CN116842337A - Transformer fault diagnosis method based on LightGBM (gallium nitride based) optimal characteristics and COA-CNN (chip on board) model - Google Patents

Transformer fault diagnosis method based on LightGBM (gallium nitride based) optimal characteristics and COA-CNN (chip on board) model Download PDF

Info

Publication number
CN116842337A
CN116842337A CN202310697655.8A CN202310697655A CN116842337A CN 116842337 A CN116842337 A CN 116842337A CN 202310697655 A CN202310697655 A CN 202310697655A CN 116842337 A CN116842337 A CN 116842337A
Authority
CN
China
Prior art keywords
coa
transformer
fault diagnosis
cnn
feature
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310697655.8A
Other languages
Chinese (zh)
Inventor
白兴东
张学亮
牛欢欢
蒋迪楠
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Super High Voltage Co Of State Grid Gansu Electric Power Co
State Grid Gansu Electric Power Co Ltd
Original Assignee
Super High Voltage Co Of State Grid Gansu Electric Power Co
State Grid Gansu Electric Power Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Super High Voltage Co Of State Grid Gansu Electric Power Co, State Grid Gansu Electric Power Co Ltd filed Critical Super High Voltage Co Of State Grid Gansu Electric Power Co
Priority to CN202310697655.8A priority Critical patent/CN116842337A/en
Publication of CN116842337A publication Critical patent/CN116842337A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
    • G01R31/12—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing
    • G01R31/1227—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing of components, parts or materials
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention relates to the technical field of transformer fault diagnosis, in particular to a transformer fault diagnosis method based on a preferred feature of a LightGBM and a COA-CNN model, which comprises the following steps: step one, data processing of transformer fault diagnosis; firstly, constructing an initial data set by collected fault sample data of a transformer and corresponding fault type labels, and then carrying out data processing and characteristic variable optimization on the initial data set; then, performing iterative optimization and updating on the CNN super parameters through a COA algorithm; constructing a model for transformer fault diagnosis and performing fault diagnosis on the transformer; and inputting the acquired data of the dissolved gas in the transformer oil into a trained COA-CNN model, and diagnosing the current state and fault type of the transformer. The model has certain reference significance for timely mastering the state of the transformer by maintenance personnel.

Description

Transformer fault diagnosis method based on LightGBM (gallium nitride based) optimal characteristics and COA-CNN (chip on board) model
Technical Field
The invention relates to the technical field of transformer fault diagnosis, in particular to a transformer fault diagnosis method based on a preferred feature of a LightGBM and a COA-CNN model.
Background
The power transformer is a junction in the links of power generation, transmission and distribution in a power system, is one of the most expensive devices, and the reliability of the power transformer is the basis of stable operation of a power grid. The fault of the transformer is rapidly and accurately diagnosed, the reliability of electric energy transmission and the operation safety of the whole power grid are directly related, and the huge economic loss caused by accident expansion can be avoided, so that an effective transformer fault diagnosis model is established, the fault diagnosis precision is improved, and the method has important significance for guaranteeing the safe, reliable and high-quality operation of the power grid.
At present, analysis of dissolved gas in oil is an effective method for diagnosing the state and fault type of an oil immersed transformer, and is mainly divided into a traditional method and a machine learning method. The traditional methods comprise an IEC ratio method, a Dornenburg judging method, a Rogers ratio method, a code-free ratio method and the like, the principles of the methods are simple, the number of characteristic variables is enriched, and the dimension among gas data is unified, but the fault diagnosis accuracy is lower due to incomplete coding, over absolute characteristic variables and too few characteristic variables. With rapid development of intelligent algorithms and machine learning, methods such as neural networks, support Vector Machines (SVMs), extreme Learning Machines (ELMs), random forests, expert systems, decision trees and the like are widely applied to transformer fault diagnosis in recent years, and good effects are obtained. However, the SVM essentially belongs to two classifiers, and the classification accuracy depends on the selection of penalty coefficients and kernel functions; ELM has poor robustness and cannot meet the diagnostic requirements for long-term stability; the training space and time of the random forest depend on the number of decision trees, and fitting is easy when the noise of the data set is large; expert systems require a large number of knowledge rules and have poor learning ability.
CNN is a novel classification model applied to transformer fault diagnosis, solves the defect of insufficient expression capacity of shallow networks such as BP neural network (BPNN), and is theoretically suitable for diagnosing transformer faults according to the content of dissolved gas in oil, compared with LSTM and other networks, the CNN has no requirement on the time sequence and continuity of data. Because the dissolved gas in the transformer oil is used as a characteristic variable less, the maximum information coefficient method and the information gain method are adopted in the prior art to conduct optimization after the quantity of the characteristic variable is enriched, and the fault diagnosis accuracy of the transformer is effectively improved, so that the feasibility is realized by enriching the characteristic variable and conducting optimization; meanwhile, as the CNN structure and parameters are improved, higher accuracy can be obtained, the prior art increases the attention mechanism on the basis of the convolution layer to distribute the weights, and the fault diagnosis level is improved; in the prior art, the ratio method is fused with the improved CNN to improve the fault diagnosis accuracy; the CNN model is optimized from the structural point of view, and from the parameter point of view, the combination of excellent intelligent algorithm and CNN still needs to be researched.
Disclosure of Invention
The invention aims to provide a transformer fault diagnosis method based on the preferred characteristics of the LightGBM and the COA-CNN model, aiming at solving the problems existing in the prior art, and aiming at diagnosing the running state and the fault type of a transformer as accurately as possible.
In order to achieve the above object, the present invention provides the following technical solutions:
the transformer fault diagnosis method based on the preferred characteristics of the LightGBM and the COA-CNN model is characterized by comprising the following steps of:
step one, data processing of transformer fault diagnosis; firstly, constructing an initial data set by collected fault sample data of a transformer and corresponding fault type labels, and then carrying out data processing and characteristic variable optimization on the initial data set; then, performing iterative optimization and updating on the CNN super parameters through a COA algorithm;
constructing a model for transformer fault diagnosis and performing fault diagnosis on the transformer; and inputting the acquired data of the dissolved gas in the transformer oil into a trained COA-CNN model, and diagnosing the current state and fault type of the transformer.
Preferably, in the first step, in the transformer fault diagnosis data processing, since the operation state of the transformer can be classified into normal, low temperature overheat, medium temperature overheat, high temperature overheat, low energy discharge, high energy discharge, etc., since the insulating oil in the transformer mainly consists of hydrocarbons, the difference in the chemical reaction speed occurs when different fault types occur, leading to H 2 、CH 4 、C 2 H 2 、C 2 H 4 、C 2 H 6 The components, the content and the ratio of the dissolved gas can be changed differently, and analyzing the dissolved gas in the transformer oil is an important means for diagnosing the fault type of the transformer; in view of the possible problems of sample data missing, redundancy, too few or too many feature variables, etc., it is necessary to process sample data with preferred feature variables.
Specifically, the processing the sample data includes: because the number is less when 5 gas components are used as characteristic variables, the relevance between the gases and faults cannot be accurately reflected, CNN (carbon nanotubes) underfitting is easy to cause, and the number of the characteristic variables can be enriched when methods such as an IEC (International electrotechnical Commission) ratio method and the like make up the difference between the orders of the gases, so that the ideas of the IEC ratio method and the non-coding ratio method are used for reference, and the ratio of different gases is provided as a supplementary characteristic variable on the basis of the 5 gases as the characteristic variables. As shown in Table 1, n represents the gas content, C 1 Represents CH 4 First-order hydrocarbon, C 2 Represents C 2 H 2 、C 2 H 4 、C 2 H 6 And second order hydrocarbons. Because the ratio idea is adopted, NAN values possibly appear in the result when the content of certain gas is 0, and the situation can influence the training of the model, the invention fills the data by adopting 0.
Table 1 supplement characteristic variables
Because the dimension of the characteristic variables with different dimensions is different, if the normalization processing is not performed, the algorithm result and the performance are adversely affected, and therefore, the invention adopts a common min-max normalization processing method to realize the mapping of the characteristic variable values in the [0,1] interval.
Specifically, the preferred feature variables include: after the feature variables are enriched, the too long training time and the time consumption of fault diagnosis can be increased when the number of the feature variables is too large, the problems can be generated when the number of the feature variables is too large, and even the fault diagnosis accuracy is reduced due to fitting caused by redundant information; therefore, in order to solve the defect of subjectively selecting the feature variables and extract important features to reduce training time, the invention adopts the LightGBM algorithm to quantify the importance of the feature variables and screens a dataset formed by the feature variables ranked at the front as an input set.
Specifically, the LightGBM is a fast, high performance framework for implementing gradient-lifted decision trees (Gradient Boosting Decision Tree, GBDT), suitable for datasets with higher feature dimensions. The algorithm constructs a series of decision trees into a strong learner by a weighted linear combination method. The training method is characterized in that samples with larger gradients are reserved during training, and samples with smaller gradients are randomly sampled and constant coefficients are introduced. For feature set I: { X 1 ,X 2 ,K X n Negative gradient of each round of iterative loss function is used { g } 1 ,g 2 ,K,g n The sampling rates of the large gradient and the small gradient samples are respectively represented by a and b, and the characteristic variable importance evaluation and calculation steps are as follows:
(1) Sorting the characteristic variables in descending order according to the absolute value of the gradient;
(2) Extracting a×n feature variables with the top order to form a large gradient feature set I 1 ；
(3) B×n are then extracted from the remaining feature variables to form a small gradient feature set I 2 ；
(4) Feature set I using sampling 1 UI 2 When learning a new decision tree and calculating the information gain of a certain node, a small sample gradient weight (1-a)/b is given, and then the information gain of a division point d of a division feature j is：
(5) Repeating the steps (1) - (4) until the preset iteration times are reached, traversing the whole model, and obtaining the importance evaluation result of the feature, namely the sum of all node information gains of the feature serving as the splitting feature.
Preferably, in the second step, in the model building of the transformer fault diagnosis and the fault diagnosis of the transformer, because the CNN network needs to set super parameters before training, the method mainly comprises a learning rate, the size and the number of convolution kernels of each layer and the number of neurons of the full-connection layer, and the proper super parameters have important influence on the accuracy and the convergence rate of the CNN, the COA is a newly proposed intelligent optimization algorithm based on a population, and compared with classical intelligent optimization algorithms such as a Genetic Algorithm (GA) and a Particle Swarm Optimization (PSO), the COA has strong optimizing capability and rapid convergence capability, and has no basic parameter setting. The present invention therefore proposes a combination of COA and CNN, i.e. a COA-CNN model.
Specifically, the specific steps of the COA (raccoon optimization algorithm) are as follows:
(1) Initializing a population, and determining the population number N, the maximum iteration number T, the independent variable number m, the independent variable upper limit ub, the independent variable lower limit lb and an objective function f. The location information for each raccoon represents a set of super-parametric settings for the CNN, using equation (6) to randomly generate the location of each raccoon in the population:
X i ＝lb+r·(ub-lb)，i＝1,2,3K,N (6)
wherein X is i Positional information representing the i-th raccoon; r represents [0,1]]Is a random number of (a) in the memory.
After initializing the population, calculating fitness value f (X) i ) Storing a globally optimal solutionAnd (3) withThen searching is started in a search space, and a search mechanism comprises two stages of exploration and development, wherein the global search capability and the local optimizing capability of the COA are respectively embodied;
(2) The exploration phase is based on the behavior of raccoon prey on prey. At this stage, the population is equally divided into two categories, one climbing up the tree and the other waiting for the prey on the ground. For individuals on a tree, the mathematical expression is:
in the method, in the process of the invention,positional information representing the i-th raccoon at the t+1st iteration; r represents [0,1]]Random numbers of (a); i represents [1,2]]Random integers of (a); t represents the current iteration number;
for individuals waiting on a game on the ground, the mathematical expression is:
Iguana t ＝lb+r·(ub-lb) (9)
wherein r represents a random number of [0,1 ]; i represents a random integer of [1,2 ].
Greedy selection of the population is performed as shown in equation (10):
(3) And (3) a development stage. The development phase is based on the natural behavior of raccoons to escape predators. At this stage, each raccoon will be predated, i.e., i=1, 2,3, k, n, expressed mathematically as
Wherein r represents a random number of [0,1 ]; t represents the current iteration number.
And (5) carrying out greedy selection on the population by the formula (10), and obtaining the global optimal solution. The flow chart is shown in fig. 2.
Specifically, the CNN (convolutional neural network) is:
the one-dimensional CNN is generally used for processing a one-dimensional array, the convolution kernel slides along one direction, accords with the input characteristics of the dissolved gas of the transformer, and mainly comprises an input layer, a convolution layer, a pooling layer, a full connection layer and an output layer, and the general mathematical model expression is as follows:
x output ＝f Softmax (f fc (f pooling (f conv (x input )))) (12)
wherein x is input Representing an input feature set; x is x output Representing the output classification result; f (f) conv (-) represents convolutional layer calculations, including convolutional operations and nonlinear activations; f (f) pooling (. Cndot.) represents pooling layer computation; f (f) fc (. Cndot.) represents a full connection layer calculation; f (f) Softmax Representing the output result calculated by the Softmax function.
Specifically, the COA-CNN model is as follows:
the COA solving CNN network structure superparameter process is shown in figure 3. In COA-CNN, the super parameter optimization of the CNN is realized by utilizing the exploration-development mechanism of the COA, the core goal of the COA is to find the maximum/minimum value of the fitness function, and the invention adopts the fault diagnosis accuracy of the transformer as the fitness function of the COA to calculate the maximum value of the fitness function;
the fault diagnosis flow chart of the COA-CNN model is shown in figure 4. The algorithm mainly comprises 3 big modules, namely: the system comprises a data processing module, a COA module and a CNN module. The main steps of transformer fault diagnosis based on the COA-CNN model are as follows:
(1) Data processing; enriching characteristic variables by using the thought of a ratio method, and preprocessing sample data by a data filling and min-max method; then, carrying out importance evaluation on the characteristic variables by adopting a LightGBM algorithm, optimizing the characteristic variables to form an input set, and dividing a training set and a testing set according to a certain proportion;
(2) COA optimizes CNN super parameters; initializing population quantity and iteration times of a COA algorithm, taking the iteration times, learning rate, hidden layer node number and other super parameters of a CNN neural network as independent variables of the COA, taking the fault diagnosis accuracy of a transformer as an adaptability function of the COA, calculating an adaptability value and updating the population until the attribute value corresponding to an individual with the largest adaptability function value when the maximum iteration times are met is the optimal super parameter of the CNN model;
(3) Performing COA-CNN fault diagnosis; based on an input set obtained by data processing, and combining with super parameters optimized by a COA algorithm, constructing a CNN model to perform fault diagnosis on the transformer.
The invention has the following beneficial effects: the invention provides a transformer fault diagnosis method based on a preferred feature of a LightGBM algorithm and a COA-CNN model, which comprises the steps of firstly processing sample data and enriching feature variables, and then carrying out feature preference by using the LightGBM algorithm; then optimizing the learning rate, the convolution kernel size and quantity and the quantity of neurons of the full-connection layer in the CNN by using a COA algorithm, so as to avoid subjectivity of CNN super-parameter setting; finally, the performance of the model provided by the invention is verified through experiments. The model has certain reference significance for timely mastering the state of the transformer by maintenance personnel.
Drawings
FIG. 1 is a general frame diagram for diagnosing faults of a transformer in embodiment 1 of the present invention;
FIG. 2 is a flowchart of the raccoon optimization algorithm of example 1 of the present invention;
FIG. 3 is a diagram showing the process of COA solving CNN network structure superparameter in embodiment 1 of the present invention;
FIG. 4 is a flow chart of fault diagnosis of the COA-CNN model in embodiment 1 of the invention;
FIG. 5 is a graph showing the evaluation result of importance of feature variables in embodiment 1 of the present invention;
FIG. 6 is a diagram of the fault diagnosis result of the COA-CNN model transformer in the embodiment 1 of the invention;
FIG. 7 is a confusion matrix diagram of fault diagnosis results of the COA-CNN model transformer in the embodiment 1 of the invention;
FIG. 8 is a chart showing the convergence of different models in example 1 of the present invention;
fig. 9 is a comparison chart of transformer fault diagnosis accuracy of different classification models in embodiment 1 of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
Example 1 example analysis
1. Data processing and model setting
The transformer fault diagnosis model provided by the invention is used for analysis by document data and 839 groups of data records of transformer monitoring and historical faults in a certain region in northwest, and the sample distribution situation and fault codes are shown in table 2. After sample data processing and feature variable optimization are performed, the importance of all feature variables is quantified through a LightGBM algorithm, and as shown in FIG. 5, the larger the importance evaluation result score of the feature variables is, the more important the importance is. Taking a data set formed by the feature variables with higher importance as the input of the model, and taking the input set as 7:3 is randomly divided into a training set and a testing set.
Table 2 sample data distribution and coding
As can be seen from an analysis of fig. 5, the first few feature variables of higher importance include: CH (CH) 4 /H 2 (feature 6), C 2 H 2 /C 2 H 4 (feature 11), C 2 H 4 /C 2 H 6 And (feature 14) coinciding with the three-ratio feature variable of the IEC ratio method, the preferred feature variable of the LightGBM algorithm has larger consistency with the traditional theory, and the preferred result of the algorithm is more comprehensive.
According to the invention, a basic CNN model is built according to manual experience: the convolution layers adopt 2 layers, the convolution kernel of each layer is respectively 3 multiplied by 3 and 2 multiplied by 2, and the number is respectively 32 and 64; setting a pooling layer and a batch standardization layer behind each convolution layer, wherein the size of a pooling core is 2 multiplied by 2, the step length is 1, and the filling mode is Same; the activation function adopts a ReLU function; the total connection layer adopts 2 layers, and the number of neurons is 64 and 32 respectively; activated by a Softmax function, the output is 6; the optimizer is Adam; the learning rate was 0.001. And for the construction of the COA-CNN model, the convolution layer, the full connection layer, the optimizer and the like are the same as those of the basic CNN model, but the COA optimization is adopted by the super parameters such as the size and the number of convolution kernels, the number of neurons of the full connection layer, the initial learning rate and the like.
In order to comprehensively and effectively evaluate the performance of the model, the invention adopts the accuracy eta accuracy Accuracy eta precision Recall ratio eta recall As evaluation indexes, the misjudgment, misjudgment and missed judgment capability of the model on the transformer faults are respectively reflected. The higher the accuracy rate is, the stronger the fault diagnosis capability of the transformer is, the higher the accuracy rate is, the higher the reliability of the fault diagnosis is, and the higher the recall rate is, the higher the identification sensitivity of the fault diagnosis is. The mathematical expression is:
wherein n represents predictionThe classification result of (2) is the same as the actual sample number; n represents the total number of samples; n is n T A sample prediction correct number representing a certain fault type; n is n P Representing the total number of samples predicted to be of a certain fault type; n is n R Representing the total number of samples for a certain fault type.
2. Model performance analysis
(1) Model validity analysis
In order to evaluate the effectiveness of the proposed model, the invention adopts an optimal network obtained by COA-CNN to carry out fault diagnosis on a test set, takes the accuracy, the precision and the recall as evaluation indexes, the fault diagnosis result is shown in figure 6, the confusion matrix is shown in figure 7, and the evaluation index result is shown in table 3.
TABLE 3 evaluation index of COA-CNN model
As can be seen from analysis of fig. 7 and table 3, the fault diagnosis accuracy reaches 96.4%, which indicates that the fault type diagnosis error judgment condition of the transformer by the model is less and the overall capability is stronger; the accuracy of the medium-temperature overheat type is lowest and is 91.2%, the highest accuracy of the other types can reach 98.1%, the overall misjudgment condition of the model in the diagnosis of the fault type of the transformer is less, and the reliability of the diagnosis is higher; the recall ratio of the low-temperature overheat type is 73.1 percent, the other types can reach 100 percent, and the method shows that the overall missed judgment condition of the model in the diagnosis of most fault types of the transformer is less, and the sensitivity of overall identification is higher.
(2) Model generalization analysis
Since the generalization of a network depends mainly on 3 factors: the model architecture of the invention has been determined, and the complexity of the problem will not change, so in order to evaluate the generalization of the proposed model, the data sets are respectively according to 6: 4. 8: the ratio of 2 is divided into a training set and a testing set, and the accuracy of the training set and the testing set is compared with that of a basic CNN model, and the result is shown in table 4.
Table 4 comparison of transformer fault diagnosis accuracy for different ratio datasets
Analysis table 4 shows that the accuracy of the model provided by the invention is at a higher level under the condition that the training set and the testing set are set to be in different proportions, so that the good performance of the model is further illustrated. Whether the basic CNN model or the model provided by the invention is adopted, when the training set is in a lower proportion, the sample data of the training network are relatively less, the neural network cannot be fully learned, so the accuracy is relatively lower, and as the proportion of the training set is increased, the accuracy of the basic CNN model is respectively improved by 2.4 percent and 3.6 percent, and the accuracy of the COA-CNN model is respectively improved by 3.9 percent and 4.5 percent, and the accuracy of the basic CNN model and the accuracy of the COA-CNN model are respectively improved to different degrees. The model provided by the invention has good generalization.
(3) Feasibility analysis of preferred feature variables in a model
In order to evaluate the necessity of the preferred feature variables of the LightGBM algorithm in the model, a data set composed of basic feature variables (5 types), all feature variables (30 types) and preferred feature variables (9 types) is selected as an input set of an optimal network obtained by the COA-CNN model, and the fault diagnosis accuracy of the test set is used as an evaluation index, and the result is shown in table 5.
Table 5 transformer fault diagnosis accuracy for different characteristic variables
As shown in the analysis table 5, when the basic characteristic variable dataset is used as the input set, the model is difficult to extract the key information for representing the fault type, the fault diagnosis accuracy of the training set and the test set is low, and the accuracy of the test set is only 84.5%; the model accuracy reaches 96.4% when the optimized characteristic variable data set obtained by the LightGBM algorithm is used as an input set, and the time is 0.739s. The time consumption is increased by 5.8 percent compared with the time consumption of adopting basic characteristic variables, but the accuracy is improved by 11.9 percent; compared with the method adopting all characteristic variables, the method has no change in accuracy, but not only greatly reduces the time for offline training of the model, but also reduces the time consumption for fault diagnosis by 18.1%. It is meaningful to illustrate the preferred feature variables, with feasibility, and the LightGBM algorithm has validity.
(4) Model convergence analysis
In order to evaluate the convergence of the COA in the model, a data set formed by the preferable characteristic variables is used as an input set, the fitness function is used as an evaluation index, classical intelligent optimization algorithms (GA and PSO) are respectively adopted to optimize CNN super parameters, an optimal network is built, the accuracy and the convergence speed of different models are compared, and the results are shown in fig. 8 and table 6.
Table 6 results of transformer fault diagnosis for different models
As can be seen from analysis of FIG. 8 and Table 6, for the improved CNN models such as GA-CNN, PSO-CNN, COA-CNN, etc., the fault diagnosis accuracy of the transformer is improved to different degrees compared with the basic CNN model due to the optimized super-parameters; the algorithm performance of GA is related to the crossover and mutation probability, and is easy to fall into local optimum; PSO performance is related to weight and learning factors, so that convergence times of the PSO performance and the learning factors are relatively more, the PSO performance and the learning factors tend to be stable in 23 rd iteration and 20 th iteration respectively, and the final accuracy reaches 96.0% and 96.4% respectively; the COA is not required to be set with basic parameters, and the global optimizing capability and the convergence capability of the algorithm are improved, so that the model tends to be stable in 7 iterations, and the model has better convergence.
(5) Model accuracy analysis
In order to evaluate the accuracy of the model provided by the invention, a data set formed by the preferable characteristic variables is used as an input set, the accuracy is used as an evaluation index, and the result is compared with a classical classification model (BPNN, decision tree, random forest and basic CNN), and the results are shown in FIG. 9 and Table 7.
Table 7 transformer fault diagnosis accuracy of different classification models
Analysis of fig. 9 and table 7 shows that, compared with the models such as BPNN, decision tree, random forest, etc., the accuracy of the basic CNN model is improved by 9.4%, 1.9% and 1.1%, respectively, which indicates that the CNN model has strong nonlinear relation mapping capability in transformer fault diagnosis, the fault diagnosis is accurate, and the accuracy of COA-CNN is improved by 3.6% on the basic CNN model, which indicates that the model provided by the invention has higher transformer fault diagnosis accuracy.
Conclusion(s)
Aiming at the diagnosis of the state and fault type of the transformer, the invention provides a COA-CNN model based on the optimized characteristic variable of the LightGBM algorithm, and the performance of the model is verified through an example, so that the conclusion is as follows:
(1) Aiming at the influence of the characteristic variable on the transformer fault diagnosis in the transformer fault diagnosis, the invention uses the lightGBM algorithm to optimize the characteristic variable after enriching the characteristic variable by referring to the thought of the ratio method, so that the relevance between the characteristic variable and the fault type is fully excavated, the fault diagnosis precision is almost unchanged, the fault diagnosis time is reduced by 18.1%, and the feasibility of the optimized characteristic variable and the effectiveness of the lightGBM algorithm are shown.
(2) Aiming at the problem that the CNN hyper-parameters are difficult to construct by artificial experience selection, the invention adopts the COA algorithm to optimize the learning rate, the convolution kernel size and number and the number of neurons of the full-connection layer in the CNN, and experimental results show that the COA-CNN model has better convergence, stronger global searching capability and higher fault diagnosis accuracy.
(3) The accuracy rate of fault diagnosis of the COA-CNN model provided by the invention on the transformer reaches 96.4%, and the model has higher accuracy rate compared with the basic CNN under different proportion training sets; compared with GA-CNN and PSO-CNN, the convergence times are respectively reduced by 16 times and 13 times; compared with classical classification models such as BPNN, decision trees, random forests and the like, the accuracy is improved by 13.0%, 5.5% and 4.7%, respectively, and the model has the advantages of being high in generalization, good in convergence and high in precision, and has a certain reference significance for operation and maintenance repair staff to identify the current state of the transformer.
The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, and alternatives falling within the spirit and principles of the invention.

Claims (9)

1. The transformer fault diagnosis method based on the preferred characteristics of the LightGBM and the COA-CNN model is characterized by comprising the following steps of:
step one, data processing of transformer fault diagnosis; firstly, constructing an initial data set by collected fault sample data of a transformer and corresponding fault type labels, and then carrying out data processing and characteristic variable optimization on the initial data set; then, performing iterative optimization and updating on the CNN super parameters through a COA algorithm;
constructing a model for transformer fault diagnosis and performing fault diagnosis on the transformer; and inputting the acquired data of the dissolved gas in the transformer oil into a trained COA-CNN model, and diagnosing the current state and fault type of the transformer.
2. The transformer fault diagnosis method based on the preferred characteristics of LightGBM and the COA-CNN model as claimed in claim 1, wherein in the first step, in the transformer fault diagnosis data processing, since the operation state of the transformer can be classified into normal, low temperature overheat, medium temperature overheat, high temperature overheat, low energy discharge, high energy discharge, etc., since the insulating oil in the transformer is mainly composed of hydrocarbons, different results occurThe chemical reaction speed is different when the barrier type is adopted, resulting in H 2 、CH 4 、C 2 H 2 、C 2 H 4 、C 2 H 6 The components, the content and the ratio of the dissolved gas can be changed differently, and analyzing the dissolved gas in the transformer oil is an important means for diagnosing the fault type of the transformer; in view of the possible problems of sample data missing, redundancy, too few or too many feature variables, etc., it is necessary to process sample data with preferred feature variables.
3. The transformer fault diagnosis method based on the LightGBM preferred feature and COA-CNN model of claim 2, wherein the processing the sample data comprises: because the number is less when 5 gas components are used as characteristic variables, the relevance between the gases and faults cannot be accurately reflected, CNN (carbon nanotubes) underfitting is easy to cause, and the number of the characteristic variables can be enriched when methods such as an IEC (International electrotechnical Commission) ratio method and the like make up the difference between the orders of the gases, so that the ideas of the IEC ratio method and the non-coding ratio method are used for reference, and the ratio of different gases is provided as a supplementary characteristic variable on the basis of the 5 gases as the characteristic variables.
4. The transformer fault diagnosis method based on the LightGBM preferred feature and COA-CNN model of claim 2, wherein the preferred feature variables comprise: after the feature variables are enriched, the too long training time and the time consumption of fault diagnosis can be increased when the number of the feature variables is too large, the problems can be generated when the number of the feature variables is too large, and even the fault diagnosis accuracy is reduced due to fitting caused by redundant information; therefore, in order to solve the defect of subjectively selecting the feature variables and extract important features to reduce training time, the invention adopts the LightGBM algorithm to quantify the importance of the feature variables and screens a dataset formed by the feature variables ranked at the front as an input set.
5. The transformer fault diagnosis method based on the preferred characteristics of the LightGBM and the COA-CNN model as claimed in claim 3, wherein the LightGBM is realized by gradient extractionA fast, high performance framework of ascending decision trees (Gradient Boosting Decision Tree, GBDT) applicable to datasets with higher feature dimensions; the algorithm constructs a series of decision trees into a strong learner by a weighted linear combination method; the method comprises the steps of retaining samples with larger gradients during training, randomly sampling samples with smaller gradients, and introducing constant coefficients; for feature set I: { X 1 ,X 2 ,K X n Negative gradient of each round of iterative loss function is used { g } 1 ,g 2 ,K,g n The sampling rates of the large gradient and the small gradient samples are respectively represented by a and b, and the characteristic variable importance evaluation and calculation steps are as follows:
(1) Sorting the characteristic variables in descending order according to the absolute value of the gradient;
(2) Extracting a×n feature variables with the top order to form a large gradient feature set I 1 ；
(3) B×n are then extracted from the remaining feature variables to form a small gradient feature set I 2 ；
(4) Feature set I using sampling 1 UI 2 When a new decision tree is learned and the information gain of a certain node is calculated, a small sample gradient weight (1-a)/b is given, and then the information gain of a division point d of a division feature j is as follows:
(5) Repeating the steps (1) - (4) until the preset iteration times are reached, traversing the whole model, and obtaining the importance evaluation result of the feature, namely the sum of all node information gains of the feature serving as the splitting feature.
6. The transformer fault diagnosis method based on the LightGBM optimized feature and the COA-CNN model according to claim 1, wherein in the second step, in the model construction of the transformer fault diagnosis and the fault diagnosis of the transformer, since the CNN network needs to be set with super parameters before training, mainly including learning rate, convolution kernel size and number of each layer, and number of neurons of the full-connection layer, the appropriate super parameters have important influence on accuracy and convergence speed of the CNN, and COA is a newly proposed intelligent optimization algorithm based on population, compared with classical intelligent optimization algorithms such as Genetic Algorithm (GA) and particle swarm algorithm (PSO), COA has strong optimizing ability and rapid convergence ability, and no basic parameter setting, so the invention proposes COA and CNN combination, namely COA-CNN model.
7. The transformer fault diagnosis method based on the LightGBM preferred feature and COA-CNN model according to claim 6, characterized in that the specific steps of COA (raccoon optimization algorithm) are as follows:
(1) Initializing a population, and determining the population number N, the maximum iteration number T, the independent variable number m, the independent variable upper limit ub, the independent variable lower limit lb and an objective function f; the location information for each raccoon represents a set of super-parametric settings for the CNN, using equation (6) to randomly generate the location of each raccoon in the population:
X i ＝lb+r·(ub-lb)，i＝1,2,3K,N (6)
wherein X is i Positional information representing the i-th raccoon; r represents [0,1]]Random numbers of (a);
after initializing the population, calculating fitness value f (X) i ) Storing a globally optimal solutionAnd->Then searching is started in a search space, and a search mechanism comprises two stages of exploration and development, wherein the global search capability and the local optimizing capability of the COA are respectively embodied;
(2) In the exploration phase, a mathematical model is established according to the behavior of raccoon prey on the prey, and in the exploration phase, the population is evenly divided into two types, one part climbs the tree, the other part waits for the prey on the ground, and for individuals on the tree, the mathematical expression is as follows:
in the method, in the process of the invention,positional information representing the i-th raccoon at the t+1st iteration; r represents [0,1]]Random numbers of (a); i represents [1,2]]Random integers of (a); t represents the current iteration number;
for individuals waiting on a game on the ground, the mathematical expression is:
Iguana t ＝lb+r·(ub-lb) (9)
wherein r represents a random number of [0,1 ]; i represents a random integer of [1,2 ];
greedy selection of the population is performed as shown in equation (10):
(3) A development stage in which the raccoons are each prey according to their natural behavior, i.e., i=1, 2,3, k, n, the mathematical expression being
Wherein r represents a random number of [0,1 ]; t represents the current iteration number;
and (5) carrying out greedy selection on the population by the formula (10), and obtaining the global optimal solution.
8. The transformer fault diagnosis method based on the LightGBM preferred feature and COA-CNN model of claim 7, wherein the CNN (convolutional neural network) is: the one-dimensional CNN is generally used for processing a one-dimensional array, the convolution kernel slides along one direction, accords with the input characteristics of the dissolved gas of the transformer, and mainly comprises an input layer, a convolution layer, a pooling layer, a full connection layer and an output layer, and the general mathematical model expression is as follows:
x output ＝f Softmax (f fc (f pooling (f conv (x input )))) (12)
wherein x is input Representing an input feature set; x is x output Representing the output classification result; f (f) conv (-) represents convolutional layer calculations, including convolutional operations and nonlinear activations; f (f) pooling (. Cndot.) represents pooling layer computation; f (f) fc (. Cndot.) represents a full connection layer calculation; f (f) Softmax Representing the output result calculated by the Softmax function.
9. The transformer fault diagnosis method based on the LightGBM optimized feature and the COA-CNN model according to claim 8, wherein in COA-CNN, the super-parameter optimization of the COA is realized by using the exploration-development mechanism of COA, and the core objective of COA is to find the maximum/minimum value of the fitness function, and the invention adopts the transformer fault diagnosis accuracy as the fitness function of COA to find the maximum value thereof; the algorithm mainly comprises 3 big modules, namely: the system comprises a data processing module, a COA module and a CNN module; the main steps of transformer fault diagnosis based on the COA-CNN model are as follows:
(1) Data processing; enriching characteristic variables by using the thought of a ratio method, and preprocessing sample data by a data filling and min-max method; then, carrying out importance evaluation on the characteristic variables by adopting a LightGBM algorithm, optimizing the characteristic variables to form an input set, and dividing a training set and a testing set according to a certain proportion;
(2) COA optimizes CNN super parameters; initializing population quantity and iteration times of a COA algorithm, taking the iteration times, learning rate, hidden layer node number and other super parameters of a CNN neural network as independent variables of the COA, taking the fault diagnosis accuracy of a transformer as an adaptability function of the COA, calculating an adaptability value and updating the population until the attribute value corresponding to an individual with the largest adaptability function value when the maximum iteration times are met is the optimal super parameter of the CNN model;
(3) Performing COA-CNN fault diagnosis; based on an input set obtained by data processing, and combining with super parameters optimized by a COA algorithm, constructing a CNN model to perform fault diagnosis on the transformer.
CN202310697655.8A 2023-06-13 2023-06-13 Transformer fault diagnosis method based on LightGBM (gallium nitride based) optimal characteristics and COA-CNN (chip on board) model Pending CN116842337A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310697655.8A CN116842337A (en) 2023-06-13 2023-06-13 Transformer fault diagnosis method based on LightGBM (gallium nitride based) optimal characteristics and COA-CNN (chip on board) model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310697655.8A CN116842337A (en) 2023-06-13 2023-06-13 Transformer fault diagnosis method based on LightGBM (gallium nitride based) optimal characteristics and COA-CNN (chip on board) model

Publications (1)

Publication Number Publication Date
CN116842337A true CN116842337A (en) 2023-10-03

Family

ID=88162638

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310697655.8A Pending CN116842337A (en) 2023-06-13 2023-06-13 Transformer fault diagnosis method based on LightGBM (gallium nitride based) optimal characteristics and COA-CNN (chip on board) model

Country Status (1)

Country Link
CN (1) CN116842337A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117150374A (en) * 2023-10-27 2023-12-01 国网江西省电力有限公司电力科学研究院 Transformer fault identification method and system based on feature importance sorting
CN117261599A (en) * 2023-10-18 2023-12-22 北京航空航天大学 Fault detection method and device of electric automobile, electronic equipment and electric automobile
CN117911799A (en) * 2024-03-19 2024-04-19 贵州师范大学 Feature classification method for improving shrimp algorithm based on multiple strategies

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117261599A (en) * 2023-10-18 2023-12-22 北京航空航天大学 Fault detection method and device of electric automobile, electronic equipment and electric automobile
CN117261599B (en) * 2023-10-18 2024-05-03 北京航空航天大学 Fault detection method and device of electric automobile, electronic equipment and electric automobile
CN117150374A (en) * 2023-10-27 2023-12-01 国网江西省电力有限公司电力科学研究院 Transformer fault identification method and system based on feature importance sorting
CN117911799A (en) * 2024-03-19 2024-04-19 贵州师范大学 Feature classification method for improving shrimp algorithm based on multiple strategies
CN117911799B (en) * 2024-03-19 2024-05-17 贵州师范大学 Feature classification method for improving shrimp algorithm based on multiple strategies

Similar Documents

Publication Publication Date Title
Liao et al. Fault diagnosis of power transformers using graph convolutional network
CN116842337A (en) Transformer fault diagnosis method based on LightGBM (gallium nitride based) optimal characteristics and COA-CNN (chip on board) model
CN110929847A (en) Converter transformer fault diagnosis method based on deep convolutional neural network
CN112070128B (en) Transformer fault diagnosis method based on deep learning
CN110542819B (en) Transformer fault type diagnosis method based on semi-supervised DBNC
CN111237134A (en) Offshore double-fed wind driven generator fault diagnosis method based on GRA-LSTM-stacking model
CN110689171A (en) Turbine health state prediction method based on E-LSTM
CN110879377B (en) Metering device fault tracing method based on deep belief network
CN115563563A (en) Fault diagnosis method and device based on transformer oil chromatographic analysis
CN116562114A (en) Power transformer fault diagnosis method based on graph convolution neural network
CN115409369A (en) Comprehensive energy system reliability evaluation method based on mechanism and data hybrid driving
CN114358116A (en) Oil-immersed transformer fault diagnosis method and system and readable storage medium
CN112559741B (en) Nuclear power equipment defect record text classification method, system, medium and electronic equipment
CN113762591A (en) Short-term electric quantity prediction method and system based on GRU and multi-core SVM counterstudy
CN117520809A (en) Transformer fault diagnosis method based on EEMD-KPCA-CNN-BiLSTM
CN116488151A (en) Short-term wind power prediction method based on condition generation countermeasure network
CN114997475A (en) Short-term prediction method for photovoltaic power generation of fusion model based on Kmeans
Jianli et al. Wind power forecasting by using artificial neural networks and Grubbs criterion
CN110837932A (en) Thermal power prediction method of solar heat collection system based on DBN-GA model
CN117633456B (en) Marine wind power weather event identification method and device based on self-adaptive focus loss
CN115587644B (en) Photovoltaic power station performance parameter prediction method, device, equipment and medium
CN116451049B (en) Wind power prediction method based on agent assisted evolutionary neural network structure search
CN111144431B (en) Transformer fault diagnosis method based on CBBO-SVM
CN117081064B (en) Wind power combination prediction method and system based on error evaluation
Li Construction of Transformer Fault Diagnosis and Prediction Model Based on Deep Learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination