CN114118248A - Power transformer fault diagnosis method based on width learning under unbalanced sample - Google Patents
Power transformer fault diagnosis method based on width learning under unbalanced sample Download PDFInfo
- Publication number
- CN114118248A CN114118248A CN202111385191.4A CN202111385191A CN114118248A CN 114118248 A CN114118248 A CN 114118248A CN 202111385191 A CN202111385191 A CN 202111385191A CN 114118248 A CN114118248 A CN 114118248A
- Authority
- CN
- China
- Prior art keywords
- sample
- stage
- classifier
- training
- samples
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 56
- 238000003745 diagnosis Methods 0.000 title claims abstract description 22
- 238000012549 training Methods 0.000 claims abstract description 53
- 230000008569 process Effects 0.000 claims abstract description 14
- 239000007789 gas Substances 0.000 claims description 36
- 230000004927 fusion Effects 0.000 claims description 11
- VNWKTOKETHGBQD-UHFFFAOYSA-N methane Chemical compound C VNWKTOKETHGBQD-UHFFFAOYSA-N 0.000 claims description 10
- 238000010891 electric arc Methods 0.000 claims description 8
- 238000010606 normalization Methods 0.000 claims description 7
- 238000005070 sampling Methods 0.000 claims description 7
- 230000006870 function Effects 0.000 claims description 6
- 238000004868 gas analysis Methods 0.000 claims description 6
- OTMSDBZUPAUEDD-UHFFFAOYSA-N Ethane Chemical compound CC OTMSDBZUPAUEDD-UHFFFAOYSA-N 0.000 claims description 5
- VGGSQFUCUMXWEO-UHFFFAOYSA-N Ethene Chemical compound C=C VGGSQFUCUMXWEO-UHFFFAOYSA-N 0.000 claims description 5
- 239000005977 Ethylene Substances 0.000 claims description 5
- HSFWRNGVRCDJHI-UHFFFAOYSA-N alpha-acetylene Natural products C#C HSFWRNGVRCDJHI-UHFFFAOYSA-N 0.000 claims description 5
- 125000002534 ethynyl group Chemical group [H]C#C* 0.000 claims description 5
- 238000000605 extraction Methods 0.000 claims description 5
- 239000001257 hydrogen Substances 0.000 claims description 5
- 229910052739 hydrogen Inorganic materials 0.000 claims description 5
- 125000004435 hydrogen atom Chemical class [H]* 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 238000007636 ensemble learning method Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 238000004422 calculation algorithm Methods 0.000 abstract description 6
- 230000008901 benefit Effects 0.000 abstract description 5
- 238000013528 artificial neural network Methods 0.000 abstract description 3
- 238000013135 deep learning Methods 0.000 abstract description 3
- 239000003921 oil Substances 0.000 description 8
- 238000010586 diagram Methods 0.000 description 5
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 239000010749 BS 2869 Class C1 Substances 0.000 description 1
- 229910006119 NiIn Inorganic materials 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 230000032683 aging Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000009413 insulation Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000010561 standard procedure Methods 0.000 description 1
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
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/20—Administration of product repair or maintenance
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Business, Economics & Management (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Biology (AREA)
- Economics (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Human Resources & Organizations (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Marketing (AREA)
- Health & Medical Sciences (AREA)
- Tourism & Hospitality (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Public Health (AREA)
- Primary Health Care (AREA)
- General Health & Medical Sciences (AREA)
- Water Supply & Treatment (AREA)
- Probability & Statistics with Applications (AREA)
- Entrepreneurship & Innovation (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Other Investigation Or Analysis Of Materials By Electrical Means (AREA)
Abstract
The invention discloses a method for diagnosing faults of a power transformer based on width learning under unbalanced samples, which comprises the steps of firstly, carrying out first-stage classification on the faults of the transformer by adopting a first-stage BLS classifier, and then carrying out second-stage classification on the faults by using a second-stage Softmax classifier on the basis of the first-stage classification so as to improve the diagnosis accuracy and efficiency of the faults; in the process of training the classifier, a width learning system (BLS) which uses an incremental algorithm of network horizontal expansion to construct a network model can effectively solve the problem of deep learning. Compared with a deep neural network, the method has the advantages of high precision and high training speed, and has the advantages of simple parameter setting and easiness in operation in practical application.
Description
Technical Field
The invention relates to the technical field of transformer fault diagnosis, in particular to a width learning-based power transformer fault diagnosis method under an unbalanced sample.
Background
The role of power transformers in power systems is irreplaceable, and the normal operation of the power transformers is closely related to the normal transmission of electric energy by the whole power grid system. In recent years, although the research and manufacturing technology of electrical equipment in China is advanced, the faults of power transformers with different sizes can be induced due to insulation aging, severe environment, too high operation load and the like, so that social and economic losses are caused, and serious accidents such as large-scale power failure, power grid collapse and the like can be caused even seriously.
A method for analyzing (DGA) dissolved gas in transformer oil is used for analyzing the content and gas production rate difference of characteristic gases such as hydrogen, methane, ethylene, acetylene, ethane and the like in insulating oil in different operating states, so that an important basis is provided for the evaluation of the operating state of a transformer, and the method has the advantage of supporting charged on-line detection, and is widely applied to the field of transformer state monitoring and fault diagnosis in China.
According to theoretical analysis and actual experience, researchers establish basic method systems with simple flows such as a three-ratio method, a Rogers ratio method and a David triangle method in the initial stage, but due to the limitation of problems such as code loss, threshold absolute and field investigation, the methods are gradually used as auxiliary means for transformer fault diagnosis. With the development of the machine learning theory, the fault diagnosis method based on artificial intelligence becomes a popular research direction in the field of science by virtue of higher classification accuracy of the fault diagnosis method on the operation state types of the transformer. However, as the model expands, the weight of deep learning is updated slowly, and the success rate of unsupervised and semi-supervised network classification is low.
Therefore, how to develop a fault diagnosis method suitable for a power transformer based on artificial intelligence becomes a problem to be solved urgently.
Disclosure of Invention
In view of the above, the invention provides a power transformer fault diagnosis method based on width learning under an unbalanced sample, so as to solve the problem of low diagnosis success rate of the conventional artificial intelligent fault diagnosis method.
The invention provides a technical scheme, in particular to a width learning-based power transformer fault diagnosis method under an unbalanced sample, which comprises the following steps:
s1: performing characteristic extraction on dissolved gas in the transformer oil based on a dissolved gas analysis method to obtain primary input data;
s2: inputting the primary input data into a trained primary BLS classifier to obtain a primary classification of the transformer state;
s3: and after feature fusion is carried out on the primary input data and the primary classification of the transformer state corresponding to the primary input data, inputting the primary input data and the primary classification of the transformer state into a trained second-stage Softmax classifier, and obtaining a second-stage classification of the transformer state.
Preferably, the first-level classification includes: normal, discharge fault, and overheat fault; the second level of classification includes: normal, partial discharge, spark discharge, arc discharge, spark discharge with superheat, arc discharge with superheat, low temperature superheat, medium temperature superheat, and high temperature superheat.
Further preferably, in step S1, the characteristic extraction is performed on the dissolved gas in the transformer oil based on a dissolved gas analysis method to obtain primary input data, specifically:
the volume fractions of 5 gases including hydrogen, methane, ethylene, acetylene and ethane in the transformer oil are obtained based on a dissolved gas analysis method, and acquired data are obtained
The collected data g are processediCarrying out normalization processing according to the following formula to obtain primary input data;
wherein, i is 1,2, …, G, indicating a sample number; g is the total number of samples; j-1, 2,3,4,5, each represent 5 gases.
Further preferably, the specific training process of the first-stage BLS classifier in step S2 is as follows:
s201: acquiring historical dissolved gas characteristics in transformer oil, and performing normalization processing to obtain sample characteristic data X;
s202: obtaining a first-stage classification Y of the transformer state corresponding to each sample characteristic data X;
s203: converting the sample characteristic data X into characteristic nodes according to the following formula;
Zi=φi(XWei+βei),i=1,2,…,n;
wherein, Wei∈RM×p,βei∈R1×pAnd W iseiWeight, β, representing characteristic node of i-th groupeiRepresenting the deviation of the ith group of feature nodes, M representing the dimension of sample feature data, p representing the number of each group of feature nodes, and n representing the group number of the feature nodes;
s204: generating an enhanced node by the characteristic node according to the following formula;
Hj=ξj(ZnWhj+βhj),j=1,2,3,…,m;
wherein Z isn=[Z1,Z2,…,Zn],Whj∈Rnp×q,βhj∈R1×qAnd W ishjWeight, β, representing the jth set of enhancement nodeshjExpressing the deviation of the jth group of enhanced nodes, wherein m expresses the group number of the enhanced nodes, and q expresses the number of the enhanced nodes in each group;
s205: splicing the characteristic nodes and the enhanced nodes to be used as an input layer, and calculating an output weight by adopting the following formula according to a first-stage classification Y of the transformer state corresponding to the sample characteristic data X to obtain a trained first-stage BLS classifier;
Wm=[Zn|Hm]+Y;
wherein Hm=[H1,H2,…,Hm]。
Further preferably, in step S3, the training process of the second-stage Softmax classifier specifically includes the following steps:
s301: performing feature fusion on the sample feature data X and a first class classification Y of the transformer state corresponding to the sample feature data X;
s302: separating the samples after the characteristics are fused to obtain a minority sample set and a majority sample set;
s303: obtaining a plurality of balance training subsets according to the minority sample set and the majority sample set;
s304: and inputting a plurality of balance training subsets into a second-stage Softmax classifier for training by adopting an easy Ensemble ensemble learning method to obtain the trained second-stage Softmax classifier.
More preferably, step S302: separating the sample after the characteristic fusion to obtain a minority sample and a majority sample, specifically:
s3021: acquiring the category of the unbalanced number set;
s3022: separating the samples after feature fusion according to the categories of the unbalanced number set, wherein the category with the least number of the corresponding samples in all the categories is a minority sample set P, and the others are a majority sample set N1,N2,…,NK-1。
More preferably, S303: obtaining a plurality of balance training subsets according to the minority sample set and the majority sample set, specifically:
s3031: for the majority sample set NiRandom independent undersampling with putting back is carried out for T times repeatedly, and a plurality of sample sets NiSeveral subsets N can be generated during each sampling titObtaining a plurality of sample subsets Nt={N1t,N2t,…,NK-1tIs and | NitI | ═ P |, where | Nit| is a subset N of majority class samplestThe number of samples in the sample set P, | P | the number of samples in the minority sample set P;
s3032: a plurality of sample subsets Nt={N1t,N2t,…,NK-1tA sum ofAnd combining the several sample sets P to obtain a plurality of balance training subsets.
More preferably, step S304: inputting the balance training subset into a second-stage Softmax classifier for training to obtain the trained second-stage Softmax classifier, which specifically comprises the following steps:
s3041: taking a plurality of balance data subsets as training samples, and respectively inputting the training samples into corresponding sub-classifiers for training;
s3042: and calculating the multi-classification probability of the transformer state by taking softmax as an activation function, and taking the second-level class code serial number k (k is 1,2,3,4,5,6,7,8) with the highest probability as a diagnosis result, wherein,
s3043: by passingCross entropy loss function of2Updating the parameters of the sub-classifier to obtain the trained sub-classifier, wherein Y2,tIs the output characteristic of the second stage;
s3044: the trained sub-classifier weight and the bias parameter are processed according to the preset weight value alphatMerge into the final second stage Softmax classifier H2Obtaining a second-stage Softmax classifier after training
According to the method for diagnosing the fault of the power transformer based on the width learning under the unbalanced sample, the fault of the transformer is firstly classified in a first stage by adopting a first-stage BLS classifier, and then classified in a second stage by adopting a second-stage Softmax classifier on the basis of the first-stage classification, so that the accuracy and the efficiency of fault diagnosis are improved.
In the process of training the classifier, a width learning system (BLS) for constructing a network model by using an incremental algorithm of network horizontal expansion can effectively solve the problem of deep learning. Compared with a deep neural network, the method has the advantages of high precision and high training speed, and has the advantages of simple parameter setting and easiness in operation in practical application.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and together with the description, serve to explain the principles of the invention.
In order to more clearly illustrate the embodiments or technical solutions in the prior art of the present invention, the drawings used in the description of the embodiments or prior art will be briefly described below, and it is obvious for those skilled in the art that other drawings can be obtained based on these drawings without creative efforts.
FIG. 1 is a schematic diagram of hierarchical classification provided by the disclosed embodiment of the present invention;
FIG. 2 is a DAG diagram of transformer states provided by the disclosed embodiment of the invention;
FIG. 3 is a block diagram of a first stage classifier width learning system (BLS) in accordance with a disclosed embodiment of the invention;
FIG. 4 is a schematic diagram of an Easyensemble algorithm provided by the disclosed embodiment of the invention;
FIG. 5 is a flow chart of a training process for providing a first-stage classifier and a second-stage classifier according to an embodiment of the present disclosure.
Detailed Description
Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. The following description refers to the accompanying drawings, in which like numerals refer to the same or similar elements throughout the different views, unless otherwise specified. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of methods consistent with certain aspects of the invention, as detailed in the appended claims.
The embodiment provides a width learning-based power transformer fault diagnosis method under an unbalanced sample, which comprises the following steps:
s1: performing characteristic extraction on dissolved gas in the transformer oil based on a dissolved gas analysis method to obtain primary input data;
s2: inputting the primary input data into a trained primary BLS classifier to obtain a primary classification of the transformer state;
s3: after feature fusion is carried out on the primary input data and the primary classification of the transformer state corresponding to the primary input data, the primary input data and the primary classification of the transformer state are input into a trained second-stage Softmax classifier, and second-stage classification of the transformer state is obtained;
in the diagnosis method, a first-stage BLS classifier is adopted to perform first-stage classification on the fault of the transformer, and a second-stage Softmax classifier is adopted to perform second-stage classification on the fault on the basis of the first-stage classification, and fig. 1 is a schematic diagram of specific hierarchical classification.
Referring to fig. 2, according to the transformer state type hierarchical classification structure, the state hierarchy can be respectively classified into 3 first-stage classifications Y of normal, discharge fault and overheat fault and a second-stage classification Y of 9 subdivided state types of normal, partial discharge, spark discharge, arc discharge, spark discharge and overheat, arc discharge and overheat, low-temperature overheat, medium-temperature overheat and high-temperature overheatz。
In the above embodiment, 5 kinds of gases of hydrogen, methane, ethylene, acetylene, and ethane are used as characteristic gases. In order to reduce the influence of the concentration of each characteristic gas on the absolute value fluctuation of a sample and ensure the convergence of a model network, a normalization method is adopted to combine the extreme value of the global characteristic gas, and a target interval is [0, 1 ] according to the maximum value and the minimum value of the characteristic gas per se]Scaling the value of (c). Under the condition of quantifying a certain characteristic gas distribution among samples and simultaneously keeping the relative size relation of the volume fractions of the component gases in the samples, each original gas is subjected toGas sample giThe treatment is carried out, and the treatment is carried out,i is 1,2, …, G. And the normalized gas data are recorded as:wherein
In the formula: i ═ 1,2, …, G, indicating the sample number; g is the total number of samples; j-1, 2,3,4,5, each represent 5 gases.
Because the number of cases among fault classes of the power transformer is unbalanced, if historical data is directly adopted as a sample to train a first-stage Softmax classifier and a second-stage Softmax classifier, the accuracy of fault diagnosis is low, therefore, the embodiment provides a width learning system (BLS) which uses an incremental algorithm of network transverse expansion to construct a network model aiming at the unbalance of the sample so as to solve the problem of the unbalance of the sample, and the specific training process of the specific classifier is as follows:
1. specific training procedure for first stage BLS classifier
Because in the first-stage classification of the power transformer, the balance degree of the sample can meet the training requirement, therefore, the sample data is collected and can be directly used for the training of the first-stage BLS classifier after being normalized, and the method specifically comprises the following steps:
s201: acquiring historical dissolved gas characteristics in transformer oil, and performing normalization processing to obtain sample characteristic data X;
as characteristic gases, 5 gases of hydrogen, methane, ethylene, acetylene, and ethane were used. In order to reduce the influence of the concentration of each characteristic gas on the absolute value fluctuation of a sample and ensure the convergence of a model network, a normalization method is adopted to combine the extreme values of the global characteristic gas, and a target interval of [0, 1 ] is carried out according to the maximum value and the minimum value of the characteristic gas per se]Reduction of numerical values ofAnd (4) placing. Under the condition of quantifying a certain characteristic gas distribution among samples and simultaneously keeping the relation of volume fractions and relative sizes of all component gases in the samples, g is taken as each original gas sampleiThe treatment is carried out, and the treatment is carried out,i is 1,2, …, G. And the normalized gas data are recorded as:wherein,
in the formula: i ═ 1,2, …, G, indicating the sample number; g is the total number of samples; j-1, 2,3,4,5, each represent 5 gases.
S202: obtaining a first-stage classification Y of the transformer state corresponding to each sample characteristic data X;
s203: converting the sample characteristic data X into characteristic nodes according to the following formula;
the normalized input data is set as { X, Y }. epsilon.RN×(M+C)Where X is sample feature data, Y is a first level of classification, N represents the number of samples, M represents the dimension of the input sample feature, and C represents the number of classes. Before converting the input sample features into feature nodes, parameters n and p need to be defined. Where n represents the number of sets of feature nodes and p represents p nodes per set of feature nodes. The process of converting the input sample features into feature nodes:
Zi=φi(XWei+βei),i=1,2,…,n
wherein X ∈ RN×M,Wei∈RM×p,βei∈R1×pAnd W isei,βeiGenerated in a random manner, representing the weights and biases used to generate the ith set of feature nodes, respectively. ZiRepresenting the ith set of generated feature nodes. All n groups of characteristic nodes are spliced together to obtain the mostFinal set of feature nodes ZnWherein Z isn=[Z1,Z2,…,Zn]。
S204: generating an enhanced node by the characteristic node according to the following formula;
similarly, the number of groups of enhancement nodes is defined as m, and the number of enhancement nodes in each group is defined as q. The process of generating the enhancement node from the feature node is as follows:
Hj=ξj(ZnWhj+βhj),j=1,2,3,…,m
wherein Z isn∈RN×np,Whj∈Rnp×q,βhj∈R1×q. And Whj,βhjGenerated in a random manner, representing the weights and biases, respectively, used to generate the jth set of enhancement nodes. HjRepresenting the characteristic nodes generated by the jth group. All m groups of feature nodes are spliced together to obtain a final feature node set HmIn which H ism=[H1,H2,…,Hm]。;
S205: splicing the characteristic nodes and the enhancement nodes to be used as an input layer, and obtaining a trained first-stage BLS classifier according to a first-stage classification Y of the transformer state corresponding to the sample characteristic data X;
after the enhanced nodes are generated, the feature nodes and the enhanced nodes need to be spliced together to be used as an input layer, and an output result is obtained through further calculation. Let us assume the weight between the connection input layer and output layer to be WmThe output Y of the first stage BLS classifier is then obtained, which is also the first stage classification, as shown in fig. 3.
Y=[Z1,Z2,…,Zn|H1…Hm]Wm
=[Zn|Hm]Wm
The pseudo-inverse algorithm can be considered as a convenient method to obtain a random weight flat neural network of output layer weights. Thus, the output layer weights can be quickly calculated as:
Wm=[Zn|Hm]+Y
however, due to the dimension and speed of the training data, it is too costly to use standard methods such as orthogonal projection, iteration, and singular value decomposition to compute the generalized inverse. The invention adopts another method to solve the pseudo-inverse.
Wherein σ1>0,σ2> 0, v and u are regularized norms. Under the condition of the original generalized inverse, the least square estimation A ═ Z of the original generalized inverse estimationn|Hm]And adding a constraint term lambda to solve the pseudo-inverse of the original generalized inverse estimation. When λ is 0, the inverse problem degenerates to the least squares problem and a solution to the original pseudo-inverse problem is obtained. Take sigma1=σ2And setting the optimal problem as a ridge regression learning algorithm, wherein v is 2, and u is 2. Ridge regression is a biased estimate of the regression parameter a, which results in larger residuals, but makes the model more general. Thus, the connection weight of the first stage BLS classifier may be approximated as:
Wm=(λI+AAT)-1ATY
then, there are:
and finally, obtaining the trained first-stage BLS classifier through the calculated weight.
2. Specific training process of second-stage Softmax classifier
Because there is a serious unbalanced sample problem in the second-stage classification of the power transformer, the training of the second classifier is performed after the acquired training samples need to be equalized, which is specifically as follows:
s301: performing feature fusion on the sample feature data X and a first class classification Y of the transformer state corresponding to the sample feature data X;
the original characteristics X can obtain three primary labels through a primary classifier, wherein the primary labels are respectively normal, discharge fault and overheat fault, and the codes are respectively 1,0 and 0 as shown in the following table 1; 0,1, 0; 0,0,1, and then performing feature fusion on the obtained first-stage classification Y and the original features X to serve as the input of a next-stage classifier.
TABLE 1
S302: separating the samples after the characteristics are fused to obtain a minority sample set and a majority sample set;
unbalanced dataset class C1,C2,…,C9The sample set with the least number of samples in the normal, partial discharge, spark discharge, arc discharge, low-temperature overheat, medium-temperature overheat, high-temperature overheat, spark discharge and overheat, and arc discharge and overheat is a least number of sample sets P, the number is | P |, and the other most sample sets are N1,N2,…,NK-1。
S303: obtaining a plurality of balance training subsets according to the minority sample set and the majority sample set;
repeating all majority sample set for T times with replaced random independent undersampling, majority sample set NiSeveral subsets N can be generated during each sampling titAnd satisfy | NitI ═ P |, where i ═ 1,2, …, K-1; t ═ 1,2, …, T, as shown in table 2.
Table 2:
in the above sampling process, most samplesThis NiArbitrary sample (x) of (a)ij,yij)∈NiIs acquired at least once to a subset NitThe probability of (1) - (1-1/| N)i|)|P|Probability L of occurring at least once in T subsetsoneAnd all occurrences of the possibility LallRespectively as follows:
under the condition of constant | P |, the larger the sampling frequency T is, the more the majority of samples (x)ij,yij) Likelihoods L occurring in the ensemble of training subsetsoneThe larger the sample is, so that most samples can be guaranteed to be sampled to a training set, and information loss in the undersampling process is prevented; meanwhile, there is a partial majority sample set sample number | NiIn the case that the number of samples | P | is not different from the number of samples | P | in the minority sample set by multiple times, the samples (x)ij,yij) Probability of occurrence in the ensemble of training subsets LallThe sampling quality of a small number of most samples can be guaranteed to a certain extent, compared with the sampling quality of a plurality of samples with larger difference multiples.
Then, a plurality of types of samples obtained in the process of each undersampling t are subjected to a subset Nt={N1t,N2t,…,NK-1tCombine P with the minority class sample set as training subset D of each sub-classifiert,Dt=NtAnd E, at the moment, the number of each category of the training subset is the same, namely the training subset is a balanced data set.
S304: and inputting a plurality of balance training subsets into a second-stage Softmax classifier for training by adopting an easy Ensemble ensemble learning method to obtain the trained second-stage Softmax classifier.
As shown in FIG. 4, when the easy Ensemble ensemble learning method is adopted, first, X is input to the second level2Constructing a data-balanced training subset X2,tThe fault type with small number of samples is used as a few class samples, the other classes are used as a majority class sample to carry out undersampling to obtain T training subsets, and each training subset can be recorded as
Dt=(X2,t,Y2,t),t=1,2,...,T
In the formula Y2,tIs the output characteristic of the second stage.
As shown in FIG. 5, T sub-classifiers h are trained in parallel2,tAnd calculating the multi-classification probability of the transformer state by taking softmax as an activation function, and taking the second-level class code serial number k (k is 1,2,3,4,5,6,7 and 8) with the highest probability as a diagnosis result.
after the parallel training of the sub-classifiers is finished, the weight and the bias parameter of the sub-classifiers are set according to the preset weight alphatMerge into the final second stage Softmax classifier H2The final output of the second stage Softmax classifier is:
other embodiments of the invention will be apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention following, in general, the principles of the invention and including such departures from the present disclosure as come within known or customary practice within the art to which the invention pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the invention being indicated by the following claims.
It is to be understood that the present invention is not limited to what has been described above, and that various modifications and changes may be made without departing from the scope thereof. The scope of the invention is limited only by the appended claims.
Claims (8)
1. A power transformer fault diagnosis method based on width learning under an unbalance sample is characterized by comprising the following steps:
s1: performing characteristic extraction on dissolved gas in the transformer oil based on a dissolved gas analysis method to obtain primary input data;
s2: inputting the primary input data into a trained primary BLS classifier to obtain a primary classification of the transformer state;
s3: and after feature fusion is carried out on the primary input data and the primary classification of the transformer state corresponding to the primary input data, inputting the primary input data and the primary classification of the transformer state into a trained second-stage Softmax classifier, and obtaining a second-stage classification of the transformer state.
2. The method for diagnosing the fault of the power transformer based on the width learning under the unbalance sample according to claim 1, wherein the first-stage classification comprises: normal, discharge fault, and overheat fault; the second level of classification includes: normal, partial discharge, spark discharge, arc discharge, spark discharge with superheat, arc discharge with superheat, low temperature superheat, medium temperature superheat, and high temperature superheat.
3. The method for diagnosing the fault of the power transformer based on the width learning under the unbalanced sample according to claim 1, wherein in step S1, the characteristic extraction is performed on the dissolved gas in the transformer oil based on the dissolved gas analysis method to obtain primary input data, specifically:
based on dissolved gas analysisThe method obtains the volume fractions of 5 gases including hydrogen, methane, ethylene, acetylene and ethane in the transformer oil to obtain the collected data
The collected data g are processediCarrying out normalization processing according to the following formula to obtain primary input data;
wherein, i is 1,2, …, G, indicating a sample number; g is the total number of samples; j-1, 2,3,4,5, each represent 5 gases.
4. The method for diagnosing the fault of the power transformer based on the width learning under the unbalanced sample as recited in claim 1, wherein the specific training process of the first-stage BLS classifier in the step S2 is as follows:
s201: acquiring historical dissolved gas characteristics in transformer oil, and performing normalization processing to obtain sample characteristic data X;
s202: obtaining a first-stage classification Y of the transformer state corresponding to each sample characteristic data X;
s203: converting the sample characteristic data X into characteristic nodes according to the following formula;
Zi=φi(XWei+βei),i=1,2,…,n;
wherein, Wei∈RM×p,βei∈R1×pAnd W iseiWeight, β, representing characteristic node of i-th groupeiExpressing the deviation of the ith group of feature nodes, M expressing the dimension of sample feature data, p expressing the number of each group of feature nodes, and n expressing the group number of the feature nodes;
s204: generating an enhanced node by the characteristic node according to the following formula;
Hj=ξj(ZnWhj+βhj),j=1,2,3,…,m;
wherein Z isn=[Z1,Z2,…,Zn],Whj∈Rnp×q,βhj∈R1×qAnd W ishjWeight, β, representing the jth set of enhancement nodeshjExpressing the deviation of the jth group of enhanced nodes, wherein m expresses the group number of the enhanced nodes, and q expresses the number of the enhanced nodes in each group;
s205: splicing the characteristic nodes and the enhanced nodes to be used as an input layer, and calculating an output weight by adopting the following formula according to a first-stage classification Y of the transformer state corresponding to the sample characteristic data X to obtain a trained first-stage BLS classifier;
Wm=[Zn|Hm]+Y;
wherein Hm=[H1,H2,…,Hm]。
5. The method for diagnosing the fault of the power transformer based on the width learning under the unbalance sample according to claim 4, wherein the training process of the second-stage Softmax classifier in the step S3 is as follows:
s301: performing feature fusion on the sample feature data X and the first-stage classification Y of the transformer state corresponding to the sample feature data X;
s302: separating the samples after the characteristics are fused to obtain a minority sample set and a majority sample set;
s303: obtaining a plurality of balance training subsets according to the minority sample set and the majority sample set;
s304: and inputting a plurality of balance training subsets into a second-stage Softmax classifier for training by adopting an easy Ensemble ensemble learning method to obtain the trained second-stage Softmax classifier.
6. The method for diagnosing the fault of the power transformer based on the width learning under the unbalance sample according to claim 5, wherein the step S302: separating the sample after the characteristic fusion to obtain a minority sample and a majority sample, specifically:
s3021: acquiring the category of the unbalanced number set;
s3022: separating the samples after feature fusion according to the categories of the unbalanced number set, wherein the category with the least number of the corresponding samples in all the categories is a minority sample set P, and the others are a majority sample set N1,N2,…,NK-1。
7. The method for diagnosing the fault of the power transformer based on the width learning under the unbalance sample according to claim 5, wherein S303: obtaining a plurality of balance training subsets according to the minority sample set and the majority sample set, specifically:
s3031: for the majority sample set NiPerforming repeated T times of random independent undersampling with release, and collecting multiple types of samples NiSeveral subsets N can be generated during each sampling titObtaining a plurality of sample subsets Nt={N1t,N2t,…,NK-1tIs and | NitI | ═ P |, where | Nit| is a subset N of majority class samplestThe number of samples in the sample set P, | P | the number of samples in the minority sample set P;
s3032: a plurality of sample subsets Nt={N1t,N2t,…,NK-1tAnd respectively combining the training data with the minority class sample sets P to obtain a plurality of balanced training subsets.
8. The method for diagnosing the fault of the power transformer based on the width learning under the unbalance sample according to claim 5, wherein the step S304: inputting the balance training subset into a second-stage Softmax classifier for training to obtain the trained second-stage Softmax classifier, which specifically comprises the following steps:
s3041: taking a plurality of balanced data subsets as training samples, and respectively inputting the training samples into corresponding sub-classifiers for training;
s3042: calculating the multi-classification probability of the transformer state by taking softmax as an activation function, and maximizing the probabilityAs a diagnosis result, the second-level class code number k (k ═ 1,2,3,4,5,6,7,8) of (a),
s3043: by passingCross entropy loss function of2Updating the parameters of the sub-classifier to obtain the trained sub-classifier, wherein Y2,tIs the output characteristic of the second stage;
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111385191.4A CN114118248A (en) | 2021-11-22 | 2021-11-22 | Power transformer fault diagnosis method based on width learning under unbalanced sample |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111385191.4A CN114118248A (en) | 2021-11-22 | 2021-11-22 | Power transformer fault diagnosis method based on width learning under unbalanced sample |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114118248A true CN114118248A (en) | 2022-03-01 |
Family
ID=80439426
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111385191.4A Pending CN114118248A (en) | 2021-11-22 | 2021-11-22 | Power transformer fault diagnosis method based on width learning under unbalanced sample |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114118248A (en) |
-
2021
- 2021-11-22 CN CN202111385191.4A patent/CN114118248A/en active Pending
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110162018B (en) | Incremental equipment fault diagnosis method based on knowledge distillation and hidden layer sharing | |
CN110175386B (en) | Method for predicting temperature of electrical equipment of transformer substation | |
CN112116058B (en) | Transformer fault diagnosis method for optimizing multi-granularity cascade forest model based on particle swarm algorithm | |
CN109978079A (en) | A kind of data cleaning method of improved storehouse noise reduction self-encoding encoder | |
CN110879373B (en) | Oil-immersed transformer fault diagnosis method with neural network and decision fusion | |
CN114167180B (en) | Oil-filled electrical equipment fault diagnosis method based on graph attention neural network | |
CN112289391B (en) | Anode aluminum foil performance prediction system based on machine learning | |
CN114219147A (en) | Power distribution station fault prediction method based on federal learning | |
CN113505655A (en) | Bearing fault intelligent diagnosis method for digital twin system | |
CN112147432A (en) | BiLSTM module based on attention mechanism, transformer state diagnosis method and system | |
CN113343581B (en) | Transformer fault diagnosis method based on graph Markov neural network | |
CN115563563A (en) | Fault diagnosis method and device based on transformer oil chromatographic analysis | |
CN111737907A (en) | Transformer fault diagnosis method and device based on deep learning and DGA | |
CN116562114A (en) | Power transformer fault diagnosis method based on graph convolution neural network | |
CN114661905A (en) | Power grid fault diagnosis method based on BERT | |
CN111091141B (en) | Photovoltaic backboard fault diagnosis method based on layered Softmax | |
CN115146739A (en) | Power transformer fault diagnosis method based on stacked time series network | |
Yin et al. | Deep learning based feature reduction for power system transient stability assessment | |
Jian et al. | Transformer fault diagnosis model with unbalanced samples based on SMOTE algorithm and focal loss | |
CN114581699A (en) | Transformer state evaluation method based on deep learning model in consideration of multi-source information | |
CN117473424A (en) | Transformer fault diagnosis method, system, equipment and medium based on random forest | |
CN117150383A (en) | New energy automobile power battery fault classification method of SheffleDarkNet 37-SE | |
CN115659258B (en) | Power distribution network fault detection method based on multi-scale graph roll-up twin network | |
CN110348489B (en) | Transformer partial discharge mode identification method based on self-coding network | |
CN116760015A (en) | Large power grid transient power angle stability evaluation method based on high-dimensional graph annotation force network |
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 |