embodiment, for RVM1, This indicates that it is normal. Indicates a fault; for RVM2, This indicates that there is axial looseness. This indicates that there is no axial loosening; for RVM3, This indicates the presence of radial deformation. This indicates that there is no axial looseness; for RVM4, This indicates the presence of an inter-turn short circuit. This indicates that there is no inter-turn short circuit; for RVM5, This indicates that insulation aging has occurred. This indicates that there is no insulation aging.

[0111] The output of the RVM model is defined as:

[0112] (13)

[0113] in Represents the first in the training set One sample; This represents the input to the RVM; Represents the kernel function; For bias terms; For the sample The corresponding weight parameters; This is the weight vector.

[0114] Determine the kernel function of the binary correlation vector machine classifier.

[0115] In this embodiment, the kernel function in RVM is typically the Gaussian radial basis function, i.e.:

[0116] (14)

[0117] in The parameter of the Gaussian kernel is called the bandwidth.

[0118] Target variable Following a Bernoulli distribution, a likelihood function is constructed, which describes the model output. Labels are generated after applying the sigmoid function. The probability is expressed as:

[0119] (15)

[0120] in Let be the likelihood function, describing the likelihood of a given input. Generate tags The probability of; Indicates fault label; For the logistic sigmoid function; This is the input to the logistic sigmoid function.

[0121] Initialize the weight vector of the binary correlation vector machine classifier. and for the weight vector Set a zero-mean Gaussian prior for the weight vector. Each weight parameter in Set an initial hyperparameter. .

[0122] To ensure the sparsity of the model, this prior is based on the Automatic Relevance Determination (ARD) mechanism: Very large, then The posterior probabilities are concentrated around 0, and the corresponding features are removed from the model, achieving sparsity. RVM is the weight... Set a zero-mean Gaussian prior for each weight. Each has a corresponding hyperparameter with independent precision. The Gaussian prior distribution is expressed as:

[0123] (16)

[0124] in It is a hyperparameter vector; It is a normal distribution function.

[0125] Then, through Bayesian posterior inference, its computational expression is:

[0126] (17)

[0127] in This is the posterior probability; This is the marginal likelihood function, used to optimize hyperparameters; The Gaussian prior distribution is represented to achieve automatic sparsity. Since the sigmoid likelihood function is not conjugate with the Gaussian prior, the posterior distribution cannot be obtained in an exact analytical form. RVM uses the Laplace approximation to address this, first finding the posterior mode (maximum a posteriori estimate, MAP):

[0128] (18)

[0129] Since the posterior is proportional to the likelihood multiplied by the prior, this is equivalent to:

[0130] (19)

[0131] The posterior mode is approximated by a Gaussian distribution: First, the covariance matrix is ​​the log-posterior probability... The inverse of the negative Hessian matrix at point is:

[0132] (20)

[0133] in Weight vector The posterior covariance matrix; For vector differential operators, it means that the weight vector is used as the basis for differentiation. Given a function of the independent variable, find its gradient.

[0134] Through derivation, the log-posterior Hessian matrix is... for:

[0135] (twenty one)

[0136] in , ; yes The design matrix; The precision matrix composed of hyperparameters; It is A diagonal matrix, where medium elements .

[0137] Therefore, the approximate posterior covariance is:

[0138] (twenty two)

[0139] Optimizing hyperparameters by maximizing the marginal likelihood function (evidence) :

[0140] (twenty three)

[0141] Since direct calculation of the integral is difficult, after applying the Laplace approximation to the posterior, the logarithm of the marginal likelihood function can be approximated as:

[0142] (twenty four)

[0143] Where const is a constant term.

[0144] By maximizing the evidence, hyperparameters can be obtained. Update rules:

[0145] (25)

[0146] in ; It is the maximum a posteriori estimate The One component; It is the posterior covariance matrix The One diagonal element.

[0147] Fix current Solve according to formula (18) Calculate the approximate posterior covariance ,according to and Update hyperparameters If hyperparameters If convergence is not achieved, repeat equations (18) to (24) until convergence is achieved. Convergence (or reaching the maximum number of iterations).

[0148] During the training phase, based on the known weighted feature data of normal and fault categories, the hyperparameters are adjusted. and related weight vectors The updated hyperparameters are then calculated and applied to the five classifiers. and corresponding weight vector Then for the new input , The predicted distribution is obtained by integrating the weights posteriorly:

[0149] (26)

[0150] in .

[0151] Because integrals are difficult to calculate, approximations are often used, namely, the posterior Gaussian distribution model. Perform point estimation:

[0152] (27)

[0153] The weighted monitoring indicators are input into the trained intelligent diagnostic model for transformer winding faults for fault identification, and the output transformer winding status is used as the fault diagnosis result.

[0154] In a binary classification fault diagnosis model, the decision-making rule is based on a preset probability threshold (usually set to 0.5). If the probability value calculated by the algorithm is higher than this threshold, it is classified as fault type one; otherwise, it is classified as fault type two. This threshold and the probability correlation between the corresponding fault types are learned during the model training phase using labeled fault samples. For example, when training the RVM1 model, given feature data of known fault states, the model learns that when the output probability is greater than 0.5, it is significantly correlated with the fault state. Therefore, when faced with new unknown feature data, RVM1 uses the same threshold standard, that is, if the calculated probability exceeds 0.5, it is judged as a fault state. The training process for the other four classifiers is the same as that for RVM1.

[0155] like Figure 2 As shown, for any transformer, the fault diagnosis process uses a multi-level correlation vector machine (RVM) model for layer-by-layer discrimination. The specific process is as follows:

[0156] (1) Before fault diagnosis, feature data under known normal state and fault type needs to be normalized. Then, the normalized data is weighted using the structural entropy weighting method. The weighted data samples are used to train the correlation vector machine classifier. The trained classifier can identify the fault type corresponding to different weighted feature data.

[0157] (2) To diagnose the fault of the transformer winding, the weighted feature vector of the unknown fault is first used as the input signal and input into the already trained correlation vector machine classifier 1. Then, the diagnosis is performed in classifier 1. The output is either normal or faulty. If the output is normal, the diagnosis ends. If the output is faulty, it enters the correlation vector machine classifier 2.

[0158] (3) In the Relevant Vector Machine Classifier 2, the winding axial looseness fault and other fault states can be diagnosed. If the output is the winding axial looseness fault, the diagnosis ends. If the output is other fault states, it enters the Classifier 3.

[0159] (4) In the Relevant Vector Machine Classifier 3, the winding radial deformation fault and other fault states can be diagnosed. If the output is the winding radial deformation fault, the diagnosis ends. If the output is other fault states, it enters the Classifier 4.

[0160] (5) In the relevant vector machine classifier 4, the winding inter-turn short circuit fault and other fault states can be diagnosed. If the output is the winding inter-turn short circuit fault, the diagnosis ends. If the output is other fault states, it enters the classifier 5.

[0161] (6) In the relevant vector machine classifier 5, the winding insulation aging fault can be diagnosed. If the output is the winding insulation aging fault, the diagnosis ends.

[0162] Complete the entire graded fault identification process.

[0163] Example 2:

[0164] This embodiment presents a transformer winding fault diagnosis system based on the structural entropy weight method, such as... Figure 3 As shown, the system includes:

[0165] The data acquisition and processing module is used to collect data on the transformer's real-time operating status. The monitoring indicators were normalized to obtain the normalized monitoring indicators;

[0166] The indicator weight allocation module uses the Delphi method to collect expert data. The importance of each monitoring indicator is ranked, and based on the ranking results, the weight value of each monitoring indicator is determined using the structural entropy weight method.

[0167] The indicator weighting module uses the weight values ​​to perform weighting processing on the normalized monitoring indicators to obtain the weighted monitoring indicators.

[0168] The diagnostic model training module is used to build and train an intelligent diagnostic model for transformer winding faults, resulting in a trained intelligent diagnostic model for transformer winding faults.

[0169] The intelligent fault diagnosis module is used to input the weighted monitoring indicators into the trained intelligent fault diagnosis model of transformer windings for fault identification, and to use the output transformer winding status as the fault diagnosis result.

[0170] Example 3:

[0171] This embodiment proposes an electronic device, including: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the transformer winding fault diagnosis method based on the structural entropy weight method.

[0172] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the transformer winding fault diagnosis method based on the structural entropy weight method as described in the embodiments. It is understood that the electronic device may also include input / output (I / O) interfaces and communication components.

[0173] The processor is used to execute all or part of the steps in the transformer winding fault diagnosis method based on the structural entropy weighting method as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.

[0174] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the transformer winding fault diagnosis method based on structural entropy weighting described in the above embodiments.

[0175] Example 4:

[0176] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0177] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the transformer winding fault diagnosis method based on the structural entropy weight method described in the various embodiments of this application.

[0178] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) application stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the transformer winding fault diagnosis method based on the structural entropy weight method described above.

[0179] Example 5:

[0180] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the transformer winding fault diagnosis method based on the structural entropy weight method.

[0181] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0182] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0183] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of this disclosure and its equivalents, then the intent of this disclosure also includes these modifications and variations.

Claims

1. A transformer winding fault diagnosis method based on structural entropy weight method, characterized in that, The method comprises the following processes: The transformer in real-time operation state is collected Item monitoring indicators are normalized to obtain normalized monitoring indicators; Experts used the Delphi method to collect data. The importance of each monitoring indicator is ranked, and based on the ranking results, the weight value of each monitoring indicator is determined using the structural entropy weight method. The normalized monitoring indicators are weighted using the weight values to obtain weighted monitoring indicators; A transformer winding fault intelligent diagnosis model is constructed and trained to obtain a trained transformer winding fault intelligent diagnosis model; The weighted monitoring indicators are input into the trained transformer winding fault intelligent diagnosis model for fault identification, and the output transformer winding state is taken as a fault diagnosis result.

2. The transformer winding fault diagnosis method based on the structural entropy weight method according to claim 1, characterized in that, The monitoring indicators include, but are not limited to, a fundamental frequency energy ratio, a low-frequency harmonic energy, an axial-radial energy ratio, a C2H2 / H2 ratio, a micro-water growth rate, a furfural content, a short-circuit impedance deviation, a dielectric loss angle increment, and a high-frequency current harmonic.

3. The transformer winding fault diagnosis method based on the structural entropy weight method according to claim 1, characterized in that, The method comprises the following steps: collecting the importance ranking of each monitoring index of the monitoring system by the Delphi method; and determining the weight value of each monitoring index based on the ranking result by the structural entropy weight method. The specific content of the method is as follows: collecting the importance ranking of each monitoring index of the monitoring system by the Delphi method; and determining the weight value of each monitoring index based on the ranking result by the structural entropy weight method. Based on the Delphi method, invitation Each expert or expert group fills out a ranking table, and... The monitoring indicators were ranked by importance, and based on... The ranking list of individual experts or expert groups yields the ranking matrix of the experts. Based on the structural entropy weight method, the ranking matrix of the experts is converted into a membership matrix using a membership function; For any monitoring indicator Calculate based on the membership matrix Each expert or expert group on monitoring indicators average awareness And based on average awareness calculate Each expert or expert group on monitoring indicators Cognitive blindness ; using the calculated average knowledge and the knowledge blindness , calculating the overall knowledge of the individual expert or expert group about each monitoring indicator ;​ Based on The overall awareness degree of the individual or the expert group to all the monitoring indexes, by normalizing the overall awareness degree of each monitoring index, the weight value of each monitoring index is obtained respectively. The weight value of each monitoring index is obtained respectively.

4. The transformer winding fault diagnosis method based on the structural entropy weight method according to claim 3, characterized in that, The transformer winding state includes a normal state and a fault state, and the fault state includes axial looseness, radial deformation, inter-turn short circuit, and insulation aging.

5. The transformer winding fault diagnosis method based on the structural entropy weight method according to claim 4, characterized in that, The specific content of constructing the transformer winding fault intelligent diagnosis model and training the transformer winding fault intelligent diagnosis model to obtain the trained transformer winding fault intelligent diagnosis model is as follows: Obtaining a plurality of groups of transformer under different transformer winding states The monitoring indexes are weighted according to the weight values of the monitoring indexes, and a set of weighted monitoring indexes is taken as a sample to construct a sample set. The sample set is divided into a training set and a verification set according to a preset ratio; The transformer winding fault intelligent diagnosis model is constructed; The transformer winding fault intelligent diagnosis model is trained using the training set to obtain the trained transformer winding fault intelligent diagnosis model, and the verification set is used for verification.

6. The transformer winding fault diagnosis method based on the structural entropy weight method according to claim 5, characterized in that, The transformer winding fault intelligent diagnosis model is composed of five binary classification related vector machine classifiers in series, denoted as RVM1, RVM2, RVM3, RVM4, and RVM5, wherein RVM1 is used to identify whether the transformer winding is in a normal state, RVM2 is used to identify whether the transformer winding has axial looseness, RVM3 is used to identify whether the transformer winding has radial deformation, RVM4 is used to identify whether the transformer winding has an inter-turn short circuit, and RVM5 is used to identify whether the transformer winding has insulation aging.

7. The transformer winding fault diagnosis method based on the structural entropy weight method according to claim 6, characterized in that, The specific content of training the transformer winding fault intelligent diagnosis model using the training set is as follows: For any binary classification related vector machine classifier in the transformer winding fault intelligent diagnosis model, according to the binary classification task performed by the binary classification related vector machine classifier, samples belonging to two categories in the binary classification task are extracted from the training set to construct the training set of the binary classification related vector machine classifier ; wherein is a feature vector of dimensionality d, used to represent a sample; is a label of the sample , used to represent the fault category under the binary classification task, and ; The kernel function of the binary classification related vector machine classifier is determined; initializing a weight vector of the binary classification relevance vector machine classifier and setting a zero-mean Gaussian prior for the weight vector for each weight parameter , setting an initial hyperparameter , resulting in a hyperparameter vector ;​ The binary classification related vector machine classifier is iteratively optimized, and the following processes are performed in each iteration: Fixing the current hyperparameters, and solving for the weight vector by Bayesian posterior inference Maximum a posteriori estimation of ; At maximum a posteriori estimation where the negative Hessian matrix of the log posterior probability is computed and inverted to obtain the posterior covariance matrix ; According to maximum a posteriori estimation and the posterior covariance matrix updating the hyperparameter vector ; If the current hyper-parameter vector does not converge or does not reach the preset maximum iteration number, the next round of iteration is started; if the current hyper-parameter vector converges or reaches the preset maximum iteration number, a trained binary classification relevance vector machine classifier is constructed according to the current hyper-parameter vector and the maximum a posteriori estimation . All the trained binary classification related vector machine classifiers are connected in series according to the task execution order to obtain the trained transformer winding fault intelligent diagnosis model.

8. A transformer winding fault diagnosis system based on structural entropy weight method, for implementing the transformer winding fault diagnosis method based on structural entropy weight method according to any one of claims 1-7, characterized in that, The system comprises: The data acquisition and processing module is used for acquiring monitoring indexes of the transformer under real-time operation state and normalizing the monitoring indexes to obtain normalized monitoring indexes; The index weight distribution module collects the importance ordering of each monitoring index by the Delphi method through experts, and determines the weight value of each monitoring index based on the ordering result by the structure entropy weight method. The index weight distribution module collects the importance ordering of each monitoring index by the Delphi method through experts, and determines the weight value of each monitoring index based on the ordering result by the structure entropy weight method. An indicator weighting module that weights the normalized monitoring indicators using the weight values to obtain weighted monitoring indicators; A diagnosis model training module that is configured to construct a transformer winding fault intelligent diagnosis model and train the transformer winding fault intelligent diagnosis model to obtain a trained transformer winding fault intelligent diagnosis model; An intelligent fault diagnosis module that is configured to input the weighted monitoring indicators into the trained transformer winding fault intelligent diagnosis model for fault identification, and take the output transformer winding state as a fault diagnosis result.

9. An electronic device, comprising: The system comprises: One or more processors, and a memory for storing instructions which, when executed by the one or more processors, cause the one or more processors to perform the transformer winding fault diagnosis method based on the structural entropy weight method of any one of claims 1-7.

10. A computer program product, characterised in that, A computer program product including computer program or instructions which, when executed by a processor, implement the transformer winding fault diagnosis method based on the structural entropy weight method of any one of claims 1-7.

Images