CN107491792B - Power grid fault classification method based on feature mapping transfer learning - Google Patents

Power grid fault classification method based on feature mapping transfer learning Download PDF

Info

Publication number
CN107491792B
CN107491792B CN201710756382.4A CN201710756382A CN107491792B CN 107491792 B CN107491792 B CN 107491792B CN 201710756382 A CN201710756382 A CN 201710756382A CN 107491792 B CN107491792 B CN 107491792B
Authority
CN
China
Prior art keywords
fault
auxiliary source
data
feature
target
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.)
Active
Application number
CN201710756382.4A
Other languages
Chinese (zh)
Other versions
CN107491792A (en
Inventor
张化光
刘鑫蕊
孙秋野
于晓婷
杨珺
王智良
赵鑫
吴泽群
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.)
Northeastern University China
Original Assignee
Northeastern University China
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 Northeastern University China filed Critical Northeastern University China
Priority to CN201710756382.4A priority Critical patent/CN107491792B/en
Publication of CN107491792A publication Critical patent/CN107491792A/en
Application granted granted Critical
Publication of CN107491792B publication Critical patent/CN107491792B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • 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
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting

Abstract

The invention discloses a power grid fault classification method based on feature mapping transfer learning, which comprises the following steps: 1. selecting target domain data and auxiliary source domain data; 2. respectively extracting fault features of target field data and auxiliary source field data based on micro-increment wavelet singular entropy, and taking each micro-increment wavelet singular entropy as a fault feature to respectively form a feature vector space corresponding to the target field and a feature vector space corresponding to the auxiliary source field; 3. finding out base vectors corresponding to the axis features, the specific features of the auxiliary source field and the specific features of the target field based on a feature mapping migration learning method; 4. taking the obtained base vector corresponding to the auxiliary source field as a support vector; and simultaneously setting a similarity penalty item and adding a constraint condition of a support vector training set to jointly train a classifier to obtain a corresponding classification result. The method can accurately and quickly find the three groups of base vectors which can reflect the fault types most.

Description

Power grid fault classification method based on feature mapping transfer learning
Technical Field
The invention belongs to the technical field of power transmission and distribution, and particularly relates to a power grid fault classification method based on feature mapping migration learning.
Background
The increasing scale of the power grid and the increasing transmission capacity and voltage level bring huge economic and social benefits, but meanwhile, the fault of the power grid can cause more serious harm to social economy and people's life. The rapid and accurate classification of the grid faults is a precondition for rapidly recovering the power supply of the grid and is an important part of fault analysis, so that the research on the rapid and reliable fault classification method has important significance for guaranteeing the safety and the economy of a power system.
Classification has been widely studied and applied as an important machine learning method; the method mainly comprises the steps of training a classification model according to source field data and then predicting the type of target field data by using the classification model. In order to ensure that the trained classification model has accuracy and high reliability, the traditional classification learning needs to satisfy two basic assumptions: (1) the training sample for learning and the new test sample meet the condition of independent and same distribution; (2) there must be enough training samples available to learn a good classification model. However, in practical applications, we find that these two conditions are often not satisfied.
In order to solve the problems of insufficient data quantity and characteristic difference, most machine learning algorithms adopt the method of re-marking fault samples, but a large amount of experiments and professional knowledge are needed, and the collected marked data and the fault data in the target field cannot be distributed consistently due to the change of factors such as the operation mode of a power grid, load and the like, so that the reliability of a diagnosis result is reduced.
The applicant researches and discovers that migration learning as a cross-field and cross-task learning method attracts more and more attention of scholars in the field of machine learning. Transfer learning is a new machine learning method that solves problems in different but related fields using existing knowledge. The method relaxes two basic assumptions in the traditional machine learning, aims to transfer knowledge learned from the source field to the target field under the condition that the source field data and the target field data have different data distributions, and solves the learning problem that only a small amount of labeled sample data exists in the target field or even the target field does not exist. When a power grid fails, the network topology structure changes, the data distribution changes, and based on a transfer learning method, knowledge of auxiliary data which is different from target data but related to the target data is fully utilized, so that the fault classification performance of a machine learning algorithm on the power grid can be effectively improved.
Therefore, the power grid fault classification based on the transfer learning has certain theoretical basis and practical significance.
Disclosure of Invention
In view of the defects in the prior art, the invention aims to provide a power grid fault classification method based on feature mapping migration learning, which can be used for mapping data of each field from an original high-dimensional feature space to a low-dimensional feature space by analyzing the correlation between the characteristic features of an auxiliary source field and the characteristic features of a target field and axis features in an abstract manner, so that the source field data and the target field data have similar distribution in the low-dimensional space; and solving the maximum value of the relation coefficient by a Lagrange multiplier method, and further finding out three groups of base vectors which can reflect the fault types most.
In order to achieve the purpose, the technical scheme of the invention is as follows:
a power grid fault classification method based on feature mapping migration learning is characterized by comprising the following steps:
step 1, selecting target domain data to be classified and auxiliary source domain data, wherein the target domain data at least comprises: three-phase current data of each fault line at each fault moment; the auxiliary source domain data includes: three-phase current data of each fault line at the previous fault moment corresponding to each fault moment, three-phase current data of each fault line at the previous normal operation moment corresponding to each fault moment and three-phase current data of adjacent lines of the fault line at each fault moment;
step 2, fault feature extraction based on micro-increment wavelet singular entropy is respectively carried out on target field data and auxiliary source field data to extract respective corresponding micro-increment wavelet singular entropy, each micro-increment wavelet singular entropy is taken as a fault feature, and then a feature vector space corresponding to the target field and a feature vector space corresponding to the auxiliary source field are respectively formed;
step 3, based on a feature mapping migration learning method, using the intersection of the auxiliary source field and the target field as an axis feature, and finding out base vectors corresponding to the axis feature, the characteristic feature of the auxiliary source field and the characteristic feature of the target field based on a Lagrange multiplier method to obtain an extreme value;
step 4, in the fault classification process based on the SVM, taking the base vector corresponding to the auxiliary source field obtained in the step 3 as a support vector; meanwhile, a similarity punishment item of a support vector training set corresponding to the auxiliary source field is added into an original target function of the support vector machine SVM, and a constraint condition of the added support vector training set is added into an original target function constraint condition, so that a classifier is trained together to obtain a corresponding classification result.
Further, the step 2 comprises:
step 21, respectively performing m-layer wavelet multi-resolution signal decomposition on the target field data and the auxiliary source field data to obtain a wavelet transformation coefficient matrix corresponding to a wavelet transformation result, and performing singular value decomposition calculation to obtain a singular value feature matrix corresponding to the wavelet transformation coefficient matrix, and marking the singular value feature matrix as Λ ═ diag (λ ═ diag)1,λ2,…λn)；
Step 22, respectively constructing n-order micro-increment wavelet singular entropies of the target field data and the auxiliary source field data, wherein the corresponding formula is
Figure BDA0001392261310000031
In the formula, λiIs the i-th order non-zero singular eigenvalue, XiIs λiThe ith micro-increment wavelet singular entropy of (a);
step 23, constructing a feature vector X by using the n-order micro-increment wavelet singular entropy element of the auxiliary source field datas1Is marked as Xs1＝[X1,X2…Xn]Simultaneously order
Figure BDA0001392261310000032
The corresponding normalized wavelet packet feature vector Xs1 *Is represented by Xs1 *＝[X1/X,X2/X,…,Xn/X]And forming a vector space X of the auxiliary source domain datas *＝[Xs1 *,Xs2 *,…Xsn *](ii) a Vector space X which similarly constitutes target domain datat *＝[Xt1 *,Xt2 *,…Xtn *]。
Further, n ═ m in the singular value feature matrix2-1 and such that λnAnd the constraint condition is satisfied.
Further, the step 3 comprises:
step 31, defining auxiliary source field Xs *The fault identifier of the known fault type is Y, so that a certain fault type identifier Y belongs to Y; auxiliary source field Xs *And the target area Xt *The intersection of (A) is the corresponding axis feature or is called the domain axis feature X∩ *∈Xs *∩Xt *Simultaneous calculation of axial features X∩ *And the correlation coefficient between Y, the corresponding calculation formula is as follows:
Figure BDA0001392261310000033
wherein, I (X)∩ *Y) represents an axial feature X∩ *Correlation coefficient with Y, P (X)∩ *Y) field axis feature X∩ *Joint distribution probability with fault identity y, P (X)∩ *) Representing axial characteristics X∩ *Data X appearing in the auxiliary Source Domains *P (y) indicates that the fault mark y appears in the target area data Xt *And selecting the axis feature with the maximum correlation coefficient value in m-layer wavelet multi-resolution signal decomposition to form an axis feature set, and recording as X∩＝{X∩1 *,X∩2 *,…,X∩m *}；
And selecting the axis feature with the maximum correlation coefficient value in m-layer wavelet multi-resolution signal decomposition to form an axis feature set, and recording as X∩＝{X∩1 *,X∩2 *,…,X∩m *}；
Step 32, firstly, based on the union formed by the extracted fault features in the auxiliary source domain data and the target domain data
Figure BDA0001392261310000041
Three sets of paired sample sets of random variables α, γ, were constructed and labeled
Figure BDA0001392261310000042
Wherein | X∩|，
Figure BDA0001392261310000043
Figure BDA0001392261310000044
Respectively representing the dimensions of the axis features, the dimensions of the fault features of the auxiliary source domain data, the dimensions of the fault features of the target domain data, and
Figure BDA0001392261310000045
representing sample points X in auxiliary source domain datas *In the axial feature space X∩The value of (a) is selected from,
Figure BDA0001392261310000046
representing auxiliary source domain data sample points Xs *In a feature space
Figure BDA0001392261310000047
The value of (a) is selected from,
Figure BDA0001392261310000048
representing sample points in target domain data
Figure BDA0001392261310000049
In a feature space
Figure BDA00013922613100000410
The value of (a) is selected from,
then linearly combined according to
Figure BDA00013922613100000411
Find three groups of base vectors by the principle that the correlation coefficient between the three groups reaches the maximum
Figure BDA00013922613100000412
Namely, based on the following formula
Figure BDA00013922613100000413
Corresponding constraint condition
Figure BDA00013922613100000414
Wherein C isAA＝(AS∪At)(AS∪At)T
Figure BDA00013922613100000415
Figure BDA00013922613100000416
Figure BDA00013922613100000417
Figure BDA00013922613100000418
Figure BDA00013922613100000419
Figure BDA00013922613100000420
Figure BDA00013922613100000421
Wherein: wAIs a set of basis vectors corresponding to the axial features; wSThe method comprises the following steps of (1) acquiring a base vector set corresponding to the characteristic features of the auxiliary source field; wTA base vector set corresponding to the characteristic features of the target field; cssIs a fault characteristic D in auxiliary source field datasA covariance matrix of medial axis features; a. theSIs | X about α∩|×nsA matrix of dimensions; a. thetIs | X about α∩|×ntMatrix of dimensions, S is about β
Figure BDA0001392261310000051
A matrix of dimensions, T being about β
Figure BDA0001392261310000052
A matrix of dimensions; cTTRefer to the failure characteristics D in the target domain datatA covariance matrix of medial axis features; cAAIs a fault characteristic D in auxiliary source field datasAnd fault signature D in target domain datatUnion D ofs∪DtA covariance matrix of medial axis features;
step 33, finding out the base vectors corresponding to the axis features, the fault features in the auxiliary source field and the fault features in the target field based on the method for solving the extreme value by the Lagrange multiplier method, namely finding out the base vectors corresponding to the axis features, the fault features in the auxiliary source field and the fault features in the target field based on the following formulas:
Figure BDA0001392261310000053
Figure BDA0001392261310000054
Figure BDA0001392261310000055
Figure BDA0001392261310000056
Figure BDA0001392261310000057
the eigenvectors corresponding to the first m generalized eigenvalues of the matrix are the base vectors WA，WS，WT。
Further, the step 4 comprises:
step 41, in the fault classification process based on the SVM, firstly, the base vector W corresponding to the auxiliary source field obtained in the step 3 is usedSAs a support vector; meanwhile, adding a similarity penalty term of a support vector training set corresponding to the field of the auxiliary source into the original target function in the support vector machine SVM, and recording the similarity penalty term as
Figure BDA0001392261310000058
Adding the constraint condition of the added support vector training set into the constraint condition of the original objective function; then the support vector training set V containing the auxiliary source field data in the support vector machine SVMsThe optimization process of the training sample T is
Figure BDA0001392261310000061
Constraint conditions
Figure BDA0001392261310000062
Figure BDA0001392261310000063
Wherein N istIs the number of i, Ns-NtIs the number of the j's,
Figure BDA0001392261310000064
k is the number of training sets of target domain data,
Figure BDA0001392261310000065
is the support vector of the jth auxiliary source field data, DtIs the training data corresponding to the target domain data,
Figure BDA0001392261310000066
representing the distance, γ, of the jth support vector from the training datat、γsRegularization coefficients for the target domain data and the auxiliary source domain data respectively,
Figure BDA0001392261310000067
Figure BDA0001392261310000068
is the squared term of the error function;
then, a Lagrange multiplier method is used for optimization, namely, in order to achieve the minimum loss function between the predicted value and the real category label, an SVM function estimation expression added with an auxiliary support vector set, namely an improved SVM function estimation expression is as follows:
Figure BDA0001392261310000069
and step 42, obtaining corresponding classification results by constructing and combining a plurality of two classifiers.
Further, said step 42 includes obtaining a corresponding classification result by using a decision binary tree method.
Compared with the prior art, the invention has the beneficial effects that:
the invention relaxes the conditions that the training and testing data distribution is the same and the target diagnosis data quantity is sufficient for a data source, and increases the auxiliary source field data, so that the auxiliary source field data can effectively help the target field to realize classification by a migration learning method, in particular to the method which is characterized in that a characteristic value diagonal array can simply reflect the time-frequency distribution characteristics of fault signals, a micro-increment wavelet singular entropy can quantitatively distinguish signals with different time-frequency distributions, quantitatively express the characteristics of the data on the distribution trend, and quantitatively reflect the characteristics of system uncertainty, complexity and the like by the statistical analysis of information. By analyzing the correlation between the specific characteristics of the auxiliary source field and the specific characteristics of the target field and the axis characteristics in an abstract way, the data of each field are effectively mapped to a low-dimensional characteristic space from an original high-dimensional characteristic space, the maximum value of a relation coefficient is solved by a Lagrange multiplier method, three groups of base vectors which can reflect fault categories most are found, and finally, the base vectors in the auxiliary source field are used as support vectors, and the support vectors are given certain weight through penalty terms and are trained with a training set of the target field together with a classifier, so that the base vectors with classification discrimination capability can greatly improve the classification precision.
Drawings
FIG. 1 is a flow chart of the steps corresponding to the method of the present invention;
FIG. 2 is a diagram of core steps corresponding to an embodiment of the method of the present invention;
FIG. 3 is a simplified model of a grid line according to the embodiment of the present invention;
FIG. 4 is a structural diagram of a decision binary tree multi-classification according to the embodiment of the present invention;
fig. 5 shows the projection result of the basis vectors based on the transfer learning according to the embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
1-2, the grid fault classification method based on feature mapping migration learning is characterized by comprising the following steps:
step 1, selecting target domain data to be classified and auxiliary source domain data, wherein the target domain data at least comprises: three-phase current data of each fault line at each fault moment, namely the size and direction of the three-phase current; the auxiliary source domain data includes: three-phase current data of each fault line at the previous fault moment corresponding to each fault moment, three-phase current data of each fault line at the normal operation moment corresponding to each fault moment and three-phase current data of each fault line at the fault moment corresponding to the adjacent line corresponding to the fault line; if the calculation is carried out in 24 hours, the current fault moment is the current fault moment when the fault occurs today, the last fault moment is the previous fault moment, if the fault occurs yesterday, the three-phase current data when the fault occurs today is used as the target field data, and the fault data related yesterday is contained in the source field data;
step 2, fault feature extraction based on micro-increment wavelet singular entropy is respectively carried out on target field data and auxiliary source field data to extract respective corresponding micro-increment wavelet singular entropy, each micro-increment wavelet singular entropy is taken as a fault feature, and then a feature vector space corresponding to the target field and a feature vector space corresponding to the auxiliary source field are respectively formed; further, the step 2 comprises:
step 21, performing m-layer wavelet multi-resolution signal decomposition on the target domain data and the auxiliary source domain data respectively to obtain a wavelet transform coefficient matrix corresponding to a wavelet transform result,obtaining a singular value feature matrix corresponding to the wavelet transform coefficient matrix after singular value decomposition calculation (the singular value feature matrix represents the basic modal feature of the wavelet transform coefficient matrix), and marking as lambda ═ diag (lambda ═ diag)1,λ2,…λn)；
Step 22, organically combining the wavelet transformation, singular value decomposition and information entropy to form a micro-increment wavelet singular entropy, specifically, n-order micro-increment wavelet singular entropy for respectively constructing target field data and auxiliary source field data, and the corresponding formula is
Figure BDA0001392261310000081
In the formula, XiFor the non-zero singular value λ of the ith orderiThe micro-increment wavelet singular entropy;
step 23, constructing a feature vector X by using the n-order micro-increment wavelet singular entropy element of the auxiliary source field datas1Is marked as Xs1＝[X1,X2…Xn]Simultaneously order
Figure BDA0001392261310000082
The corresponding normalized wavelet packet feature vector Xs1 *Is represented by Xs1 *＝[X1/X,X2/X,…,Xn/X]And forming a vector space X of the auxiliary source domain datas *＝[Xs1 *,Xs2 *,…Xsn *](ii) a Vector space X which similarly constitutes target domain datat *＝[Xt1 *,Xt2 *,…Xtn *]. Furthermore, m is often selected according to different fault conditions, and n is generally equal to m2-1, so that the number of layers of the wavelet decomposition can be dynamically adjusted according to the complexity of the fault, and λnSatisfies the constraint condition lambdan/λ1The singular value feature matrix obtained in the way can reflect fault information most simply.
Step 3, based on the feature mapping migration learning method, the method willThe intersection of the auxiliary source field and the target field is used as an axis feature, and a base vector corresponding to the axis feature, the fault feature of the auxiliary source field and the fault feature of the target field is found out based on a Lagrange multiplier method for solving an extreme value; further, based on the idea of feature mapping migration learning, the intersection of the auxiliary source field and the target field is used as an axis feature, and data of each field is mapped from the original high-dimensional feature space to the low-dimensional feature space, in the low-dimensional space, the source field data and the target field data have similar distribution, so that the source field data and the target field data can be abstracted to analyze the correlation between the characteristic features of the auxiliary source field and the characteristic features of the target field and the axis features, and the maximum value of the relation coefficient is solved by using a Lagrange multiplier method, so that three groups of basis vectors which can most reflect the fault category are found, specifically, the step 3 includes: step 31, defining auxiliary source field Xs *The fault identifier of the known fault type is Y, so that a certain fault type identifier Y belongs to Y; auxiliary source field Xs *And the target area Xt *The intersection of (A) is the corresponding axis feature or is called the domain axis feature X∩ *∈Xs *∩Xt *Simultaneous calculation of axial features X∩ *And the correlation coefficient between Y, the corresponding calculation formula is as follows:
Figure BDA0001392261310000091
wherein, P (X)∩ *Y) field axis feature X∩ *Joint distribution probability with fault sign y, correlation coefficient I (X)∩ *And y) the axis features with large values have stronger discriminability for fault types, so the axis features with the maximum correlation coefficient values in m-layer wavelet multi-resolution signal decomposition are selected to form an axis feature set which is recorded as X∩＝{X∩1 *,X∩2 *,…,X∩m *}; step 32, firstly, based on the union formed by the extracted fault features in the auxiliary source domain data and the target domain data
Figure BDA0001392261310000092
Three sets of paired sample sets of random variables α, γ, were constructed and labeled
Figure BDA0001392261310000093
Wherein | X∩|，
Figure BDA0001392261310000094
Respectively representing the dimensions of the axis features, the dimensions of the fault features of the auxiliary source domain data, the dimensions of the fault features of the target domain data, and
Figure BDA0001392261310000095
representing sample points X in auxiliary source domain datas *In the axial feature space X∩The value of (a) is selected from,
Figure BDA0001392261310000096
representing auxiliary source domain data sample points Xs *In a feature space
Figure BDA0001392261310000097
The value of (a) is selected from,
Figure BDA0001392261310000098
representing sample points in target domain data
Figure BDA0001392261310000099
In a feature space
Figure BDA00013922613100000910
And then linearly combining the values
Figure BDA00013922613100000911
Find three groups of base vectors by the principle that the correlation coefficient between the three groups reaches the maximum
Figure BDA00013922613100000912
Namely, based on the following formula
Figure BDA00013922613100000913
The constraint condition is
Figure BDA00013922613100000914
Wherein C isAA＝(AS∪At)(AS∪At)T
Figure BDA00013922613100000915
Figure BDA00013922613100000916
Figure BDA00013922613100000917
Figure BDA00013922613100000918
Figure BDA0001392261310000101
Figure BDA0001392261310000102
Figure BDA0001392261310000103
Step 33, finding out the basis vectors corresponding to the axis features, the auxiliary source field fault features, and the target field fault features based on the method for solving the extremum by Lagrange multiplier method, that is, finding out the basis vectors corresponding to the axis features, the auxiliary source field unique features, and the target field unique features based on the following formulas, where the source field unique features refer to the parts (axis features) of the source field features excluding the intersection between the source field and the target field, and the remaining features, and the target field unique features refer to the parts (axis features) of the target field features excluding the intersection between the source field and the target field, and the remaining features:
Figure BDA0001392261310000104
Figure BDA0001392261310000105
Figure BDA0001392261310000106
Figure BDA0001392261310000107
Figure BDA0001392261310000108
the eigenvectors corresponding to the first m generalized eigenvalues of the matrix are the base vectors WA，WS，WT。
Step 4, in the fault classification process based on the SVM, taking the base vector corresponding to the auxiliary source field obtained in the step 3 as a support vector; meanwhile, a similarity punishment item of a support vector training set corresponding to the auxiliary source field is added into an original target function of the support vector machine SVM, and a constraint condition of the added support vector training set is added into an original target function constraint condition, so that a classifier is trained together to obtain a corresponding classification result. Further, the step 4 comprises: step 41, in the fault classification process based on the SVM, firstly, the base vector W corresponding to the auxiliary source field obtained in the step 3 is usedSAs a support vector; meanwhile, adding a similarity penalty term of a support vector training set corresponding to the field of the auxiliary source into the original target function in the support vector machine SVM, and recording the similarity penalty term as
Figure BDA00013922613100001111
And in the original object boxAdding the constraint conditions of the added support vector training set into the number constraint conditions; then the support vector training set V containing the auxiliary source field data in the support vector machine SVMsThe optimization process of the training sample T is
Figure BDA0001392261310000111
Figure BDA0001392261310000112
Figure BDA0001392261310000113
Wherein
Figure BDA0001392261310000114
k is the number of training sets of target domain data,
Figure BDA0001392261310000115
is the support vector of the jth auxiliary source field data, DtIs the training data corresponding to the target domain data,
Figure BDA0001392261310000116
represents the distance of the jth support vector from the training data, if the smaller its value, then
Figure BDA0001392261310000117
The larger the value is, the greater the classification effect of the support vector on the target domain is, and gamma ist、γsRegularization coefficients for the target domain data and the auxiliary source domain data respectively,
Figure BDA0001392261310000118
Figure BDA0001392261310000119
the square term of the error function is used for replacing the original relaxation variable, so that the calculation can be simplified;
then, a Lagrange multiplier method is used for optimization, namely, in order to achieve the minimum loss function between the predicted value and the real category label, an SVM function estimation expression added with an auxiliary support vector set, namely an improved SVM function estimation expression is as follows:
Figure BDA00013922613100001110
wherein sgn represents a sign function, and if the number of the corresponding return value is greater than 0, sgn returns 1, if the number is equal to 0, then 0 is returned, and if the number is less than 0, then-1 is returned.
And step 42, multi-classification of the power grid fault can obtain corresponding classification results by constructing and combining a plurality of two classifiers. Further, the step 42 includes obtaining a corresponding classification result by using a decision binary tree method, for example, dividing all categories into two subclasses by using the decision binary tree method, each subclass further being divided into two subclasses, where a fault can be divided into a ground and a non-ground, and the ground is divided into a single-phase ground (a/b/c) and a two-phase ground (ab/ac/bc); and the method is divided into two-phase short circuit (ab/ac/bc), three-phase short circuit (abc) and the like without ground until the final class is divided.
The scheme of the invention is further illustrated below by taking the specific example as an example:
as shown in fig. 3 to 5, the specific steps in the power grid model are as follows:
setting parameters: as shown in fig. 4, the power grid model is a simplified 500kV double-end power supply transmission system with a total length of 200 km; the circuit model adopts a frequency correlation model to ensure that a calculation result obtained in the transient simulation is more accurate, and the model considers that signals with different frequencies have different attenuation degrees in the transmission process; in the case of power frequency, the positive sequence parameter is r1＝0.035W/km，x1＝0.424W/km，b1＝2.726×10-6S/km; zero sequence parameter is r0＝0.3W/km，x0＝1.143W/km，b0＝1.936×10-6S/km; meanwhile, A, B, C three-phase current data of 10 faults under the working conditions of different fault positions, different transition resistances and different fault moments are generated on the power grid model1089 groups in total as samples for fault classification, where AgFault 105 group, BgFault 145 group, CgFailed 90 group, ABgFault 95 group, BCgFailed 118 group, ACgFailure 102 group, AB failure 129 group, BC failure 109 group, AC failure 111 group, and ABC failure 85 group.
Step 2: respectively carrying out m-layer wavelet multi-resolution signal decomposition on target field data and auxiliary source field data to obtain a wavelet transformation coefficient matrix corresponding to a wavelet transformation result, and obtaining a singular value characteristic matrix corresponding to the wavelet transformation coefficient matrix after singular value decomposition calculation, wherein the singular value characteristic matrix is marked as lambda ═ diag (lambda ═ diag)1,λ2,…λn) (ii) a Taking 3-layer wavelet resolution decomposition of C-phase current in the field of auxiliary sources as an example, the singular value feature matrix is lambda ═ diag (lambda)1,λ2,…λ8) The singular characteristic values obtained after SVD conversion of the C-phase current signals under different types of faults are shown in table 1 (the bold data indicates fault data on the C-phase). As can be seen from Table 1, for the C-phase related faults, the 8 singular values are relatively averaged; and the fault data which is not related to the phase C is relatively uneven.
TABLE 1C-phase current singular diagonal matrix singular eigenvalues of each order
Figure BDA0001392261310000131
Taking A-phase single-phase grounding as an example, calculating the singular entropy of the micro-increment wavelet as
Figure BDA0001392261310000132
The same can be obtained
Figure BDA0001392261310000133
By analogy, X can be obtaineds1＝[X1,X2,…,X6,X7,X8]＝[2.198,0.341,-0.345,-0.187,-0.108,-0.196,-0.084,-0.056]，
Figure BDA0001392261310000141
Xs1 *＝[X1/X,X2/X,…,X8/X]＝[0.970,0.151,-0.152,-0.083,-0.047,-0.092,-0.003,-0.001]Similarly, X can be obtained when the B-phase single-phase short circuit occurss2 *The auxiliary source field vector space X can be obtained by the singular eigenvalues of 10 fault typess *＝[Xs1 *,Xs2 *,…Xs10 *]8×10And target domain vector space Xt *＝[Xt1 *,Xt2 *,…Xt10 *]8×10。
And step 3: based on the feature mapping migration learning method, three groups of basis vectors which can reflect the fault category most are found, and the method mainly comprises the following steps:
(1) obtaining an auxiliary source domain vector space Xs *＝[Xs1 *,Xs2 *,…Xs10 *]8×10And target domain vector space Xt *＝[Xt1 *,Xt2 *,…Xt10 *]8×10；
(2) From X∩ *∈Xs *∩Xt *Selecting the axis features with the maximum m phase relation values to form an axis feature set, and recording the axis feature set as X∩＝{X∩1 *,X∩2 *,…,X∩m *}；
(3) Construction of paired sample sets
Figure BDA0001392261310000142
(4) Then the eigenvector corresponding to the first m generalized eigenvalues of the above formula matrix is selected as the base vector WA，WS，WT(ii) a According to (1) - (4), the axial feature number is taken as 100, the projection vector dimension is 70, and the obtained basis vector projection result is shown in fig. 5:
and finally, adding a support vector training set to obtain a classification result, specifically, if the C phase is taken as a special phase at first, 1 represents that the phase is a fault phase, and 0 represents that the phase is a non-fault phase, table 2 lists part of training samples and coding conditions with the C phase as the special phase, and other conditions are similar.
TABLE 2C phase Fault codes for Special phases
Figure BDA0001392261310000143
Figure BDA0001392261310000151
The classification test results of the grid faults after the support vector training set is added are as follows: as can be seen from Table 3, after the support vector training set is added, various faults can be correctly identified, the average accuracy of fault classification can reach more than 99%, and the method is obviously improved compared with the method without the support vector training set.
TABLE 3 Fault Classification test results statistics
Figure BDA0001392261310000152
Table 4 shows that the fault classification method of the improved SVM after the support vector training set is added is basically not influenced by the fault time, the fault position and the transition resistance, and the misjudgment of the algorithm is possibly caused only when the high-resistance fault occurs at the tail end of the power transmission line through the discovery of the misjudgment samples.
TABLE 4 Fault Classification results under different conditions
Figure BDA0001392261310000153
Figure BDA0001392261310000161
In order to verify the adaptability of the fault classification method based on the transfer learning to the parameter change of the power grid line, the trained improved SVM model is used for testing 3 pieces of line fault sample data with different parameters, the line parameters are shown in a table 5, and the test result of each line is shown in a table 6. As can be seen from Table 6: the fault classification accuracy of different power transmission lines can reach more than 98% by using a fault classification method based on transfer learning, which shows that the fault classification method can be well adapted to the change of line parameters; meanwhile, the method can quickly realize the whole process of extracting the characteristics to classify the faults, the time required for classifying and identifying 1 sample data is less than 0.2 s, and the requirement of fault diagnosis on diagnosis time is met.
TABLE 5 parameters of 3 lines in the grid model
Figure BDA0001392261310000162
TABLE 6 Fault Classification results for different grid lines
Figure BDA0001392261310000163
Figure BDA0001392261310000171
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art should be considered to be within the technical scope of the present invention, and the technical solutions and the inventive concepts thereof according to the present invention should be equivalent or changed within the scope of the present invention.

Claims (6)

1. A power grid fault classification method based on feature mapping transfer learning is characterized by comprising the following steps:
step 1, target field data to be classified and auxiliary source field data are selected, wherein the target field data comprise: three-phase current data of each fault line at each fault moment; the auxiliary source domain data includes: three-phase current data of each fault line at the previous fault moment corresponding to each fault moment, three-phase current data of each fault line at the previous normal operation moment corresponding to each fault moment and three-phase current data of adjacent lines of the fault line at each fault moment;
step 2, fault feature extraction based on micro-increment wavelet singular entropy is respectively carried out on target field data and auxiliary source field data to extract respective corresponding micro-increment wavelet singular entropy, each micro-increment wavelet singular entropy is taken as a fault feature, and then a feature vector space corresponding to the target field and a feature vector space corresponding to the auxiliary source field are respectively formed;
step 3, based on a feature mapping migration learning method, finding out base vectors corresponding to the axis features, the characteristic features of the auxiliary source field and the characteristic features of the target field by using the intersection of the auxiliary source field and the target field as the axis features and solving an extreme value based on a Lagrange multiplier method;
step 4, in the fault classification process based on the SVM, taking the base vector corresponding to the auxiliary source field obtained in the step 3 as a support vector; meanwhile, a similarity punishment item of a support vector training set corresponding to the auxiliary source field is added into an original target function of the support vector machine SVM, and a constraint condition of the added support vector training set is added into an original target function constraint condition, so that a classifier is trained together to obtain a corresponding classification result.
2. The grid fault classification method according to claim 1, characterized in that:
the step 2 comprises the following steps:
step 21, respectively performing m-layer wavelet multi-resolution signal decomposition on the target field data and the auxiliary source field data to obtain a wavelet transformation coefficient matrix corresponding to a wavelet transformation result, and obtaining a singular value feature matrix corresponding to the wavelet transformation coefficient matrix after singular value decomposition calculation, and recording the singular value feature matrix as Λ ═ diag (λ ═ diag)1,λ2,…λn)；
Step 22, respectively constructing n-order micro-increment wavelet singular entropies of the target field data and the auxiliary source field data, wherein the corresponding formula is
Figure FDA0002149454780000011
In the formula, λiIs the i-th order non-zero singular eigenvalue, XiIs λiThe ith micro-increment wavelet singular entropy of (a);
step 23, constructing a feature vector X by using the n-order micro-increment wavelet singular entropy element of the auxiliary source field datas1Is marked as Xs1＝[X1,X2…Xn]Simultaneously order
Figure FDA0002149454780000021
The corresponding normalized wavelet packet feature vector Xs1 *Is represented by Xs1 *＝[X1/X,X2/X,…,Xn/X]And forming a feature vector space X of the auxiliary source field datas *＝[Xs1 *,Xs2 *,…Xsn *](ii) a Repeating the above steps to form a feature vector space X of the target field datat *＝[Xt1 *,Xt2 *,…Xtn *]。
3. The grid fault classification method according to claim 2, characterized in that:
n-m in the singular value feature matrix2-1 and such that λnAnd the constraint condition is satisfied.
4. The grid fault classification method according to claim 1, characterized in that:
the step 3 comprises the following steps:
step 31, defining auxiliary source field data Xs *The fault identifier of the known fault type is Y, so that a certain fault type identifier Y belongs to Y; auxiliary source domain data Xs *And target domain data Xt *The intersection of (A) is the corresponding axis feature or the domain axis feature, denoted as X∩ *∈Xs *∩Xt *Simultaneous calculation of axial features X∩ *And the correlation coefficient between Y, the corresponding calculation formula is as follows:
Figure FDA0002149454780000022
wherein, I (X)∩ *Y) represents an axial feature X∩ *Correlation coefficient with Y, P (X)∩ *Y) field axis feature X∩ *Joint distribution probability with fault identity y, P (X)∩ *) Representing axial characteristics X∩ *Data X appearing in the auxiliary Source Domains *P (y) indicates that the fault mark y appears in the target area data Xt *And selecting the axis feature with the maximum correlation coefficient value in m-layer wavelet multi-resolution signal decomposition to form an axis feature set, and recording as X∩＝{X∩1 *,X∩2 *,…,X∩m *}；
Step 32, firstly, based on the union formed by the extracted fault features in the auxiliary source domain data and the target domain data
Figure FDA0002149454780000023
Three groups of random variables α are set, and the number of gamma is nsThree sets of paired sample sets of random variables α, γ, were constructed and labeled
Figure FDA0002149454780000024
Wherein | X∩|，
Figure FDA0002149454780000025
Respectively representing the dimensions of the axis features, the dimensions of the fault features of the auxiliary source domain data, the dimensions of the fault features of the target domain data, and
Figure FDA0002149454780000026
representing sample points X in auxiliary source domain datas *In the axial feature space X∩The value of (a) is selected from,
Figure FDA0002149454780000027
representing auxiliary source domain data sample points Xs *In a feature space
Figure FDA0002149454780000031
The value of (a) is selected from,
Figure FDA0002149454780000032
representing sample points in target domain data
Figure FDA0002149454780000033
In a feature space
Figure FDA0002149454780000034
The value of (a) is selected from,
then linearly combining
Figure FDA0002149454780000035
Find three groups of base vectors by the principle that the correlation coefficient between the three groups reaches the maximum
Figure FDA0002149454780000036
Namely, based on the following formula
Figure FDA0002149454780000037
The constraint condition corresponding to the formula is
Figure FDA0002149454780000038
CAA＝(AS∪At)(AS∪At)T
Figure FDA0002149454780000039
Figure FDA00021494547800000310
CASST＝STASTT
Figure FDA00021494547800000311
Figure FDA00021494547800000312
Figure FDA00021494547800000313
Figure FDA00021494547800000314
Wherein: wAIs a set of basis vectors corresponding to the axial features; wSThe method comprises the following steps of (1) acquiring a base vector set corresponding to the characteristic features of the auxiliary source field; wTA base vector set corresponding to the characteristic features of the target field; cssIs a fault characteristic D in auxiliary source field datasA covariance matrix of medial axis features; a. theSIs | X about α∩|×nsA matrix of dimensions; a. thetIs | X about α∩|×ntMatrix of dimensions, S is about β
Figure FDA00021494547800000315
A matrix of dimensions, T being about β
Figure FDA00021494547800000316
A matrix of dimensions; cTTRefer to the failure characteristics D in the target domain datatA covariance matrix of medial axis features; cAAIs a fault characteristic D in auxiliary source field datasAnd fault signature D in target domain datatUnion D ofs∪DtA covariance matrix of medial axis features;
step 33, finding out the basis vectors corresponding to the axis features, the characteristic features of the auxiliary source field and the characteristic features of the target field based on the method of solving the extremum by the lagrange multiplier method, that is, finding out the basis vectors corresponding to the axis features, the characteristic features of the auxiliary source field and the characteristic features of the target field based on the following formulas:
Figure FDA0002149454780000041
Figure FDA0002149454780000042
Figure FDA0002149454780000043
Figure FDA0002149454780000044
Figure FDA0002149454780000045
the eigenvectors corresponding to the first m generalized eigenvalues of the matrix are the base vectors WA，WS，WT。
5. The grid fault classification method according to claim 1, characterized in that:
the step 4 comprises the following steps:
step 41, in the fault classification process based on the SVM, firstly, the base vector W corresponding to the auxiliary source field obtained in the step 3 is usedSAs a support vector; meanwhile, adding a similarity penalty term of a support vector training set corresponding to the field of the auxiliary source into the original target function in the support vector machine SVM, and recording the similarity penalty term as
Figure FDA0002149454780000046
Adding the constraint condition of the added support vector training set into the constraint condition of the original objective function; then the support vector training set V containing the auxiliary source field data in the support vector machine SVMsThe optimization process of the training sample T is
Figure FDA0002149454780000047
Wherein N istIs the number of i, Ns-NtIs the number of the j's,
Figure FDA0002149454780000051
k is the number of training sets of target domain data,
Figure FDA0002149454780000052
is the support vector of the jth auxiliary source field data, DtIs the training data corresponding to the target domain data,
Figure FDA0002149454780000053
representing the distance, γ, of the jth support vector from the training datat、γsRegularization coefficients for the target domain data and the auxiliary source domain data respectively,
Figure FDA0002149454780000054
is the squared term of the error function;
then, optimizing by using a Lagrange multiplier method, namely adding an SVM function estimation expression of an auxiliary support vector set when the loss function between the predicted value and the real category label is minimum, wherein the SVM function estimation expression is as follows:
Figure FDA0002149454780000055
and step 42, obtaining corresponding classification results by constructing and combining a plurality of two classifiers.
6. The grid fault classification method according to claim 5, characterized in that:
said step 42 comprises obtaining the corresponding classification result using a decision binary tree method.
CN201710756382.4A 2017-08-29 2017-08-29 Power grid fault classification method based on feature mapping transfer learning Active CN107491792B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710756382.4A CN107491792B (en) 2017-08-29 2017-08-29 Power grid fault classification method based on feature mapping transfer learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710756382.4A CN107491792B (en) 2017-08-29 2017-08-29 Power grid fault classification method based on feature mapping transfer learning

Publications (2)

Publication Number Publication Date
CN107491792A CN107491792A (en) 2017-12-19
CN107491792B true CN107491792B (en) 2020-04-07

Family

ID=60650761

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710756382.4A Active CN107491792B (en) 2017-08-29 2017-08-29 Power grid fault classification method based on feature mapping transfer learning

Country Status (1)

Country Link
CN (1) CN107491792B (en)

Families Citing this family (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108509850B (en) * 2018-02-24 2022-03-29 华南理工大学 Intrusion signal identification method based on distributed optical fiber system
CN108764915B (en) * 2018-04-26 2021-07-30 创新先进技术有限公司 Model training method, data type identification method and computer equipment
CN108805206A (en) * 2018-06-13 2018-11-13 南京工业大学 A kind of modified LSSVM method for building up for analog circuit fault classification
CN108798641B (en) * 2018-06-19 2021-06-11 东北大学 Rod pump pumping well fault diagnosis method based on subspace migration learning
CN108875918B (en) * 2018-08-14 2021-05-04 西安交通大学 Mechanical fault migration diagnosis method based on adaptive shared depth residual error network
CN109635837A (en) * 2018-11-10 2019-04-16 天津大学 A kind of carefree fall detection system of scene based on commercial wireless Wi-Fi
CN111600734B (en) * 2019-02-21 2021-11-02 烽火通信科技股份有限公司 Network fault processing model construction method, fault processing method and system
CN110365583B (en) * 2019-07-17 2020-05-22 南京航空航天大学 Symbol prediction method and system based on bridge domain transfer learning
CN110363763B (en) * 2019-07-23 2022-03-15 上饶师范学院 Image quality evaluation method and device, electronic equipment and readable storage medium
CN110736968B (en) * 2019-10-16 2021-10-08 清华大学 Radar abnormal state diagnosis method based on deep learning
CN110726958B (en) * 2019-11-05 2022-06-28 国网江苏省电力有限公司宜兴市供电分公司 Fault diagnosis method of dry-type reactor
CN112036301B (en) * 2020-08-31 2021-06-22 中国矿业大学 Driving motor fault diagnosis model construction method based on intra-class feature transfer learning and multi-source information fusion
CN112255500A (en) * 2020-10-12 2021-01-22 山东翰林科技有限公司 Power distribution network weak characteristic fault identification method based on transfer learning

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101672873A (en) * 2009-10-20 2010-03-17 哈尔滨工业大学 Detection method of transient harmonic signals of power system based on combination of Tsallis wavelet singular entropy and FFT computation
CN104361396A (en) * 2014-12-01 2015-02-18 中国矿业大学 Association rule transfer learning method based on Markov logic network
CN105469111A (en) * 2015-11-19 2016-04-06 浙江大学 Small sample set object classification method on basis of improved MFA and transfer learning
CN106841910A (en) * 2016-12-20 2017-06-13 国网辽宁省电力有限公司沈阳供电公司 Imitative electromagnetism algorithm is melted into the Fault Diagnosis Method for Distribution Networks of timing ambiguity Petri network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101672873A (en) * 2009-10-20 2010-03-17 哈尔滨工业大学 Detection method of transient harmonic signals of power system based on combination of Tsallis wavelet singular entropy and FFT computation
CN104361396A (en) * 2014-12-01 2015-02-18 中国矿业大学 Association rule transfer learning method based on Markov logic network
CN105469111A (en) * 2015-11-19 2016-04-06 浙江大学 Small sample set object classification method on basis of improved MFA and transfer learning
CN106841910A (en) * 2016-12-20 2017-06-13 国网辽宁省电力有限公司沈阳供电公司 Imitative electromagnetism algorithm is melted into the Fault Diagnosis Method for Distribution Networks of timing ambiguity Petri network

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
A Novel Protection Scheme against Fault Resistance for;Xinrui Liu等;《Mathematical Problems in Engineering》;20170202;第2017卷;第1-11页 *
基于多源信息的智能电网动态层次化故障诊断;刘鑫蕊等;《东北大学学报(自然科学版)》;20140915;第35卷(第9期);第1221-1224页 *
基于自组织特征映射网络的配电网故障类型识别;赵智等;《自动化技术与应用》;20120925;第31卷(第9期);第64-68页 *
迁移学习方法研究及其在跨领域数据分类中的应用;覃姜维;《中国博士学位论文全文数据库 信息科技辑》;20111215(第12期);I140-39 *

Also Published As

Publication number Publication date
CN107491792A (en) 2017-12-19

Similar Documents

Publication Publication Date Title
CN107491792B (en) Power grid fault classification method based on feature mapping transfer learning
CN103065160B (en) Based on the hyperspectral image classification method that the collaborative expression in local and neighborhood information retrain
CN103177265B (en) High-definition image classification method based on kernel function Yu sparse coding
CN103942749A (en) Hyperspectral ground feature classification method based on modified cluster hypothesis and semi-supervised extreme learning machine
CN105787259A (en) Method for analyzing influence correlation of multiple meteorological factors and load changes
CN111398811A (en) PMSM demagnetization fault diagnosis method based on terminal current cost sensitive learning
Bidoki et al. Comparison of several clustering methods in the case of electrical load curves classification
Li et al. Graph regularized non-negative matrix factorization by maximizing correntropy
CN109324595B (en) Industrial monitoring data classification method based on incremental PCA
Liu et al. Unsupervised object segmentation with a hybrid graph model (HGM)
Zarabie et al. Load profile based electricity consumer clustering using affinity propagation
CN117131022B (en) Heterogeneous data migration method of electric power information system
CN114117852A (en) Regional heat load rolling prediction method based on finite difference working domain division
CN113595071A (en) Transformer area user identification and voltage influence evaluation method
Xia et al. A Switch State Recognition Method based on Improved VGG19 network
CN109684749B (en) Photovoltaic power station equivalent modeling method considering operating characteristics
CN116845971A (en) Automatic identification method for topological structure of photovoltaic grid-connected low-voltage transformer area
CN116298699A (en) Power distribution network fault identification method based on wavelet optimization and graph convolution neural network
CN109886352A (en) A kind of unsupervised appraisal procedure of airspace complexity
Mounce Visualizing smart water meter dataset clustering with parametric t-distributed stochastic neighbour embedding
Roy et al. On some fast and robust classifiers for high dimension, low sample size data
CN111401519B (en) Deep neural network unsupervised learning method based on similarity distance in object and between objects
CN116538127B (en) Axial flow fan and control system thereof
Wu et al. A deep recognition network of capacitor voltage transformer based on dilated convolution and Bi-LSTM
Fansheng et al. Energy efficiency evaluation method based on multi-model fusion strategy

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant