A method, system and device for constructing a GIS intermittent discharge mode identification model

Abstract

The embodiment of the application discloses a GIS intermittent discharge mode recognition model construction method, system and equipment, the method comprises the following steps: collecting different types of partial discharge information in the GIS intermittent discharge process; extracting a training set and a test set according to different types of partial discharge information, training an initial BP neural network using the training set; obtaining partial discharge characteristics V X and V Y from the set m different feature subsets to find n optimal feature sets that meet the MRMR criteria, and form a partial discharge fault state evaluation feature set; using the partial discharge fault state evaluation feature set to test the trained initial BP neural network to obtain a GIS intermittent discharge mode recognition model. The application combines the MRMR algorithm and the BP neural network algorithm, and the overall recognition rate is better than that of a single BP neural network, and the intermittent discharge fault can be accurately recognized.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power equipment fault type identification, and more particularly to a GIS intermittent discharge mode identification model construction method, system and device. BACKGROUND

[0002] Gas insulated switchgear (GIS) has been widely applied in different voltage level systems of urban distribution network due to its excellent insulation performance, small footprint, small maintenance workload, strong environmental adaptability and other advantages. However, changes may occur in the internal of GIS during its production, transportation and maintenance, which may cause defects of GIS. The defects will cause partial discharge (PD) of GIS, and the PD may cause insulation failure of GIS and thus cause power grid outage accident. In order to avoid the impact of GIS with defects on the power grid, domestic and foreign technical personnel check the GIS with defects by disassembling the GIS with abnormal signals detected and checking it.

[0003] Although a large amount of theoretical and experimental researches have been made on the technology of detecting PD insulation defects at home and abroad, in recent years, the existing GIS PD online monitoring system has a lot of false positives and false negatives, and intermittent discharge is the main cause of false positives and false negatives. The time interval between each discharge of intermittent discharge is much larger than the time interval of stable discharge, which causes the traditional PD detection system to judge the intermittent discharge pulse as an interference signal and be ignored.

[0004] MRMR (Max-Relevance and Min-Redundancy) is a minimum redundancy and maximum correlation algorithm, which is used to solve the problem that the best m features are obtained by maximizing the correlation between the features and the target variables. The MRMR algorithm ensures the maximum correlation while having the minimum redundancy between each other. For the signals collected by multi-source detection of GIS intermittent discharge from the aspects of sound, light, electricity, magnetism and chemistry, the feature parameters that can effectively represent the fault state of GIS intermittent discharge are extracted. The redundancy between each feature parameter can be analyzed by using the MRMR algorithm, and then the PD feature parameters are sorted and selected. The PD fault state evaluation feature set composed of the optimal features can be used to construct the GIS intermittent discharge fault diagnosis system, and guide the fault diagnosis and state evaluation of the on-site power equipment.

[0005] As a mature artificial neural network model, the BP neural network is a kind of model for learning and storing input-output mode mapping relationship, and its structure contains three layers of input layer, hidden layer and output layer. The BP neural network mainly has two processes of forward propagation and backward error propagation, the input signal is transmitted to the hidden layer through the input layer, the neurons of the hidden layer have corresponding activation functions, and then the input signal is input to the output layer through the activation functions, so that the effective recognition of the discharge mode can be realized based on the discharge characteristic quantity. By using the BP neural network, not only the fault state of GIS intermittent discharge can be evaluated, but also the advantages and disadvantages of the partial discharge feature set selected by the evaluation accuracy test can be found out, and the optimal GIS intermittent discharge partial discharge feature set can be found out. SUMMARY

[0006] The application provides a GIS intermittent discharge mode recognition model construction method, system and equipment, which combines the MRMR algorithm and the BP neural network algorithm, solves the problems of GIS intermittent discharge feature optimization and mode recognition, and can accurately recognize the intermittent discharge defect type.

[0007] The application provides a GIS intermittent discharge mode recognition model construction method, which comprises the following steps:

[0008] Collecting different types of partial discharge information in the GIS intermittent discharge process;

[0009] According to different types of partial discharge information, corresponding types of partial discharge features are obtained, and preliminary feature parameters are extracted from the partial discharge features, different types of preliminary feature parameters constitute m different feature subsets, each preliminary feature parameter in the m different feature subsets is divided into a training set and a test set, and an initial BP neural network is trained by using the training set;

[0010] From the test set, the partial discharge features V X and V Y with the maximum mutual information are obtained, according to the partial discharge features V X and V Y , n optimal feature sets satisfying the MRMR criterion are searched from the set m different feature subsets, and the n optimal feature sets constitute a partial discharge fault state evaluation feature set;

[0011] The partial discharge fault state evaluation feature set is used to test the trained initial BP neural network, so as to adjust the parameters of the initial BP neural network, including adjusting the number of layers and the number of training times of the initial BP neural network, and obtaining a GIS intermittent discharge mode recognition model.

[0012] Further, the partial discharge features corresponding to different types of partial discharge information are obtained by:

[0013] The MRMR algorithm is used to reduce the dimensionality of the partial discharge information, obtain different variables X and Y from the partial discharge information, and determine the mutual information between variables X and Y in each partial discharge information after dimensionality reduction.

[0014] Furthermore, the partial discharge information includes ultra-high frequency signals, ultrasonic signals, pulsed current signals, optical signals, and chemical information.

[0015] Further, preliminary feature parameters are extracted from the partial discharge features, followed by:

[0016] The preliminary feature parameters are normalized, and m different feature subsets are constructed using the different types of preliminary feature parameters after data normalization.

[0017] Furthermore, for discrete variables X and Y, the mutual information I(X,Y) between variables X and Y is defined as follows:

[0018]

[0019] For continuous variables X and Y, the mutual information I(X,Y) between variables X and Y is defined as follows:

[0020]

[0021] Where p(x), p(y), and p(x,y) are the probability density of variable x, the probability density of variable y, and the joint probability density of variables x and y, respectively.

[0022] Furthermore, the partial discharge feature V with the largest mutual information is obtained from the test set. x and V y According to the partial discharge characteristics V x and V y Finding n optimal feature sets that satisfy the MRMR criterion from m different feature subsets includes the following steps:

[0023] Obtain the partial discharge feature V with the highest mutual information from the test set. X and V Y The partial discharge feature V, which constitutes the largest mutual information X and V Y A feature subset S with the largest mutual information is formed, and n optimal feature sets satisfying the MRMR criterion are searched from the feature subset S.

[0024] Further, finding the n optimal feature sets satisfying the MRMR criterion from the feature subset S includes the following steps:

[0025] Based on the MRMR criterion, partial discharge features V that satisfy the MRMR criterion are initially selected from the feature subset S.X a first optimal subset S1, the expression of the first optimal subset S1 is:

[0026] S1 = {Vx}

[0027] wherein V x represents the feature subset of the partial discharge feature V X with the maximum mutual information in the feature subset;

[0028] If the number of optimal feature sets in the first optimal subset S1 reaches n, then the n optimal feature sets form a partial discharge fault state evaluation feature set;

[0029] If the number of optimal feature sets in the first optimal subset S1 does not reach n, then a first remaining feature set F1 is determined from the m different feature subsets, the expression of F1 is:

[0030] F1 = F-S1

[0031] wherein F is the set of the m feature subsets, and F1 is the first remaining feature set after the first optimal subset S1 is removed from the set of the feature subsets;

[0032] And continue to find the partial discharge feature V Y that meets the MRMR criterion from the first remaining feature set F1, and fuse it with the first optimal subset S1 to obtain a second optimal subset S2, the expression of S2 is:

[0033] S2 = S1∪{Vy}

[0034] wherein V y represents the feature subset of the partial discharge feature V Y with the maximum mutual information in the feature subset, and the second optimal subset S2 is the intersection of the optimal subset S1 and the feature subset of the partial discharge feature V Y ;

[0035] When the number of features in the second optimal subset S2 does not reach n, then the related parameters in the MRMR criterion are adjusted to obtain a new MRMR criterion, and continue to determine a second remaining feature set F2 from the first remaining feature set F1, and find the partial discharge feature that meets the new MRMR criterion from the second remaining feature set F2, until the number of features in the second optimal subset S2 reaches n;

[0036] The expression of the second remaining feature set F2 is:

[0037] F2 = F1-S2

[0038] F2 represents the feature set after the second optimal subset S2 is removed on the basis of the first remaining feature set F1.

[0039] Furthermore, a model for identifying the intermittent discharge pattern of GIS is obtained, followed by the following steps:

[0040] Real-time partial discharge information of different types during the intermittent discharge process of GIS is obtained, and the intermittent discharge defect type is obtained by identification through the identification model.

[0041] This invention also provides a system for constructing a GIS intermittent discharge pattern recognition model, comprising:

[0042] The partial discharge information module is used to collect different types of partial discharge information during the intermittent discharge process of GIS;

[0043] The initial BP neural network acquisition module acquires corresponding local discharge features based on different types of local discharge information, and extracts preliminary feature parameters from the local discharge features. Different types of preliminary feature parameters constitute m different feature subsets. The preliminary feature parameters in the m different feature subsets are divided into training set and test set. The initial BP neural network is trained using the training set.

[0044] The partial discharge fault state assessment feature set acquisition module is used to obtain the partial discharge feature V with the maximum mutual information from the test set. X and V Y According to the partial discharge characteristics V X and V Y Find n optimal feature sets that satisfy the MRMR criterion from the set m different feature subsets, and form the n optimal feature sets into a partial discharge fault state assessment feature set;

[0045] The identification model acquisition module is used to test the trained initial BP neural network using the partial discharge fault state assessment feature set, so as to adjust the parameters of the initial BP neural network, including adjusting the number of layers and training times of the initial BP neural network, to obtain the identification model of the GIS intermittent discharge mode.

[0046] The present invention also provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the method for constructing a GIS intermittent discharge pattern recognition model.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] 1、The present application combines MRMR algorithm and BP neural network algorithm, solves the problem of GIS intermittent discharge feature optimization and pattern recognition, and the overall recognition rate is better than that of single BP neural network, which can not only evaluate GIS intermittent discharge fault state, but also test the advantages and disadvantages of the optimized partial discharge feature set through evaluation accuracy, and find out the optimal GIS intermittent discharge partial discharge feature set.

[0049] 2、The present application uses sound, light, electricity, magnetism and chemical discharge signals collected by multi-source detection technology to select optimal feature quantity and realize pattern recognition to evaluate GIS intermittent discharge fault state. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the present application or prior art, the following will briefly introduce the drawings needed to be used in the description of the present application or prior art. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0051] Fig. 1 The present application provides a GIS intermittent discharge pattern recognition model construction method, including the following steps:

[0052] Fig. 2 The present application provides a GIS intermittent discharge pattern recognition model construction method, including the following steps:

[0053] Fig. 3 The present application provides a GIS intermittent discharge pattern recognition model construction method, including the following steps:

[0054] Fig. 4 The present application provides a GIS intermittent discharge pattern recognition model construction method, including the following steps: DETAILED DESCRIPTION

[0055] The present application provides a GIS intermittent discharge pattern recognition model construction method, including the following steps: Figs. 1-4 The present application provides a GIS intermittent discharge pattern recognition model construction method, including the following steps:

[0056] As shown in the drawings, the present embodiment provides a GIS intermittent discharge pattern recognition model construction method, including the following steps: Fig. 1 Collecting different types of partial discharge information in GIS intermittent discharge process;

[0057] Collecting different types of partial discharge information in GIS intermittent discharge process;

[0058] Collecting different types of partial discharge information in GIS intermittent discharge process;According to different types of partial discharge information, corresponding types of partial discharge characteristics are obtained, and preliminary feature parameters are extracted from the partial discharge characteristics, different types of preliminary feature parameters constitute m different feature subsets, each preliminary feature parameter in the m different feature subsets is divided into a training set and a test set, and an initial BP neural network is trained by using the training set;

[0059] From the test set, the partial discharge characteristic V X and V Y is obtained, according to the partial discharge characteristics V X and V Y , n optimal feature sets satisfying the MRMR criterion are searched from the set m different feature subsets, and the n optimal feature sets constitute a partial discharge fault state evaluation feature set;

[0060] The trained initial BP neural network is tested by using the partial discharge fault state evaluation feature set, so as to adjust the parameters of the initial BP neural network, including adjusting the number of layers and the number of training times of the initial BP neural network, and a GIS intermittent discharge mode recognition model is obtained.

[0061] The present application combines the MRMR algorithm and the BP neural network algorithm, solves the problems of GIS intermittent discharge feature optimization and pattern recognition, and the overall recognition rate is better than that of a single BP neural network, not only can the GIS intermittent discharge fault state be evaluated, but also the advantages and disadvantages of the optimal partial discharge feature set can be tested by evaluating the accuracy, and the optimal GIS intermittent discharge partial discharge feature set is found.

[0062] In the embodiment, the MRMR algorithm is used for dimension reduction processing of the partial discharge information, the redundant information of the data is reduced, different variables X and Y are obtained from the partial discharge information, and the mutual information between the variables X and Y in each partial discharge information after the dimension reduction processing is determined.

[0063] Further, the partial discharge information in the embodiment includes very high frequency signals, ultrasonic signals, pulse current signals, optical signals and chemical information. The specific partial discharge information is respectively a PRPD discharge spectrum of very high frequency, ultrasonic waves, pulse current and a time domain spectrum of optical signals including characteristic values such as variance, standard deviation, root mean square, skewness, waveform factor, peak factor, pulse factor and the like. The PRPD discharge spectrum of very high frequency includes characteristic values such as skewness, kurtosis, phase asymmetry, discharge quantity factor and the like, and the PRPD discharge spectrum of very high frequency also includes concentration values of characteristic components of gas decomposition; the characteristic values such as variance, standard deviation, root mean square, skewness, waveform factor, peak factor, pulse factor and the like contained in the time domain spectrum are obtained by very high frequency detection method, ultrasonic detection method, pulse current detection method and optical detection method, the characteristic values such as skewness, kurtosis, phase asymmetry, discharge quantity factor and the like contained in the PRPD discharge spectrum of very high frequency are obtained by very high frequency detection method, and the concentration value information is obtained by chemical detection method.

[0064] Wherein, the peak time Tp and signal mean μt of the time domain spectrum of ultrasonic waves are variables X and Y respectively, and the random variables X and Y of other acoustic, electric, magnetic and chemical detection methods are also selected with reference to them. The present application aims at GIS intermittent discharge, and uses the discharge signals of sound, light, electricity, magnetism and chemistry collected by multi-source detection technology to select optimal characteristic quantities and realize pattern recognition to evaluate the GIS intermittent discharge fault state.

[0065] The embodiment also needs to perform data normalization processing on the preliminary characteristic parameters, and the preliminary characteristic parameters of different types and after data normalization processing constitute m different characteristic subsets.

[0066] In the embodiment, the random variables X and Y are judged to belong to discrete type or continuous type by artificial means. For the discrete type variables X and Y, the mutual information I(X, Y) between the variables X and Y is defined as:

[0067]

[0068] For the continuous type variables X and Y, the mutual information I(X, Y) between the variables X and Y is defined as:

[0069]

[0070] Wherein, p(x), p(y) and p(x, y) are respectively the probability density of variable x, the probability density of variable y and the joint probability density of variables x and y, wherein x ∈ X, y ∈ Y; the greater the value of I(X, Y), the higher the similarity of the two variables, and if the two variables are independent of each other, I(X, Y) is equal to 0.

[0071] In the embodiment, the partial discharge characteristic V with the maximum mutual information is obtained from the test setX and V Y , the partial discharge feature V X with the largest mutual information Y , the feature subset S with the largest mutual information, and n optimal feature sets satisfying the MRMR criterion are searched from the feature subset S.

[0072] In the embodiment, searching n optimal feature sets satisfying the MRMR criterion from the feature subset S includes the following steps:

[0073] Preliminarily selecting the partial discharge features V X satisfying the MRMR criterion from the feature subset S according to the MRMR criterion to obtain a first optimal subset S1, and the expression of the first optimal subset S1 is:

[0074] S1 = {Vx}

[0075] wherein V x represents the feature subset screening the partial discharge features V X with the largest mutual information in the feature subset;

[0076] If the number of optimal feature sets in the first optimal subset S1 reaches n, the n optimal feature sets are combined to form the partial discharge fault state evaluation feature set;

[0077] If the number of optimal feature sets in the first optimal subset S1 does not reach n, a first remaining feature set F1 is determined from the m different feature subsets, and the expression of F1 is:

[0078] F1 = F-S1

[0079] wherein F is the set of the m feature subsets, and F1 is the first remaining feature set after the first optimal subset S1 is removed from the set of the feature subsets;

[0080] And the partial discharge feature V Y satisfying the MRMR criterion is searched from the first remaining feature set F1, and the second optimal subset S2 is obtained by fusing the first optimal subset S1 and the partial discharge feature V y , and the expression of S2 is:

[0081] S2 = S1∪{Vy}

[0082] wherein V y represents the feature subset screening the partial discharge features V Y with the largest mutual information in the feature subset, and the second optimal subset S2 is the intersection of the optimal subset S1 and the partial discharge feature V Y ;

[0083] When the number of features in the second optimal subset S2 does not reach n, the correlation parameter in the MRMR criterion is adjusted to obtain a new MRMR criterion, and the second remaining feature set F2 is determined from the first remaining feature set F1, and the partial discharge features satisfying the new MRMR criterion are searched from the second remaining feature set F2 until the number of features in the second optimal subset S2 reaches n;

[0084] The expression of the second remaining feature set F2 is:

[0085] F2 = F1 - S2

[0086] F2 represents the feature set after removing the second optimal subset S2 on the basis of the first remaining feature set F1.

[0087] In this embodiment, as shown in Fig. 2 Adjusting the parameters of the initial BP neural network includes the following steps:

[0088] The partial discharge state evaluation feature set is input to the trained initial BP neural network.

[0089] The parameters of the overall structure of the BP neural network are initialized, wherein the transfer function in the construction of the BP neural network model is purelin, the gradient descent training method is adopted, the number of neural network layers is 20; the training times are set to 1000; the learning rate is set to 0.01, the minimum error of the training target is set to 0.00001; the additional momentum factor is set to 0.01; the minimum performance gradient is set to 0.000001.

[0090] The input and output values of the BP neural network are calculated, the input layer is the input end of information, the output layer is the output end of information, and the input value of each neuron in the hidden layer is determined by the output data of the previous layer through the corresponding weights between different nodes and the sum of the bias values of the hidden layer nodes.

[0091] The output layer error Loss is calculated, and the bias and weight are updated by deriving the error function of the output value calculated by each forward propagation and the actual value.

[0092] When the output layer error Loss reaches the minimum value, the output result is output, and the GIS internal intermittent discharge fault type recognition is realized; when the output layer error Loss is small, the weight and threshold value are corrected until the Loss reaches the minimum value, and the final output result is output.

[0093] When the output layer error Loss is large, the input and output values of the BP neural network are recalculated, and the output layer error Loss is recalculated. When it reaches the minimum value, the output result is output, and the GIS internal intermittent discharge fault type recognition is realized.

[0094] In the embodiment, the activation function of the output layer is Softmax, the Softmax function is an activation function for multi-class classification problems, the denominator of the Softmax function combines all factors of the original output value, and the obtained various probabilities are related to each other, and the expression is as follows:

[0095]

[0096] In the formula, S i is an output value of the output layer, V i is an output of a previous layer, C is a total number of categories, and i is an index of a category.

[0097] In the embodiment, the error function is a classification cross-entropy function, and the expression is as follows:

[0098]

[0099] In the formula, Loss represents an error of the output layer; t i represents a true value, C is a total number of categories, i is an index of a category, and y i represents a value of the obtained Softmax classifier.

[0100] The embodiment also provides a GIS intermittent partial discharge mode recognition model construction system, including:

[0101] A partial discharge information module is configured to collect different types of partial discharge information in a GIS intermittent partial discharge process.

[0102] An initial BP neural network acquisition module is configured to acquire partial discharge characteristics of different types according to the different types of partial discharge information, extract preliminary feature parameters from the partial discharge characteristics, divide each preliminary feature parameter in m different feature subsets into a training set and a test set, and train an initial BP neural network by using the training set.

[0103] A partial discharge fault state evaluation feature set acquisition module is configured to acquire partial discharge features V X and V Y with maximum mutual information from the test set, find n optimal feature sets that satisfy the MRMR criterion from the m different feature subsets according to the partial discharge features V X and V Y , and group the n optimal feature sets into a partial discharge fault state evaluation feature set.

[0104] The recognition model obtaining module is configured to test the trained initial BP neural network by using the partial discharge fault state evaluation feature set, so as to adjust parameters of the initial BP neural network, including adjusting the number of layers of the initial BP neural network and the number of training times, and obtain a recognition model of GIS intermittent discharge mode.

[0105] The application further provides an electronic device, comprising a memory and a processor, which are in communication connection with each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the GIS intermittent discharge mode recognition model construction method.

[0106] As Figs. 3-4 In the embodiment, a GIS intermittent discharge test platform is built in a laboratory, and time-domain high-frequency data of four typical GIS internal intermittent discharge defects, i.e., insulator surface metal contamination, insulator internal air gap, suspended electrode and metal protrusion, are collected, 100 groups of data are collected for each defect, 70 groups of defect data in each group are used as a training set, the initial BP neural network is trained by using the training set, and the remaining 30 groups are used as a test set. In the embodiment, the specific features selected in the optimal feature set are as follows: for the collected PRPD spectrum signal, the positive half cycle, negative half cycle, positive and negative skewness and steepness, positive and negative half cycle discharge times and amplitude ratio and positive and negative half cycle cross-correlation coefficient in the spectrum are extracted; and for the collected original TRPD signal, the root mean square, variance, absolute integral average, kurtosis, skewness, maximum power spectrum value, median frequency and average power frequency in the signal are extracted.

[0107] The four different types of preliminary feature parameters constitute 24 different feature subsets, the preliminary feature parameters in the feature subsets are divided into the training set and the test set, and the initial BP neural network is trained by using the training set. The BP neural network is used for pattern recognition research on the original data, the MRMR algorithm is used for feature optimization of the original data feature quantity, and the optimized feature quantity is input into the BP neural network for pattern recognition research. In the embodiment, by acquiring different types of real-time partial discharge information in the GIS intermittent discharge process, the recognition model is used for recognition, and the intermittent discharge defect type

[0108] The partial discharge features V X and V Y with the maximum mutual information are acquired from the test set, and the partial discharge features V X and V Yn optimal feature sets satisfying the MRMR criterion are searched from the set m different feature subsets, and the n optimal feature sets are combined to form the partial discharge fault state evaluation feature set. The n optimal feature sets are output to the trained BP neural network to test the trained initial BP neural network by using the partial discharge fault state evaluation feature set, and the parameters of the initial BP neural network are adjusted, including adjusting the number of neural network layers and the number of training times, and the predicted value and the true value of different defect recognition are output through the confusion matrix.

[0109] Fig. 3 , Fig. 4 respectively, are the confusion matrix diagrams of the existing BP neural network recognition output and the BP neural network recognition output based on the MRMR feature optimization of the application, wherein the true label represents the original classification of the tested data, the predicted label represents the predicted classification of the tested data, the data in the diagram represent the recognition rate, and NaN represents a "not a number" value. It can be seen from the output confusion matrix diagram that the overall recognition rate of the BP neural network based on the MRMR feature optimization can reach 95%, the overall recognition rate of the BP neural network is 85%, and the overall recognition rate of the BP neural network based on the MRMR feature optimization is better than the overall recognition rate of the single BP neural network.

Claims

1. A method for constructing a GIS intermittent discharge pattern recognition model, characterized in that, The method comprises the following steps: Collecting different types of partial discharge information in GIS intermittent discharge process; According to different types of partial discharge information, corresponding types of partial discharge characteristics are obtained, and preliminary feature parameters are extracted from the partial discharge characteristics, different types of preliminary feature parameters constitute m different feature subsets, each preliminary feature parameter in the m different feature subsets is divided into a training set and a test set, and an initial BP neural network is trained by using the training set; obtaining partial discharge features V with maximum mutual information from the test set X and V Y , according to the partial discharge features V X and V Y , finding n optimal feature sets satisfying the MRMR criterion from the set m different feature subsets, and combining the n optimal feature sets to form a partial discharge fault state evaluation feature set; The trained initial BP neural network is tested by using the partial discharge fault state evaluation feature set to adjust the parameters of the initial BP neural network, including adjusting the number of neural network layers and the number of training times, to obtain a GIS intermittent discharge mode recognition model.

2. The method according to claim 1, wherein, The partial discharge information is processed by dimension reduction using the MRMR algorithm, different variables X and Y are obtained from the partial discharge information, and the mutual information between variables X and Y in each partial discharge information after dimension reduction is determined. The partial discharge information includes ultra-high frequency signal, ultrasonic signal, pulse current signal, optical signal and chemical information.

3. The method according to claim 1, wherein, After extracting the preliminary feature parameters from the partial discharge characteristics, the following steps are further included:

4. The method according to claim 1, wherein, The preliminary feature parameters are subjected to data normalization processing, and different types of preliminary feature parameters subjected to data normalization processing constitute m different feature subsets. For discrete variables X and Y, the mutual information I(X, Y) between the variables X and Y is defined as:

5. The method according to claim 2, wherein, For continuous variables X and Y, the mutual information I(X, Y) between the variables X and Y is defined as: Where p(x), p(y) and p(x, y) are the probability density of variable x, the probability density of variable y and the joint probability density of variables x and y, respectively. Finding n optimal feature sets from the feature subset S that meet the MRMR criterion comprises the following steps:

6. The method of claim 1, wherein the method comprises: obtaining partial discharge features V from the test set X and V Y , according to partial discharge features V X and V Y , from the set of m different feature subsets, find n optimal feature sets that meet the MRMR criteria, including the following steps: obtaining a partial discharge feature V with maximum mutual information from the test set X and V Y , obtaining a partial discharge feature V with maximum mutual information from the test set X and V Y , obtaining a feature subset S with maximum mutual information, and finding n optimal feature sets satisfying MRMR criterion from the feature subset S.

7. The method according to claim 6, wherein, S1={Vx} preliminarily selecting, from the feature subset S, partial discharge features V satisfying the MRMR criterion according to the MRMR criterion X a first optimal subset S1, an expression of the first optimal subset S1 being: If the number of optimal feature sets in the first optimal subset S1 has reached n, n optimal feature sets form a partial discharge fault state evaluation feature set; wherein V x represents the partial discharge feature V X with the largest mutual information in the feature subset; If the number of optimal feature sets in the first optimal subset S1 has not reached n, a first remaining feature set F1 is determined from the m different feature subsets, and the expression of F1 is: F1=F-S1 Where F is the set of m feature subsets, and F1 is the first remaining feature set after removing the first optimal subset S1 from the set of feature subsets; S2=S1∪{Vy} and continue to find a localized feature V satisfying the MRMR criterion from the first remaining feature set F1 y and the first optimal subset S1 to obtain a second optimal subset S2, and the expression of S2 is: When the number of features in the second optimal subset S2 does not reach n, the related parameters in the MRMR criterion are adjusted to obtain a new MRMR criterion, and the second remaining feature set F2 is continuously determined from the first remaining feature set F1, the partial discharge characteristics that meet the new MRMR criterion are found from the second remaining feature set F2, until the number of features in the second optimal subset S2 reaches n; wherein V y represents the partial discharge feature V Y whose mutual information is the largest in the feature subset, and the second optimal subset S2 is the intersection of the optimal subset S1 and the partial discharge feature V Y . The expression of the second remaining feature set F2 is: F2=F1-S2 F2 represents the feature set after removing the second optimal subset S2 from the first remaining feature set F1. After obtaining the GIS intermittent discharge mode recognition model, the following steps are further included:

8. The method according to claim 1, wherein, ​ The GIS intermittent discharge process is acquired in real time, and different types of partial discharge information are obtained through the identification model to obtain the intermittent discharge defect type.

9. A GIS intermittent discharge pattern recognition model construction system, characterized in that, The method comprises the steps of: a partial discharge information module is used to collect different types of partial discharge information in the GIS intermittent discharge process; an initial BP neural network acquisition module is used to acquire partial discharge characteristics of different types according to different types of partial discharge information, and to extract preliminary feature parameters from the partial discharge characteristics, different types of preliminary feature parameters constitute m different feature subsets, each preliminary feature parameter in the m different feature subsets is divided into a training set and a test set, and the initial BP neural network is trained by using the training set; The partial discharge fault state assessment feature set acquisition module is used to obtain the partial discharge feature V with the maximum mutual information from the test set. X and V Y According to the partial discharge characteristics V X and V Y Find n optimal feature sets that satisfy the MRMR criterion from the set m different feature subsets, and form the n optimal feature sets into a partial discharge fault state assessment feature set; an identification model acquisition module is used to test the trained initial BP neural network by using the partial discharge fault state evaluation feature set, to adjust the parameters of the initial BP neural network, including adjusting the number of layers and the number of training times of the initial BP neural network, and to obtain the identification model of the GIS intermittent discharge mode.

10. An electronic device, comprising: The method comprises the steps of: a memory and a processor are connected in communication with each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the method for constructing a GIS intermittent discharge mode identification model according to any one of claims 1-8.

Images