Fault detection method and system for oil-immersed transformer
By building a decision tree model and gradient enhancement model, and optimizing hyperparameters with Bayesian optimization scheme, the misreading and inaccurate problems in oil-immersed transformer fault detection is solved, and fault detection with high accuracy and high reliability is achieved .
Patent Information
- Application Number
- CN202411550392.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-01
- Publication Date
- 2025-02-14
AI Technical Summary
Existing oil-immersed transformer fault detection methods, such as oil dissolved gas analysis (DGA) and detection schemes based on artificial intelligence algorithms, have problems such as misreading, inaccuracy and low reliability, and are highly complex and difficult to implement. .
Using a fault detection method combining decision tree model and gradient enhancement model, the training data set is constructed by obtaining gas dissolution analysis data of the oil-immersed transformer in the early stage of failure, and the model hyperparameters are optimized using Bayesian optimization scheme to optimize the model. Improve the accuracy and reliability of detection.
The detection of oil-immersed transformer faults is achieved, which avoids misreading and inaccurate problems in traditional methods, and reduces the complexity and difficulty of detection.
Smart Images
Figure CN119441887A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of electrical automation, and in particular relates to a method and system for detecting faults of an oil-immersed transformer. Background Art
[0002] With the development of economy and technology and the improvement of people's living standards, electricity has become an indispensable secondary energy source for people's production and life, bringing endless convenience to people's production and life. Therefore, ensuring the stable and reliable supply of electricity has become one of the most important tasks of the power system.
[0003] Oil-immersed transformers are important components of power systems, so fault detection of oil-immersed transformers is of great significance to power systems. At present, the fault detection scheme for oil-immersed transformers often adopts the dissolved gas analysis (DGA) scheme, that is, by analyzing the oil and gas samples collected from the faulty transformer, the early fault detection of the oil-immersed transformer is realized. However, the traditional DGA scheme will misread the gas near the comparison value boundary, resulting in inaccurate and relatively low reliability fault diagnosis results. At present, some researchers have proposed an oil-immersed transformer fault detection scheme based on artificial intelligence algorithms, but this type of detection scheme based on artificial intelligence algorithms also has the problems of high complexity and difficulty in implementation. Summary of the invention
[0004] One of the purposes of the present invention is to provide a method for detecting faults of an oil-immersed transformer which is highly reliable, accurate and easy to implement.
[0005] A second object of the present invention is to provide a system for implementing the oil-immersed transformer fault detection method.
[0006] The oil-immersed transformer fault detection method provided by the present invention comprises the following steps:
[0007] S1. Obtain gas dissolution analysis data of oil-immersed transformers in the early stage of fault and construct a training data set;
[0008] S2. Using the data obtained in step S1 as input, construct a decision tree model, calculate information gain and information gain ratio, and select nodes;
[0009] S3. construct a gradient enhancement model based on the gradient descent algorithm and the decision tree constructed in step S2;
[0010] S4. Initialize the gradient enhancement model constructed in step S3;
[0011] S5. Optimize the hyperparameters of the gradient boosting model obtained in step S4 using a Bayesian optimization scheme;
[0012] S6. Using the training data set, the gradient enhancement model is trained to obtain an oil-immersed transformer fault detection model;
[0013] S7. Using the oil-immersed transformer fault detection model obtained in step S6, actual oil-immersed transformer fault detection is performed.
[0014] The step S1 of obtaining the gas dissolution analysis data of the oil-immersed transformer in the early stage of a fault specifically includes the following steps:
[0015] Obtain gas dissolution analysis data of oil-immersed transformers in the early stages of faults to obtain the concentrations of methane, ethylene, acetylene, and hydrogen;
[0016] The relative concentrations of methane, ethylene, acetylene and hydrogen are calculated:
[0017]
[0018]
[0019] Where %CH 4 is the relative concentration of methane; %C 2 H 4 is the relative concentration of ethylene; %C 2 H 2 is the relative concentration of acetylene; %H 2 is the relative concentration of hydrogen; [CH 4 ] is the measured methane concentration; [C 2 H 4 ] is the measured ethylene concentration; [C 2 H 2 ] is the measured acetylene concentration; [H 2 ] is the measured hydrogen concentration value.
[0020] The step S2 uses the data obtained in step S1 as input to construct a decision tree model, calculate information gain and information gain ratio, and select nodes, which specifically includes the following steps:
[0021] Construct a decision tree model, and use the data obtained in step S1 as input to the decision tree model;
[0022] The information gain and information gain ratio are calculated using the following formula:
[0023] gain(S,A)=entroy(S)-entroy(S|A)
[0024]
[0025] Where gain(S,A) is the amount of %CH 4 、%C2 H 4 、%C 2 H 2 、%H 2 The information gain corresponding to the S label is calculated under the four relative concentration attributes; S is the set label, i.e., whether there is a fault, abnormal concentration, etc.; A is %CH 4 、%C 2 H 4 、%C 2 H 2 、%H 2 Four relative concentration attributes; entroy is the information entropy function; entroy(S) is the information entropy corresponding to label S; entroy(S|A) is the corresponding %CH 4 、%C 2 H 4 、%C 2 H 2 、%H 2 Conditional entropy of the four relative concentration attributes with respect to the S label; entropy(A) is the %CH calculated according to the calculation rules of information entropy. 4 、%C 2 H 4 、%C 2 H 2 、%H 2 Information entropy of four relative concentration attributes; gain_ratio(S,A) is %CH 4 、%C 2 H 4 、%C 2 H 2 、%H 2 Information gain ratio of four relative concentration attributes with respect to S label;
[0026] According to the obtained information gain ratio, the most relevant samples are weighted over other samples to select the nodes of the decision tree.
[0027] The step S3 constructs a gradient enhancement model based on the gradient descent algorithm and the decision tree constructed in step S2, which specifically includes the following steps:
[0028] Select the relative concentration of acetylene as the root node;
[0029] After selecting a root node, temporarily move the root node to a lower priority, allowing other attributes to continue partitioning the tree;
[0030] Select the relative concentration of ethylene as the second-level node;
[0031] For each branch, the most relevant attribute is selected as the next level node again, and the above segmentation process is repeated until the stopping condition is met, the predetermined depth of the first decision tree is reached, all sample attributes belong to the same category, or segmentation cannot be continued, thereby generating the first tree of the gradient enhancement model.
[0032] Step S4 initializes the gradient enhancement model constructed in step S3, specifically including the following steps:
[0033] Let F / (x) be the approximate function of the response y based on a set of predictor variables x, and express the multinomial logistic regression function as the loss function for detecting different fault categories; the approximate function is estimated as
[0034]
[0035] Where L(y,F / (x)) is the loss function of a set of prediction variables x corresponding to different fault category responses y; m is the number of prediction variables in the set; y (i) is the fault category response corresponding to the i-th predictor variable; x (i) is the i-th predictor variable;
[0036] The calculation formula of F / (x) is expressed as
[0037]
[0038] Where P(y=N|x; θ) is the function corresponding to the Nth type of fault response based on the prediction variable x and the model parameter θ; N is the number of classes; θ is the gradient enhancement model parameter; θ(j) is the gradient enhancement model parameter corresponding to the jth parameter vector; θ( N ) is the N-type parameter vector of the gradient boosting model;
[0039] Will As the training data set, the decision tree of the gradient boosting model constructs M different independent regression trees h 1 (x),h 2 (x),...,h M (x); the approximate function F is calculated using the following formula m / (x) and the gradient descent step size ρ m / To update:
[0040]
[0041] Where F m-1 / (x) is the approximate function of the response y based on a set of predictor variables x; J is the number of splits in each regression tree; b jm is a constant; I(x∈R jm ) takes the value 0 or 1 if x∈R jm , then I(x∈Rjm )=1, otherwise I(x∈R jm )=0;R jm is the region where the regression tree is located; ρ m is the gradient descent step length; L() is the estimated value of the approximate function;
[0042] The logistic regression tree of the gradient boosting model develops the model in chronological order and updates the model by minimizing the loss function; the logistic regression tree of the gradient boosting model adopts a shrinkage strategy to prevent overfitting and improve prediction accuracy; the shrinkage is defined as the learning rate, and the scaling factor ξ is introduced to help scale the contribution of each base tree model, expressed as
[0043]
[0044] Where γ jm To minimize the loss function; the value range of ξ is 0<ξ≤1;
[0045] Based on the shrinkage value and tree complexity, the optimal number of trees is obtained by checking how well the model fits the dataset; furthermore, the optimal performance of the logistic regression tree for the gradient boosting model is determined based on the choice of learning rate and tree complexity combination.
[0046] The Bayesian optimization scheme described in step S5 is used to optimize the hyperparameters of the gradient enhancement model obtained in step S4, which specifically includes the following steps:
[0047] Construct a Bayesian optimization model and optimize the hyperparameters of the gradient enhancement model obtained in step S4; the hyperparameters include the number of estimators, the maximum depth, the minimum sample split, the minimum sample leaf (controlling the complexity and generalization ability of the model by limiting the minimum number of samples of the decision tree leaf nodes) and the learning rate;
[0048] The Bayesian optimization model uses the posterior information of the objective function and Gaussian process regression to iteratively approximate the input-output relationship of the target system; based on the optimization algorithm of the sequence model, the Bayesian optimization process is iterated to determine the next optimal point; in order to obtain the global optimality of the objective function, an acquisition function that depends on previous observations and ultimately maximizes over-iteration is used;
[0049] In the optimal value search process, the expected improvement is used as the acquisition function, expressed as
[0050]
[0051] Where EI(x) is the expected improvement; μ(x) is the expected prediction; f(x+) is the maximum value reached in the current iteration of the function; x +is the design point of the maximum value reached in the current iteration of the function; Φ(Z) is the cumulative distribution function of the standard normal distribution; σ(x) is the predicted variance; φ(Z) is the probability density function of the standard normal distribution; Z is the residual, and x is a query point, and x=argmaxxE(max{0,f(x)-f(x+)});
[0052] The query point x is used to achieve the maximum improvement of f(x); f(x) is an iterative function.
[0053] The step S6 uses a training data set to train the gradient enhancement model to obtain an oil-immersed transformer fault detection model, which specifically includes the following steps:
[0054] The optimal hyperparameters of the gradient enhancement model are obtained through the Bayesian optimization model. Under the set learning rate, the optimal values of other parameters of the gradient enhancement model are determined; the obtained optimal values are used for training the gradient enhancement model;
[0055] The final probability prediction is obtained by taking the sum of the leaf values of all trees and then converting between 0 and 1 using a sigmoid function. The confusion matrix of the training data represents the prediction success value of the developed model. The correct prediction of each fault detected is added to the total number of all detected faults to obtain the overall prediction result.
[0056] Finally, the training of the gradient enhancement model is completed, and the trained gradient enhancement model is used as the oil-immersed transformer fault detection model.
[0057] The present invention also provides a system for implementing the oil-immersed transformer fault detection method, comprising a data acquisition module, a decision tree construction module, a model construction module, a model initialization module, a parameter optimization module, a model training module and a fault detection module; the data acquisition module, the decision tree construction module, the model construction module, the model initialization module, the parameter optimization module, the model training module and the fault detection module are connected in series in sequence; the data acquisition module is used to obtain the gas dissolution analysis data of the oil-immersed transformer in the early stage of the fault, and to construct a training data set, and upload the data information to the decision tree construction module; the decision tree construction module is used to construct a decision tree model based on the received data information, using the acquired data as input, calculate the information gain and the information gain ratio, and select nodes, and upload the data information to the model construction module; the model construction module is used to Data information, based on the gradient descent algorithm and the constructed decision tree, constructs a gradient enhancement model, and uploads the data information to the model initialization module; the model initialization module is used to initialize the constructed gradient enhancement model according to the received data information, and upload the data information to the parameter optimization module; the parameter optimization module is used to optimize the hyperparameters of the obtained gradient enhancement model using the Bayesian optimization scheme according to the received data information, and upload the data information to the model training module; the model training module is used to train the gradient enhancement model according to the received data information using the training data set to obtain the oil-immersed transformer fault detection model, and upload the data information to the fault detection module; the fault detection module is used to perform actual oil-immersed transformer fault detection according to the received data information using the obtained oil-immersed transformer fault detection model.
[0058] The oil-immersed transformer fault detection method and system provided by the present invention use the gas dissolution analysis data of the oil-immersed transformer in the early stage of the fault as the input of the decision tree, and construct a gradient enhancement model of the decision tree to realize fault detection, and also use a Bayesian optimization scheme to optimize the hyperparameters of the gradient enhancement model; therefore, the present invention can not only realize the fault detection of the oil-immersed transformer, but also has higher reliability and better accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0059] Figure 1 The figure is a schematic diagram of the method flow of the present invention.
[0060] Figure 2 Schematic diagram of the functional modules of the system of the present invention. DETAILED DESCRIPTION
[0061] like Figure 1 The method flow chart of the method of the present invention is shown as follows: The oil-immersed transformer fault detection method disclosed in the present invention comprises the following steps:
[0062] S1. Obtain gas dissolution analysis data of oil-immersed transformers in the early stage of fault and construct a training data set; specifically, the steps include:
[0063] Obtain gas dissolution analysis data of oil-immersed transformers in the early stages of faults to obtain the concentrations of methane, ethylene, acetylene, and hydrogen;
[0064] The relative concentrations of methane, ethylene, acetylene and hydrogen are calculated:
[0065]
[0066]
[0067] Where %CH 4 is the relative concentration of methane; %C 2 H 4 is the relative concentration of ethylene; %C 2 H 2 is the relative concentration of acetylene; %H 2 is the relative concentration of hydrogen; [CH 4 ] is the measured methane concentration; [C 2 H 4 ] is the measured ethylene concentration; [C 2 H 2 ] is the measured acetylene concentration; [H 2 ] is the measured hydrogen concentration value.
[0068] The dataset increases the number of problem variables and expands the input space using the relationship between attributes to improve the accuracy of power transformer fault diagnosis;
[0069] S2. Using the data obtained in step S1 as input, building a decision tree model, calculating information gain and information gain ratio, and selecting nodes; specifically comprising the following steps:
[0070] Construct a decision tree model, and use the data obtained in step S1 as input to the decision tree model;
[0071] The information gain and information gain ratio are calculated using the following formula:
[0072] gain(S,A)=entroy(S)-entroy(S|A)
[0073]
[0074] Where gain(S,A) is the amount of %CH 4 、%C 2 H 4 、%C 2 H 2 、%H2 The information gain corresponding to the S label is calculated under the four relative concentration attributes; S is the set label, i.e., whether there is a fault, abnormal concentration, etc.; A is %CH 4 、%C 2 H 4 、%C 2 H 2 、%H 2 Four relative concentration attributes; entroy is the information entropy function; entroy(S) is the information entropy corresponding to label S; entroy(S|A) is the corresponding %CH 4 、%C 2 H 4 、%C 2 H 2 、%H 2 Conditional entropy of the four relative concentration attributes with respect to the S label; entropy(A) is the %CH calculated according to the calculation rules of information entropy. 4 、%C 2 H 4 、%C 2 H 2 、%H 2 Information entropy of four relative concentration attributes; gain_ratio(S,A) is %CH 4 、%C 2 H 4 、%C 2 H 2 、%H 2 Information gain ratio of four relative concentration attributes with respect to S label;
[0075] According to the obtained information gain ratio, the most relevant samples are weighted to other samples to select the nodes of the decision tree;
[0076] S3. Based on the gradient descent algorithm and the decision tree constructed in step S2, a gradient enhancement model is constructed; specifically, the steps include:
[0077] The relative concentration of acetylene is selected as the root node, which has the highest gain ratio and the lowest entropy of the corresponding attribute set inserted;
[0078] After selecting a root node, temporarily move the root node to a lower priority, allowing other attributes to continue partitioning the tree;
[0079] In order to select the attribute corresponding to the second stage of the tree, the algorithm again considers the available parameters to calculate the gain ratio, and through the analysis of the calculation results, the relative concentration of ethylene is selected as the second level node;
[0080] For each branch, the most relevant attribute is selected as the next level node again, and the above segmentation process is repeated until the stopping condition is met, the predetermined depth of the first decision tree is reached, all sample attributes belong to the same category, or segmentation cannot be continued, thereby generating the first tree of the gradient boosting model;
[0081] S4. Initialize the gradient enhancement model constructed in step S3; specifically, the following steps are included:
[0082] Gradient boosting model is an ensemble method that fits successive new models according to the learning process to provide more accurate estimates of the response variable; the basic idea of gradient boosting model is to develop a new base learner that is maximally correlated with the negative gradient of the loss function in the complete ensemble process;
[0083] Let F / (x) be the approximate function of the response y based on a set of predictor variables x, and express the multinomial logistic regression function as the loss function for detecting different fault categories; the approximate function is estimated as
[0084]
[0085] Where L(y,F / (x)) is the loss function of a set of prediction variables x corresponding to different fault category responses y; m is the number of prediction variables in the set; y (i) is the fault category response corresponding to the i-th predictor variable; x (i) is the i-th predictor variable;
[0086] The calculation formula of F / (x) is expressed as
[0087]
[0088] Where P(y=N|x; θ) is the function corresponding to the Nth type of fault response based on the prediction variable x and the model parameter θ; N is the number of classes; θ is the gradient enhancement model parameter; θ(j) is the gradient enhancement model parameter corresponding to the jth parameter vector; θ( N ) is the N-type parameter vector of the gradient boosting model;
[0089] Will As the training data set, the decision tree of the gradient boosting model constructs M different independent regression trees h 1 (x),h 2 (x),...,h M (x); the approximate function F is calculated using the following formula m / (x) and the gradient descent step size ρ m / To update:
[0090]
[0091] Where Fm-1 / (x) is the approximate function of the response y based on a set of predictor variables x; J is the number of splits in each regression tree; b jm is a constant; I(x∈R jm ) takes the value 0 or 1 if x∈R jm , then I(x∈R jm )=1, otherwise I(x∈R jm )=0;R jm is the region where the regression tree is located; ρ m is the gradient descent step length; L() is the estimated value of the approximate function;
[0092] The logistic regression tree of the gradient boosting model develops the model in chronological order and updates the model by minimizing the loss function; the logistic regression tree of the gradient boosting model adopts a shrinkage strategy to prevent overfitting and improve prediction accuracy; the shrinkage is defined as the learning rate, and the scaling factor ξ is introduced to help scale the contribution of each base tree model, expressed as
[0093]
[0094] Where γ jm To minimize the loss function; the value range of ξ is 0<ξ≤1;
[0095] Based on the shrinkage value and tree complexity, the optimal number of trees is obtained by checking how well the model fits the dataset. In addition, the optimal performance of the logistic regression tree for the gradient boosting model is determined based on the choice of learning rate and tree complexity combination.
[0096] S5. Optimize the hyperparameters of the gradient enhancement model obtained in step S4 using a Bayesian optimization scheme; specifically, the following steps are included:
[0097] Construct a Bayesian optimization model and optimize the hyperparameters of the gradient enhancement model obtained in step S4; the hyperparameters include the number of estimators, the maximum depth, the minimum sample split, the minimum sample leaf (controlling the complexity and generalization ability of the model by limiting the minimum number of samples of the decision tree leaf nodes) and the learning rate;
[0098] The Bayesian optimization model uses the posterior information of the objective function and Gaussian process regression to iteratively approximate the input-output relationship of the target system; based on the optimization algorithm of the sequence model, the Bayesian optimization process is iterated to determine the next optimal point; in order to obtain the global optimality of the objective function, an acquisition function that depends on previous observations and ultimately maximizes over-iteration is used;
[0099] In the optimal value search process, the expected improvement is used as the acquisition function, expressed as
[0100]
[0101] Where EI(x) is the expected improvement; μ(x) is the expected prediction; f(x+) is the maximum value reached in the current iteration of the function; x + is the design point of the maximum value reached in the current iteration of the function; Φ(Z) is the cumulative distribution function of the standard normal distribution; σ(x) is the predicted variance; φ(Z) is the probability density function of the standard normal distribution; Z is the residual, and x is a query point, and x = argmaxxE(max { 0,f(x)-f(x+)});
[0102] The query point x is used to achieve the maximum improvement of f(x); f(x) is an iterative function;
[0103] S6. Using the training data set, the gradient enhancement model is trained to obtain an oil-immersed transformer fault detection model; specifically comprising the following steps:
[0104] The optimal hyperparameters of the gradient enhancement model are obtained through the Bayesian optimization model. Under the set learning rate, the optimal values of other parameters of the gradient enhancement model are determined; the obtained optimal values are used for training the gradient enhancement model;
[0105] The final probability prediction is obtained by taking the sum of the leaf values of all trees and then converting between 0 and 1 using a sigmoid function. The confusion matrix of the training data represents the prediction success value of the developed model. The correct prediction of each fault detected is added to the total number of all detected faults to obtain the overall prediction result.
[0106] Finally, the training of the gradient enhancement model is completed, and the trained gradient enhancement model is used as the oil-immersed transformer fault detection model;
[0107] S7. Using the oil-immersed transformer fault detection model obtained in step S6, actual oil-immersed transformer fault detection is performed.
[0108] like Figure 2The functional module schematic diagram of the system of the present invention is shown as follows: the system disclosed in the present invention for implementing the oil-immersed transformer fault detection method comprises a data acquisition module, a decision tree construction module, a model construction module, a model initialization module, a parameter optimization module, a model training module and a fault detection module; the data acquisition module, the decision tree construction module, the model construction module, the model initialization module, the parameter optimization module, the model training module and the fault detection module are connected in series in sequence; the data acquisition module is used to obtain the gas dissolution analysis data of the oil-immersed transformer in the early stage of the fault, and to construct a training data set, and upload the data information to the decision tree construction module; the decision tree construction module is used to construct a decision tree model based on the received data information, taking the acquired data as input, calculate the information gain and the information gain ratio, and select nodes, and upload the data information to the model construction module; the model construction module It is used to construct a gradient enhancement model based on the gradient descent algorithm and the constructed decision tree according to the received data information, and upload the data information to the model initialization module; the model initialization module is used to initialize the constructed gradient enhancement model according to the received data information, and upload the data information to the parameter optimization module; the parameter optimization module is used to optimize the hyperparameters of the obtained gradient enhancement model according to the received data information using the Bayesian optimization scheme, and upload the data information to the model training module; the model training module is used to train the gradient enhancement model according to the received data information using the training data set to obtain the oil-immersed transformer fault detection model, and upload the data information to the fault detection module; the fault detection module is used to perform actual oil-immersed transformer fault detection according to the received data information using the obtained oil-immersed transformer fault detection model.
Claims
1. A method for detecting faults of an oil-immersed transformer, comprising the following steps: S1. Obtain gas dissolution analysis data of oil-immersed transformers in the early stage of fault and construct a training data set; S2. Using the data obtained in step S1 as input, construct a decision tree model, calculate information gain and information gain ratio, and select nodes; S3. construct a gradient enhancement model based on the gradient descent algorithm and the decision tree constructed in step S2; S4. Initialize the gradient enhancement model constructed in step S3; S5. Optimize the hyperparameters of the gradient boosting model obtained in step S4 using a Bayesian optimization scheme; S6. Using the training data set, the gradient enhancement model is trained to obtain an oil-immersed transformer fault detection model; S7. Using the oil-immersed transformer fault detection model obtained in step S6, actual oil-immersed transformer fault detection is performed.
2. The oil-immersed transformer fault detection method according to claim 1 is characterized in that The step S1 of obtaining the gas dissolution analysis data of the oil-immersed transformer in the early stage of a fault specifically includes the following steps: Obtain gas dissolution analysis data of oil-immersed transformers in the early stages of faults to obtain the concentrations of methane, ethylene, acetylene, and hydrogen; The relative concentrations of methane, ethylene, acetylene and hydrogen are calculated: In the formula, %CH4 is the relative concentration of methane; %C2H4 is the relative concentration of ethylene; %C2H2 is the relative concentration of acetylene; %H2 is the relative concentration of hydrogen; [CH4] is the measured methane concentration value; [C2H4] is the measured ethylene concentration value; [C2H2] is the measured acetylene concentration value; [H2] is the measured hydrogen concentration value.
3. The oil-immersed transformer fault detection method according to claim 2 is characterized in that The step S2 uses the data obtained in step S1 as input to construct a decision tree model, calculate information gain and information gain ratio, and select nodes, which specifically includes the following steps: Construct a decision tree model, and use the data obtained in step S1 as input to the decision tree model; The information gain and information gain ratio are calculated using the following formula: gain(S,A)=entroy(S)-entroy(S|A) Where gain(S,A) is the information gain corresponding to the S label calculated under the four relative concentration attributes of %CH4, %C2H4, %C2H2, and %H2; S is the set label, that is, whether there is a fault, abnormal concentration, etc.; A is the four relative concentration attributes of %CH4, %C2H4, %C2H2, and %H2; entroy is the information entropy function; entroy(S) is the information entropy corresponding to the label S; entroy(S|A) is the conditional entropy corresponding to the four relative concentration attributes of %CH4, %C2H4, %C2H2, and %H2 with respect to the S label; entroy(A) is the information entropy of the four relative concentration attributes of %CH4, %C2H4, %C2H2, and %H2 calculated according to the calculation rules of information entropy; gain_ratio(S,A) is the information gain rate of the four relative concentration attributes of %CH4, %C2H4, %C2H2, and %H2 with respect to the S label; According to the obtained information gain ratio, the most relevant samples are weighted over other samples to select the nodes of the decision tree.
4. The oil-immersed transformer fault detection method according to claim 3 is characterized in that The step S3 constructs a gradient enhancement model based on the gradient descent algorithm and the decision tree constructed in step S2, which specifically includes the following steps: Select the relative concentration of acetylene as the root node; After selecting a root node, temporarily move the root node to a lower priority, allowing other attributes to continue partitioning the tree; Select the relative concentration of ethylene as the second-level node; For each branch, the most relevant attribute is selected as the next level node again, and the above segmentation process is repeated until the stopping condition is met, the predetermined depth of the first decision tree is reached, all sample attributes belong to the same category, or segmentation cannot be continued, thereby generating the first tree of the gradient enhancement model.
5. The oil-immersed transformer fault detection method according to claim 4 is characterized in that Step S4 initializes the gradient enhancement model constructed in step S3, specifically including the following steps: Let F / (x) be the approximate function of the response y based on a set of predictor variables x, and express the multinomial logistic regression function as the loss function for detecting different fault categories; the approximate function is estimated as Where L(y,F / (x)) is the loss function of a set of prediction variables x corresponding to different fault category responses y; m is the number of prediction variables in the set; y (i) is the fault category response corresponding to the i-th predictor variable; x (i) is the ith predictor variable. The calculation formula of F / (x) is expressed as Where P(y=N|x;θ) is the function corresponding to the Nth type of fault response based on the prediction variable x and the model parameter θ; N is the number of classes; θ is the gradient enhancement model parameter; θ(j) is the gradient enhancement model parameter corresponding to the jth parameter vector; θ (N) is the N-type parameter vector of the gradient boosting model; Will As the training data set, the decision tree of the gradient boosting model constructs M different independent regression trees h1(x),h2(x),...,h M (x); the approximate function F is calculated using the following formula m / (x) and the gradient descent step size ρ m / To update: Where F m-1 / (x) is the approximate function of the response y based on a set of predictor variables x; J is the number of splits in each regression tree; b jm is a constant; I(x∈R jm ) takes the value 0 or 1 if x∈R jm , then I(x∈R jm )=1, otherwise I(x∈R jm )=0;R jm is the region where the regression tree is located; ρ m is the gradient descent step length; L() is the estimated value of the approximate function; The logistic regression tree of the gradient boosting model develops the model in chronological order and updates the model by minimizing the loss function; the logistic regression tree of the gradient boosting model adopts a shrinkage strategy to prevent overfitting and improve prediction accuracy; the shrinkage is defined as the learning rate, and the scaling factor ξ is introduced to help scale the contribution of each base tree model, expressed as Where γ jm To minimize the loss function; the value range of ξ is 0<ξ≤1; Based on the shrinkage value and tree complexity, the optimal number of trees is obtained by checking how well the model fits the dataset; Furthermore, the optimal performance of the logistic regression tree for the gradient boosting model is determined based on the choice of learning rate and tree complexity combination.
6. The oil-immersed transformer fault detection method according to claim 5 is characterized in that The Bayesian optimization scheme described in step S5 is used to optimize the hyperparameters of the gradient enhancement model obtained in step S4, which specifically includes the following steps: Constructing a Bayesian optimization model to optimize hyperparameters of the gradient enhancement model obtained in step S4; the hyperparameters include the number of estimators, maximum depth, minimum sample split, minimum sample leaf, and learning rate; The Bayesian optimization model uses the posterior information of the objective function and Gaussian process regression to iteratively approximate the input-output relationship of the target system; based on the optimization algorithm of the sequence model, the Bayesian optimization process is iterated to determine the next optimal point; in order to obtain the global optimality of the objective function, an acquisition function that depends on previous observations and ultimately maximizes over-iteration is used; In the optimal value search process, the expected improvement is used as the acquisition function, expressed as Where EI(x) is the expected improvement; μ(x) is the expected prediction; f(x + ) is the maximum value reached in the current iteration of the function; x + is the design point of the maximum value reached in the current iteration of the function; Φ(Z) is the cumulative distribution function of the standard normal distribution; σ(x) is the predicted variance; φ(Z) is the probability density function of the standard normal distribution; Z is the residual, and x is the query point, and x = argmax x E(max{0,f(x)-f(x + )}); The query point x is used to achieve the maximum improvement of f(x); f(x) is an iterative function.
7. The oil-immersed transformer fault detection method according to claim 6 is characterized in that The step S6 uses a training data set to train the gradient enhancement model to obtain an oil-immersed transformer fault detection model, which specifically includes the following steps: The optimal hyperparameters of the gradient enhancement model are obtained through the Bayesian optimization model. Under the set learning rate, the optimal values of other parameters of the gradient enhancement model are determined; the obtained optimal values are used for training the gradient enhancement model; Take the sum of leaf values of all trees and then use a sigmoid function to convert between 0 and 1 to get the final probability prediction; The confusion matrix of the training data represents the prediction success value of the developed model. The correct prediction of each detected fault is added to the total number of all detected faults to obtain the overall prediction result. Finally, the training of the gradient enhancement model is completed, and the trained gradient enhancement model is used as the oil-immersed transformer fault detection model.
8. A system for implementing the oil-immersed transformer fault detection method according to any one of claims 1 to 7, characterized in that It includes a data acquisition module, a decision tree construction module, a model construction module, a model initialization module, a parameter optimization module, a model training module and a fault detection module; the data acquisition module, the decision tree construction module, the model construction module, the model initialization module, the parameter optimization module, the model training module and the fault detection module are connected in series in sequence; the data acquisition module is used to obtain the gas dissolution analysis data of the oil-immersed transformer in the early stage of the fault, and to construct a training data set, and upload the data information to the decision tree construction module; The decision tree construction module is used to construct a decision tree model based on the received data information, taking the acquired data as input, calculating the information gain and the information gain ratio, selecting nodes, and uploading the data information to the model construction module; The model building module is used to build a gradient enhancement model based on the received data information, based on the gradient descent algorithm and the constructed decision tree, and upload the data information to the model initialization module; The model initialization module is used to initialize the constructed gradient enhancement model according to the received data information, and upload the data information to the parameter optimization module; The parameter optimization module is used to optimize the hyperparameters of the obtained gradient enhancement model using a Bayesian optimization scheme based on the received data information, and upload the data information to the model training module; The model training module is used to train the gradient enhancement model according to the received data information using the training data set to obtain the oil-immersed transformer fault detection model, and upload the data information to the fault detection module; The fault detection module is used to perform actual fault detection of the oil-immersed transformer based on the received data information and using the obtained oil-immersed transformer fault detection model.