Power grid fault diagnosis method based on adasyn-dhdsd-et

Abstract

This invention discloses a power grid fault diagnosis method based on ADASYN-DHSD-ET. The method comprises: 1) extracting power grid monitoring data from a database, preprocessing it, and retaining records with complete attributes; 2) classifying the preprocessed data, and randomly sampling each class according to a certain proportion to obtain a training set D. train and test set D test 3. Using the ADASYN-DHSD-ET algorithm to analyze D train Training the fault type diagnosis training model; 4: Using D test The training model is validated by setting each parameter and the optimal model is obtained based on the evaluation index. Step 5: Repeat step 1 and identify the fault types in the power grid based on the optimal model. This invention addresses the problem of unbalanced distribution of categories in power grid fault diagnosis data by constructing a model based on the ADASYN-DHSD-ET algorithm and then obtaining the optimal model based on the corresponding evaluation index. This model can more accurately identify the fault types in the power grid and enhance the stability and reliability of the safe operation of the power system.

Description

TECHNICAL FIELD

[0001] The application relates to the field of power system fault maintenance, in particular to a power grid fault diagnosis method based on ADASYN-DHSD-ET. BACKGROUND

[0002] With the development of China's economy, the demand for electricity is increasing, and it is particularly important to ensure the safe and stable operation of the power system and effectively diagnose and predict power grid faults. In recent years, with the continuous development of computers and diagnostic technologies, power grid fault diagnosis technology has also been rapidly improved. Domestic and foreign scholars have used intelligent fault diagnosis technologies such as expert systems, neural networks, Petri nets, machine learning and deep learning. Although good results have been achieved in fault diagnosis, there are problems such as harsh prerequisites (high data preprocessing requirements), local under-learning and over-learning. However, in the sample, if normal data accounts for the majority and fault data accounts for a small part, normal and fault data are large and small class samples respectively, that is, typical unbalanced data.

[0003] Currently, there are two main aspects to solve the problem of unbalanced data classification: ① data level, up-sampling and down-sampling, changing data distribution, reducing imbalance, called resampling method; ② algorithm level, analyzing the defects of existing algorithms in the face of unbalanced data, improving the algorithm or proposing a new algorithm to improve the classification accuracy of the small class, such as cost-sensitive learning and ensemble learning.

[0004] For unbalanced sample data in fault diagnosis, many domestic and foreign scholars have studied the application in the field of fault diagnosis under its characteristics. Some have proposed a rotating machinery fault diagnosis algorithm based on fast clustering and support vector machine, which reduces data through fast clustering and balances it using support vector machine for training, with good diagnostic effect. Some propose a rotating machinery unbalanced data fault diagnosis method that combines up-sampling and feature learning, which balances data distribution using weighted up-sampling and selects features using enhanced automatic coding, which can more effectively detect fault samples.

[0005] Unbalanced learning methods have good results in the field of fault diagnosis, but they are less used in the field of power grid fault diagnosis. Power grid fault diagnosis based on electrical quantity information (voltage, current, power consumption, power, phase, frequency, etc.) (Zhao Yao. Power grid fault diagnosis method based on information fusion[D]. Shandong: Shandong University, 2018.) can provide more accurate fault diagnosis and has strong interpretability. However, in actual situations, electrical measurement is difficult, and there are many reasons for power grid faults, such as equipment aging, power load, weather conditions such as severe weather and lightning, and wireless communication, which can all affect precision power grid equipment; some of the reasons are irregular and may include many unforeseen circumstances.

[0006] At present, the improved model based on SMOTE method is mostly used in the study of unbalanced data, but SMOTE method is easy to cause the generated samples of small class to overlap, because the generated samples are blindly generated in the same number, ignoring the distribution characteristics of the adjacent samples. The number of samples generated by adaptive synthetic sampling (ADASYN) is calculated according to the density distribution of each small class, which can better enhance the learning ability of the classification model. Figure 1 In order to generate simulated samples for each sampling algorithm, it is concluded that SMOTE sampling generates new samples by linear interpolation between sample points; SVMSMOTE sampling generates new samples based on the hyperplane of SVM; BorderlineSMOTE sampling generates samples near the boundary points of small classes; and ADASYN sampling generates more samples at the sample aggregation of small classes. In the integrated algorithm, the extremely randomized tree algorithm (ET) can effectively reduce the bias and variance of classification, and has good effect on the application of small classes.

[0007] The content of the application is:

[0008] The technical problem to be solved by the application is to overcome the shortcomings of the prior art, aiming at the class distribution imbalance problem of power grid fault diagnosis data, that is, the ratio of fault class to normal class is small, by extracting the daily monitoring data in the multi-element database, after data preprocessing, a model based on ADASYN-DHSD-ET algorithm is constructed, then the optimal model is obtained according to the corresponding evaluation index, and finally the power grid fault type is identified.

[0009] The technical scheme of the application is: a power grid fault diagnosis method based on ADASYN-DHSD-ET, characterized by: extracting the daily monitoring data in the multi-element database, after data preprocessing, a model based on ADASYN-DHSD-ET algorithm is constructed, then the optimal model is obtained according to the corresponding evaluation index, and finally the power grid fault type is identified, which has the following steps: step one, extracting the daily monitoring data of power grid in the multi-element database, after preprocessing, retaining all attribute complete records;

[0010] Step two, classify the preprocessed data, and randomly sample each class according to a certain proportion to obtain the training set D train And the test set D test ;

[0011] Step three, using ADASYN-DHSD-ET algorithm to D trainTraining a fault type diagnosis training model;

[0012] Step four: using D test Verify the training model, verify according to each parameter setting, and obtain the optimal model according to the evaluation index,

[0013] Step five: repeat step one and identify the fault type in the power grid according to the optimal model obtained in step four.

[0014] Further, in the step one, the multivariate database is a historical storage database of power grid daily monitoring data or real-time monitoring data; the attribute complete record is a record with 11 characteristic attributes of voltage, current, whether repaired, working day, time.

[0015] Further, in the step two, according to normal data, signal loss fault, MIB backup misalignment fault, R loss fault and signal degradation fault, the data set is classified and formed.

[0016] Further, in the step three, the process of training model establishment is: (1) sampling the samples in the training set D train According to the density distribution of the fault class, new samples are automatically generated, the synthesis number, synthesis data and balance degree of the samples are calculated by DHSD, and the characteristics of the sampled data are kept;

[0017] (2) Parameter adjustment is performed on the number of new samples to make it the same as the number of normal class samples, and a new sample set is formed with the normal class samples;

[0018] (3) In the ET algorithm, each base classifier uses all samples in the new sample set for training, randomly selects n split attributes from N attributes when the base classifier node is split, selects the optimal attribute for splitting according to the Gini coefficient or information gain entropy, and does not prune in the splitting process until a base classifier is generated.

[0019] (4) Voting is performed on all base classifiers to obtain a training model.

[0020] Further, in the step four, the average precision, the average precision of small classes, F_score, G_mean and time complexity are used as the performance evaluation indexes of the unbalanced data set.

[0021] The beneficial effects of the present application are:

[0022] 1. The application aims at the class distribution imbalance problem of power grid fault diagnosis data, that is, the ratio of fault class to normal class is small, through extracting daily monitoring data in a multivariate database, constructing a model based on the ADASYN-DHSD-ET algorithm after data preprocessing, then obtaining the optimal model according to the corresponding evaluation index, and finally identifying the fault type of the power grid, which can accurately identify the fault type in the power grid and enhance the stability and reliability of the safe operation of the power system.

[0023] 2. The application combines the adaptive synthesis of ADASYN and the low deviation and variance of extreme random tree, and improves on this basis, proposes an ADASYN-DHSD-ET algorithm, improves the sample balance degree and the training method of the model, adaptively calculates the number of new samples according to the internal distribution density of the fault class samples in the sampling stage; when synthesizing new samples, the high-dimensional space distance between discrete data sample points is calculated, so that the synthesis number between points is negatively related to the distance, and new samples are synthesized; in the process of generating base classifiers, the features are randomly selected when the node is split, and the characteristics of strong randomness and low variance of the extreme random tree are ingeniously used to solve the influence of noise data.

[0024] 3. The application can improve randomness, reduce the deviation and variance of classification, overcome the overfitting of the model, and improve the generalization ability of the model; improve the accuracy of small classes, and ensure high accuracy of large classes.

[0025] 4. The application starts from a real multi-source database, and the preprocessed data contains normal and fault samples, and an imbalance classification method is used for research, which starts from actual data sources, has a complete detection route, and more comprehensively considers all characteristics of normal and fault data. The drawbacks of previous researches which only extract part of fault data for diagnosis are solved. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 Comparison of simulation samples generated by various sampling algorithms.

[0027] Figure 2 Modeling flowchart of the power grid fault diagnosis method based on ADASYN-DHSD-ET.

[0028] Figure 3 Performance comparison of traditional classification models and ADASYN-DHSD-ET algorithm in the application.

[0029] Figure 4 Performance comparison table of common imbalance classification models and ADASYN-DHSD-ET algorithm in the application. DETAILED DESCRIPTION

[0030] Embodiment: seeFigure 1 、 Figure 2 、 Figure 3 and Figure 4 .

[0031] The power grid fault diagnosis method based on ADASYN-DHSD-ET extracts daily monitoring data in a multivariate database, constructs a model based on the ADASYN-DHSD-ET algorithm after data preprocessing, obtains an optimal model according to corresponding evaluation indexes, and finally identifies the power grid fault type. ADASYN-DHSD-ET is the abbreviation of Adaptive Synthetic Sampling of Discrete High-dimensional Spatial Distance Extremely Randomized Trees.

[0032] The application will be described in detail below in combination with the drawings and examples.

[0033] The ADASYN-DHSD algorithm is an algorithm for improving ADASYN when generating data based on discrete high-dimensional spatial distance. This method considers the multi-class imbalance problem of data, and adaptively synthesizes new samples according to the distribution of small-class samples. When synthesizing new samples, the high-dimensional spatial distance is used to calculate the discrete feature data of the sample, and the number of synthesized sample points is determined according to the distance between the sample points. The closer the distance, the more sample points are synthesized. A matrix defining the high-dimensional spatial distance between the characteristic values of all feature vectors is created, and the distance δ between two feature vectors is defined as follows:

[0034]

[0035] In the formula: V1 and V2 are two characteristic values, C1 is the total number of times that the characteristic value V1 appears, C 1i is the number of times that the characteristic value V1 appears in class i, C2 and C 2i are similar to the above definition, and k is a constant. Equation (1) is used to calculate the difference matrix of each nominal characteristic value of the feature vector and give a certain geometric distance.

[0036] According to ADASYN, let the sample point x of the small class have t t (K) neighbor points satisfy K-neighbor, x needs to synthesize n new sample points, and the distance between the sample point and each neighbor point is δ1, δ2, … δt. The number of synthesized samples of the sample point and each neighbor point can be calculated. The generation ratio v i of x i is:

[0037] v i = 1 / δ i (2)

[0038] Regularization is:

[0039]

[0040] x and x i The number of generated nodes m i is:

[0041]

[0042] x in turn with each x i Synthetic m i New samples.

[0043] ET algorithm is an integrated algorithm. Its base classifier is trained using all samples, to enhance randomness, randomly select n features from N features when node splitting, select the optimal attribute for splitting with Gini coefficient or information gain entropy, no pruning in the splitting process, until a decision tree (base classifier) is generated. Use voting decision to count all base classifiers to produce the final classification results.

[0044] ET is better than greedy decision tree, has better smoothness on small samples, can effectively reduce bias and variance. To illustrate this point, consider a sample of size N:

[0045] ls N = {(x i , y i ): i = 1, 2,..., N}

[0046] In the formula: is an n-dimensional feature vector, y i is the corresponding output value, expressed as: j th The sample value of the attribute is obtained by increasing order, and the symbol is simplified as:

[0047]

[0048] And The hyper-interval of the characteristic function is:

[0049]

[0050] With these symbols, it can be proved that an infinite ET can be approximated as follows:

[0051]

[0052] In the formula: parameter Depends on the input sample xi and output y i and parameter n of the method min and K.

[0053] When n min = 2, the special case of complete tree, we have:

[0054]

[0055] If the input space is one-dimensional (n = 1 and x = (x1)), the above formula degenerates into a linear piecewise model:

[0056]

[0057] In the formula: I (i) (x1) is the characteristic function of the interval . and The values of and are derived from the derivation of equation (6), and the N continuous constraints are

[0058] ET is more continuous and smooth when the number of base classifiers M→∞, compared with other tree-based ensemble classification methods. From the perspective of bias and variance, the continuity of the model makes the variance and bias of the smooth region of the objective function smaller, so that the model in this region is more accurate.

[0059] The application uses the K nearest neighbors of each sample in the small class to calculate the distribution of the sample, uses DHSD to calculate the synthesis number between samples, synthesizes data, improves the balance degree of the sample, and enhances the training effect of the algorithm; meanwhile, each base classifier in the extreme random tree algorithm uses all the newly synthesized samples for training, randomly selects a split feature when splitting the base classifier node, calculates the optimal split attribute for splitting, and generates a base classifier until a base classifier is generated; finally, all base classifiers are voted to form the ADASYN-DHSD-ET algorithm.

[0060] Let the training sample set be D = {(x1, y1), …, (x m , y m )}, and the ADASYN-DHSD-ET algorithm is described as follows:

[0061] 1. Construct ADASYN_DHSD_ET(D)

[0062] Input training set D = {(x1, y1), …, (x m , y m )}

[0063] Output extreme random tree T = {t1, …, t M}

[0064] (1) for i = 1 to M do

[0065] (2) Generate decision tree, t i = Construct sub-classifier (D)

[0066] (3) end for

[0067] (4) Voting strategy

[0068] (5) Return extreme random tree T

[0069] 2. Construct sub-classifier (D)

[0070] Input training set D = {(x1, y1),..., (x m , y m )}

[0071] Output sub-classifier t

[0072] (1) Call ADASYN_DHSD sampling (D), return sampled dataset D new

[0073] (2) if stop splitting (D new ) then

[0074] (3) Return a leaf node;

[0075] (4) else

[0076] (5) Randomly select K attributes {a1,..., a K} from all candidate attributes;

[0077] (6) Generate K split thresholds {s1,..., s K}, where s i = select split point

[0078] (7) According to Score (s * , D new ) = max i＝1,...,K Score (s * , D new ), select the best test split threshold s * ;

[0079] (8) According to the test split threshold s * , divide the sample set D new into two sub-sample sets D l and D r ;

[0080] (9) Construct left child t l and right child t r using subsets D l and D l respectively. r r

[0081] (10) Create tree node t * according to s l , t r and t i are its left and right children respectively, and return decision tree t.

[0082] (11) End if

[0083] 3. ADASYN_DHSD sampling (D)

[0084] Input D is a training dataset, containing m samples {x i , y i}, x i is a n-dimensional feature set, y i ∈Y={1,…C} is a class set, representing the number of large classes with m l , the number of each small class with m s1 ,…,m sj . m si ≤m l , and ∑m si +m l =m.

[0085] Output sampled sample D new

[0086] (1) Loop each small class for i=1 to j do

[0087] (2) Calculate the imbalance degree d i of the small class:

[0088] d i =m si / m l , d i ∈(0,1] (8)

[0089] (3) If d i <d th (d th is the maximum threshold of class imbalance), then:

[0090] (4) Calculate the total number of small class samples that need to be synthesized G i , and the parameter β represents the imbalance degree of the small class sample after synthesis. β=1 means that the number of large class samples and the number of small class samples after synthesis are the same.

[0091] G i =(m l -m si )×β,β∈(0,1] (9)

[0092] (5)x i Let each sample in the subclass be a sample, and calculate its K-nearest neighbor ratio r in n-dimensional space. i , where Δ i It is x i The number of samples in the K-nearest neighbors, therefore r i ∈[0,1].

[0093] r i =Δ i / K,i=1,...m si (10)

[0094] (6) Regularization r i Then r i satisfy for:

[0095]

[0096] (7)g i Represents each subclass sample x i The desired number of synthetic samples, where G i It is the total number of synthetic samples in formula (9).

[0097]

[0098] (8) Use DHSD for each sample x i Generate g i Synthetic data.

[0099] 4. Stop dividing (D)

[0100] Input training set D

[0101] Output Boolean value

[0102] (1)if D <n min Then return TRUE;

[0103] (2) If all attributes in D are fixed, then return TRUE;

[0104] (3) If the output variable in D remains unchanged, then return TRUE;

[0105] (4) else return FALSE.

[0106] 5. Select split point (D, a)

[0107] Input training set D, attribute a

[0108] Output split attribute

[0109] (1) Calculate the maximum value of attribute a in training set D Minimum value

[0110] (2) Randomly select a split attribute a from ; c ;

[0111] (3) Return the split attribute [a < a c ].

[0112] The daily monitoring data of the power grid between September 2016 and April 2018 in Jiangxi Province are selected, and after preprocessing, the records with complete attributes are retained. The sample is shown in Table 1, and there are 30039 records, each record has 11 characteristic attributes, including voltage, current, whether repaired, repair duration, whether working day, month, date, day of the week, morning or afternoon, season, temperature and the like; the fault category is divided into 5 categories, as shown in Table 1, category 1 represents normal category, categories 2, 3, 4 and 5 represent signal loss fault category, MIB backup misalignment fault category, R loss fault category and signal degradation fault category respectively, wherein the data of the normal category is extremely large, and the data of the 4 fault categories is extremely small, and the category ratio is as shown in the table: 0.9891:0.0036:0.003:0.0025:0.0018; the characteristics of the sample are an extremely large category and multiple extremely small categories, and have obvious data imbalance characteristics.

[0113] Table 1 Characteristics of the data set

[0114]

[0115] The flowchart of the experiment of the application is shown in Figure 2 . First, data preprocessing (data cleaning, data integration) is performed based on the data of the multi-source database to obtain the above data set. The data set is randomly sampled according to a 6:4 ratio for each category to obtain a training set D train and a test set D test .

[0116] Then the ADASYN-DHSD-ET algorithm is used on D trainThe training fault diagnosis model, the algorithm oversamples four fault classes, generates new samples automatically according to the density distribution of the fault classes, changes the balance degree of the data set, and keeps the characteristics of the original samples after sampling; the number of generated new samples is adjusted, and the sampling number is set to be 10 times, 20 times, 50 times, 100 times and 200 times of the original sample in turn until the number of normal class samples; then each base classifier in the algorithm classifies the whole new data set, randomly selects a split feature in 11 features at node splitting, does not reduce branches, and fully considers the characteristics of the data set; the number of base classifiers is adjusted, and the number of generated base classifiers is set to be 100, 200, 500 and 1000 in turn; the integrated classifier votes to obtain the training model. Use D test Verify the training model, perform experiments according to each parameter setting, and obtain the optimal model according to the following evaluation indexes.

[0117] The application selects average accuracy (average accuracy, Avg_Acc), average accuracy (Avg_Acc m ) of small class, F_score, G_mean and time complexity (Time) as the performance evaluation indexes of the unbalanced data set.

[0118] The mixed matrix is expanded to a k (k>=2) class classification problem, and C1, C2, … C k represent k classes, let the number of samples actually correctly predicted as the kth class n kk , the number of samples predicted as the ith class n ki , and the mixed matrix is shown in Table 2.

[0119] Table 2: Extended confusion matrix

[0120]

[0121] The following evaluation indexes can be calculated from the mixed matrix in Table 2:

[0122] (1) Let P i be the precision of the ith class, which represents the proportion of n ii in the actual ith class sample :

[0123]

[0124] (2) Let R i be the recall rate of the ith class, which represents the proportion of n ii in all predicted ith class , that is, the correct rate of the ith class:

[0125]

[0126] (3) Overall accuracy (OA), i.e. the number of correct predictions In all samples The proportion of:

[0127]

[0128] (4) Average accuracy Avg_Acc, representing the arithmetic mean of the accuracy of each class:

[0129]

[0130] OA in the case of a large number of categories with high accuracy, cannot reflect the accuracy of small classes, and Avg_Acc can better reflect the accuracy of small classes. C1 is a large class, C2, … C k is each small class, the average accuracy (Avg_Acc m ) of each small class, representing the arithmetic mean of the accuracy of each small class:

[0131]

[0132] (5) Let F i be the harmonic mean of the recall and precision of the i-th class, and F_score represent the average of all classes F i :

[0133]

[0134]

[0135] (6) G_mean represents the geometric mean of the accuracy of each class:

[0136]

[0137] Various traditional classifiers and unbalanced classification methods are used to train and test the model and compared with the ADASYN-DHSD-ET model proposed in the application to obtain Tables 3 and 4, respectively.

[0138] The evaluation indicators are OA, Avg_Acc, R large (the accuracy of the large class), Avg_Acc m , F_score, G_mean and Time defined above. All experiments are averaged over 50 experiments to ensure repeatability and generality.

[0139] Figure 3The performance comparison results of the traditional classification model, the integrated classification model and the algorithm of the application are shown in Table 3. The traditional classification model includes multinomial naive Bayes classification (Bayes), nearest neighbor classification (KNN), decision tree classification (Decision Tree), support vector machine (SVM) and neural network multilayer perceptron (MLP). The integrated method includes Bagging, AdaBoost, Gradient Boosting, Random Forest and extreme random tree (ET). As shown in Table 3, the ADASYN-DHSD-ET model has obvious improvement in the performance indicators Avg_Acc and G_mean. Most of the algorithms perform well in the indicator R large m The model is more than 30% higher than other models in the indicators Avg_Acc m The model is more than 30% higher than other models in the indicators Avg_Acc large ET performs better in the indicator R

[0140] Figure 4 The algorithm of the application is compared with the current popular imbalance classification method. The imbalance classification method includes the combination of oversampling and ET model commonly used in imbalance: SMOTE-ET, SMOTENC-ET, SVMSMOTE-ET, BorderlineSMOTE-ET and ADASYN-ET. The imbalance integrated classification model takes internal subset sampling: BalancedRandomForest, RUSBoost, EasyEnsemble and BalancedBagging. As shown in Table 4, compared with the ET model combined with oversampling, the algorithm of the application is more than 30% higher than other models in the indicators Avg_Acc, F_score and Avg_Acc m ​The algorithm shows good performance in the identification of small classes. SMOTENC-ET is less different from the index of the present application, but the time consumption is too high, and the time complexity compared with the present application is 78.09:21.96. The time efficiency of the present application algorithm is higher. The present application algorithm and the unbalanced integrated classification are compared in the index Avg_Acc, Rlarge and G_mean. The present application algorithm has good performance in the identification of large classes and is not affected by up-sampling. Although BalancedRandomForest has good performance in the index Avg_Acc m The above is better than the model of the present application, but is significantly lower than the model of the present application in the other three indexes. This is because the data generated by BalancedRandomForest after up-sampling affects the identification rate of large classes, and performs poorly in the index Rlarge. The present application algorithm still maintains good identification rate in large classes. For power grid fault diagnosis, the performance of ADASYN-DHSD-ET model is better. The identification accuracy of large classes, i.e. normal operation classes, reaches 99%, and the average identification accuracy of small classes, i.e. fault classes, reaches 70%. The overall average accuracy reaches 84.39%, F_score reaches 84.67%, and G_mean reaches 82.6%. Therefore, the power grid fault diagnosis method based on ADASYN-DHSD-ET can accurately identify the fault type in the power grid, and enhances the stability and reliability of the safe operation of the power system.

[0141] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Any simple modification, equivalent change and modification made according to the technical essence of the present application to the above embodiment still belongs to the scope of the technical solution of the present application.

Claims

1. A power grid fault diagnosis method based on ADASYN-DHSD-ET, characterized by: The daily monitoring data in the multivariate database is extracted, and after data preprocessing, a model based on the ADASYN-DHSD-ET algorithm is constructed. Then, the optimal model is obtained according to the corresponding evaluation indicators. Finally, the power grid fault type is identified. The specific steps are as follows: Step 1: Extract the daily monitoring data of the power grid from the multivariate database, and after preprocessing, retain all records with complete attributes. Step 2: Classify the preprocessed data and randomly sample from each class according to a certain proportion to obtain the training set D. train and test set D test ; Step 3: Use the ADASYN-DHSD-ET algorithm to train the fault type diagnosis training model on Dtrain; In step three, the process of training the model is as follows: (1) The training set D is trained using ADASYN-DHSD. train The sample is sampled from the sample, and new samples are automatically generated according to the density distribution of fault classes. DHSD is used to calculate the number of composites between samples, composite data, improve the balance of samples, and make the sampled data retain the characteristics of the original samples. (2) Make the number of fault class samples the same as the number of normal class samples, and form a new sample set with the normal class samples; (3) In the ET algorithm, each base classifier is trained using all samples in the new sample set. When the base classifier node splits, n splitting attributes are randomly selected from N attributes. The optimal attribute is selected for splitting based on the Gini coefficient or information gain entropy. No pruning is performed during the splitting process until a base classifier is generated. (4) Vote for all base classifiers to obtain the training model; Step 4: Use D test Validate the trained model by performing validation for each parameter setting, and obtain the optimal model based on evaluation metrics. Step 5: Repeat Step 1 and identify the fault types in the power grid based on the optimal model obtained in Step 4.

2. The power grid fault diagnosis method based on ADASYN-DHSD-ET according to claim 1, characterized in that: In step one, the multi-dimensional database is a historical storage database of daily power grid monitoring data or real-time monitoring data; the record with complete attributes is a record with 11 characteristic attributes: voltage, current, whether it has been repaired, repair duration, whether it is a working day, month, date, day of the week, morning or afternoon, season, and temperature.

3. The power grid fault diagnosis method based on ADASYN-DHSD-ET according to claim 1, characterized in that: In step two, the data is classified into normal data, signal loss fault, MIB backup misalignment fault, R loss fault, and signal degradation fault, and a dataset is formed.

4. The power grid fault diagnosis method based on ADASYN-DHSD-ET according to claim 1, characterized in that: In step four, average precision, average precision of minor classes, F_score, G_mean, and time complexity are used as performance evaluation metrics for imbalanced datasets.

Images