CN115201640A - FI-CNN-based transformer fault diagnosis method - Google Patents
FI-CNN-based transformer fault diagnosis method Download PDFInfo
- Publication number
- CN115201640A CN115201640A CN202210703453.5A CN202210703453A CN115201640A CN 115201640 A CN115201640 A CN 115201640A CN 202210703453 A CN202210703453 A CN 202210703453A CN 115201640 A CN115201640 A CN 115201640A
- Authority
- CN
- China
- Prior art keywords
- data
- transformer
- fault diagnosis
- spatial information
- cnn
- 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
- 238000003745 diagnosis Methods 0.000 title claims abstract description 61
- 238000000034 method Methods 0.000 title claims abstract description 38
- 238000012549 training Methods 0.000 claims abstract description 47
- 239000007789 gas Substances 0.000 claims abstract description 25
- 238000010276 construction Methods 0.000 claims abstract description 7
- 230000001502 supplementing effect Effects 0.000 claims abstract description 5
- 238000012360 testing method Methods 0.000 claims description 21
- 238000000605 extraction Methods 0.000 claims description 11
- 230000006870 function Effects 0.000 claims description 8
- 238000012545 processing Methods 0.000 claims description 5
- 238000007405 data analysis Methods 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 abstract description 38
- 238000013528 artificial neural network Methods 0.000 abstract description 12
- 230000008569 process Effects 0.000 abstract description 8
- 230000000694 effects Effects 0.000 abstract description 7
- 239000000284 extract Substances 0.000 abstract description 5
- 238000007781 pre-processing Methods 0.000 abstract description 3
- 238000004422 calculation algorithm Methods 0.000 description 13
- 238000012706 support-vector machine Methods 0.000 description 9
- 230000003068 static effect Effects 0.000 description 7
- 239000003086 colorant Substances 0.000 description 6
- 238000013021 overheating Methods 0.000 description 6
- 238000004458 analytical method Methods 0.000 description 5
- 238000009826 distribution Methods 0.000 description 5
- 239000011159 matrix material Substances 0.000 description 5
- 238000013473 artificial intelligence Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- VNWKTOKETHGBQD-UHFFFAOYSA-N methane Chemical compound C VNWKTOKETHGBQD-UHFFFAOYSA-N 0.000 description 3
- 238000003062 neural network model Methods 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000004868 gas analysis Methods 0.000 description 2
- 229910052739 hydrogen Inorganic materials 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000009413 insulation Methods 0.000 description 2
- 230000007774 longterm Effects 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 101100136092 Drosophila melanogaster peng gene Proteins 0.000 description 1
- OTMSDBZUPAUEDD-UHFFFAOYSA-N Ethane Chemical compound CC OTMSDBZUPAUEDD-UHFFFAOYSA-N 0.000 description 1
- VGGSQFUCUMXWEO-UHFFFAOYSA-N Ethene Chemical compound C=C VGGSQFUCUMXWEO-UHFFFAOYSA-N 0.000 description 1
- 239000005977 Ethylene Substances 0.000 description 1
- 241001620634 Roger Species 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 239000000654 additive Substances 0.000 description 1
- 230000000996 additive effect Effects 0.000 description 1
- 230000032683 aging Effects 0.000 description 1
- HSFWRNGVRCDJHI-UHFFFAOYSA-N alpha-acetylene Natural products C#C HSFWRNGVRCDJHI-UHFFFAOYSA-N 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000004883 computer application Methods 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 230000001419 dependent effect Effects 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
- 230000002708 enhancing effect Effects 0.000 description 1
- 125000002534 ethynyl group Chemical group [H]C#C* 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000002068 genetic effect Effects 0.000 description 1
- 239000001257 hydrogen Substances 0.000 description 1
- 125000004435 hydrogen atom Chemical class [H]* 0.000 description 1
- 239000003112 inhibitor Substances 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000011089 mechanical engineering Methods 0.000 description 1
- 239000002245 particle Substances 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000002787 reinforcement Effects 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 230000008646 thermal stress Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/12—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing
- G01R31/1227—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing of components, parts or materials
- G01R31/1263—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing of components, parts or materials of solid or fluid materials, e.g. insulation films, bulk material; of semiconductors or LV electronic components or parts; of cable, line or wire insulation
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
- G01N33/26—Oils; Viscous liquids; Paints; Inks
- G01N33/28—Oils, i.e. hydrocarbon liquids
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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)
- Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Chemical & Material Sciences (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Economics (AREA)
- Data Mining & Analysis (AREA)
- Strategic Management (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Public Health (AREA)
- Water Supply & Treatment (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- Evolutionary Computation (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Artificial Intelligence (AREA)
- Chemical Kinetics & Catalysis (AREA)
- General Chemical & Material Sciences (AREA)
- Oil, Petroleum & Natural Gas (AREA)
- Food Science & Technology (AREA)
- Medicinal Chemistry (AREA)
- Analytical Chemistry (AREA)
- Biochemistry (AREA)
- Immunology (AREA)
- Pathology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a transformer fault diagnosis method based on FI-CNN, which analyzes the relation between main faults and characteristic quantities of a transformer, sorts and analyzes according to influence weights, and extracts the content of various dissolved gases in transformer oil; selecting 3 gases from the dissolved gases in the transformer oil, wherein the concentration ratios of the 3 gases are expressed in an RGB (Red Green blue) form, constructing by using DGA (differential global positioning System) data of transformer faults, configuring a corresponding spatial information map for each fault data, supplementing more characteristic information and completing the construction of a characteristic data block; and constructing and training a transformer fault diagnosis model to realize transformer fault diagnosis. The method extracts the feature data of the fault sample as much as possible in the data preprocessing process, introduces more reference feature quantities, combines the obvious advantages of RGB (red, green and blue) on color display, enables the information contained in the data to be more comprehensive, can more effectively extract the information by using a neural network, finds a more definite high-dimensional classification curve on a two-dimensional coordinate axis, and achieves a better fitting effect.
Description
Technical Field
The invention relates to the technical field of transformer fault diagnosis, in particular to a FI-CNN-based transformer fault diagnosis method.
Background
The power transformer is one of the key devices for safe and reliable operation of power systems, and undertakes the conversion of voltage and the distribution of electric energy, and the components of the power transformer comprise insulating oil and insulating paper. However, due to design, manufacture, transportation, installation, human error, equipment aging, lightning strike and the like, insulation of the transformer is deteriorated, and normal power supply is affected. Effective condition monitoring and fault diagnosis techniques are fundamental measures for maintaining and enhancing reliability of power systems, which may cause abnormal decomposition of transformer oil or degradation of paper insulation, etc. if an early fault occurs due to electrical or thermal stress, thereby generating various gases such as methane (CH) 4 ) Ethylene (C) 2 H 4 ) Acetylene (C) 2 H 2 ) Ethane (C) 2 H 6 ) Hydrogen (H) 2 ) Etc., which are dissolved in the transformer oil; therefore, condition monitoring and fault diagnosis of the transformer are closely related to these gases.
In transformer fault diagnosis, because the frequency of occurrence of faults of a large-scale transformer which is actually in operation is low, the Analysis of Gas in oil (disasolved Gas Analysis DGA) data generated by different faults has large difference and coding loss in magnitude and data distribution, so that the accuracy of transformer fault diagnosis is low.
To date, many DGA evaluation methods, such as the Doernenburg ratio method, the key gas analysis method, the roger ratio method, the IEC three ratio method, and the method of dewar triangle, have appeared, but these methods mostly depend on the personal experience of experts, and inevitably produce different conclusions, or even fail to obtain results. Therefore, accurate DGA analysis has been one of the issues that has been widely focused on in transformer fault diagnosis.
With the development of machine learning, deep learning, reinforcement learning techniques and computer hardware, artificial intelligence techniques are gradually applied to fault diagnosis of transformers with their high accuracy and generalization capability. To solve such problems, a variety of artificial intelligence techniques are applied to this field, including artificial neural networks, support vector machines, fuzzy logic algorithms, bayesian networks, and hybrid systems of these techniques, with good results. However, the artificial neural network has the disadvantages of slow convergence rate and overfitting; the Bayesian network needs a large amount of sample training to obtain a good diagnosis effect; the support vector machine has great advantages in solving small samples, nonlinear and high-dimensional pattern recognition and strong generalization capability, but the kernel function of the support vector machine meets the Mercer condition, and a punishment coefficient C needs to be set through cross validation, so that the difficulty is high. The specific examples are as follows:
documents 1 to 3 provide a Support Vector Machine (SVM) -based transformer fault diagnosis model, and a fault boundary hyperplane is searched by using an SVM model to perform data classification.
[ document 1] Han Shijun, zhu Ju, mao Jigui, etc. transformer fault diagnosis based on particle swarm optimization support vector machine [ J ]. Electric measurement and instrumentation, 2014,51 (11): 71-75+90.
[ document 2] Tsuntan, gao Wensheng, zhang Ziwei, and the like [ J ] Chihua university journal (Nature science edition), 2018,58 (07): 623-629 ] for diagnosis of transformer failure based on a support vector machine and a genetic algorithm.
[ document 3] Hao Lingling, zhu Yongli. Transformer fault diagnosis based on DCAE-SVM [ J ]. Computer applications and software, 2021,38 (05): 49-53+87.
Documents 4 to 5 construct a membership function of the measure of the fault under the fault by using a fuzzy logic algorithm, and judge the fault type according to a maximum membership principle:
[ document 4] Bhalla D, bansal R K, gupta H O. Transform inhibitor based on DGA using fuzzy logic, 2C// India International Conference on Power Electronics, IICPE, 2011.
Document 5 Liu Kai, peng Weijie, yang Xuejun application of feature optimization and fuzzy theory in transformer fault diagnosis [ J ] power system protection and control, 2016,44 (15): 54-60.
Documents 6 to 7 and the like utilize a division space with a membership function fuzzification three-ratio value, and a fuzzy Bayesian network is used for reasoning to obtain a fault type.
[ document 6] Wang Yongjiang, law square, li and ming. Power transformer fault diagnosis method based on rough set theory and bayesian network [ J ] chinese electro-mechanical engineering press, 2006, 26 (8): 137-141.
Although the above schemes have good results in practical use, in the use process of these models, the training classification is performed by using the training set corpus as the object, and the objective function takes into account the whole sample, so that the error recognition of a few classes of objects in the sample has little influence on the whole error of the model, and the recognition capability of the model on the error is weak.
A Convolutional Neural Network (CNN) is a neural network that is used specifically to process data having a grid-like structure. Convolutional networks refer to neural networks that use convolution operations in at least one layer of the network instead of the usual matrix multiplication operations. The number of convolution kernels in a Convolutional Neural Network (CNN) is the feature extraction amount of the neural network model, and the extracted features of the neural network model are also called a feature map. When the convolution kernel carries out a forward propagation process, the input variables of the neural network model are processed, and then the unit node matrix output by the output layer is obtained through solving.
[ document 8] Xianhong Jiang, guo Gongbing, showjin super [ CNN-based power transformer fault diagnosis method [ J ]. Electronic design engineering, 2020,28, (13): 189-193).
[ document 9] Wang Feng, bi Jiangang, mo Zicong, yan danfeng, transformer fault diagnosis method based on deep convolutional neural network [ J ]. Guangdong power, 2019,32 (09): 177-183.
In a word, in transformer fault diagnosis, a simple Convolutional Neural Network (CNN) preprocesses fault basic information and is constructed only in a simple coding mode, the construction of a feature vector is rough, and the boundary division of fault features and faults is fuzzy, so that the data neural network fitting effect of a fault sample is poor; the traditional BP neural network is very sensitive to initial network weight, the network is initialized by different weights, the network is often converged to different local minimum, and the defect that the expression capability is insufficient and the fitting is easy to carry out exists when the nonlinear fitting of the BP neural network is used for learning the nonlinear relation between fault data and fault categories to carry out fault diagnosis.
Interpretation of terms:
RGB: RGB color model red, green and blue color model
FI: feature Image
CNN: convolutional neural network
Disclosure of Invention
In view of the above problems, an object of the present invention is to provide a transformer fault diagnosis method based on FI-CNN, which extracts feature data of a fault sample as much as possible in a data preprocessing process, introduces more reference feature quantities, constructs a euclidean two-dimensional coordinate axis, and combines the obvious advantages of RGB in color display, so that information contained in the data is more comprehensive, a neural network can be used more effectively for information extraction, and a more definite high-dimensional classification curve is found on the two-dimensional coordinate axis, thereby achieving a better fitting effect. The technical scheme is as follows:
a transformer fault diagnosis method based on FI-CNN comprises the following steps:
step 1: analyzing the relation between the main faults and the characteristic quantity of the transformer, sorting and analyzing according to the influence weight, and extracting the content of various dissolved gases in the transformer oil;
step 2: selecting 3 gases from the dissolved gases in the transformer oil, wherein the concentration ratios of the 3 gases are expressed in an RGB (Red Green blue) form, constructing by using DGA (differential global positioning System) data of transformer faults, configuring a corresponding spatial information map for each fault data, supplementing more characteristic information and completing the construction of a characteristic data block;
and step 3: considering the structure of the characteristic data block, constructing a transformer fault diagnosis model based on a CNN convolutional network;
and 4, step 4: and training the transformer fault diagnosis model based on the CNN convolutional network to realize transformer fault diagnosis.
Further, the feature data block is specifically constructed in the following manner:
step 2.1: dividing existing transformer fault DGA data into a training set and a test set in proportion, and setting the number of data contained in the training set and the test set to be M and N respectively;
step 2.2: creating a zero array of 100 × 3, namely an RGB graph with the length and width of 100 and the number of layers of 3, wherein each element is 0;
step 2.3: taking the first data A in the training set, and filling the spatial information map with the rest M-1 DGA data; the filling rule is as follows: for M-1 data, with H for each data 2 Concentration x 0 X-axis coordinate as Euclidean plane, CH 4 Concentration y 0 As y-axis coordinate, C 2 H 6 、C 2 H 4 、C 2 H 2 The concentrations are respectively used as R, G, B parameters and are filled in a spatial information graph to obtain an RGB graph containing not more than M-1 nonzero pixel points;
step 2.4: repeating the steps 2.2-2.3 for the residual data in the training set, and finally obtaining a total of M different spatial information graphs without repetition;
step 2.5: taking one data in the training set and a corresponding spatial information graph as a feature data block, wherein the total number of the feature data blocks is M; thus, obtaining a characteristic data block of the training data set;
step 2.6: filling M training data into a spatial information graph to obtain a reference graph in the mode of the step 2.2-2.3;
step 2.7: taking one data in the test set and the reference image as a characteristic data block, wherein N different characteristic data blocks are counted; thereby, a feature data block of the test data set is obtained.
Further, the transformer fault diagnosis model comprises four layers:
one layer is a convolution layer part, the size of an input interface of the convolution layer part is 100 × 3, and the part is used for carrying out feature identification and feature extraction on the spatial information graph;
the second layer is a flat layer and is used for processing the two-dimensional array output by the convolution layer part into a one-dimensional array;
the three layers are full connection layers, the first layer is also a gathering layer and is used for gathering output data of the flat layer and fault DGA data in sequence as input of the full connection layer part; the rest layers of the full connection layer are used for data analysis and feature extraction;
the four layers are output layers, the input of the output layers is the output of the last layer of the full connection layer, the output dimensionality of the output layers is equal to the fault type of the transformer, and normalization processing is carried out through a Softmax function.
Furthermore, the transformer fault diagnosis model operates as follows:
inputting the spatial information graph into a convolutional layer as a part of characteristic data, after data fitting is carried out, accurately finding a high-dimensional curved surface of the partitioned data in the spatial information graph through learning by a CNN convolutional network, wherein the output of the CNN convolutional network is the boundary curved surface of the spatial information graph or high-dimensional representation thereof;
and simultaneously inputting the curved surface or the high-dimensional representation thereof and the corresponding fault DGA data into a full-link network for further fitting, training the full-link network to classify the fault DGA data according to the boundary curved surface, and finally obtaining correct classified data.
Compared with the prior art, the invention has the beneficial effects that:
1. the method analyzes the relation between the main faults of the transformer and the characteristic quantity, sorts and analyzes according to the influence weight, and extracts the salient factors.
2. The invention selects the concentration proportion of 3 gases to be expressed in an RGB mode, utilizes DGA data of transformer faults to construct, and configures a corresponding spatial information graph containing spatial information for each fault data, thereby supplementing more characteristic information and being beneficial to improving the fault diagnosis accuracy of the transformer.
3. The invention adopts a dynamic RGB image to divide a more reasonable boundary curved surface in the spatial information image, and effectively performs data expansion on the basis of the existing data set, thereby realizing that the model utilizes new fault DGA data to perform continuous learning and continuous updating in long-term application.
Drawings
FIG. 1 shows a RGB three-dimensional neuron structure
Fig. 2 is an RGB sample graph of failure data.
Fig. 3 is a graph of a convoluted two-dimensional plot.
Fig. 4 is a transformer fault diagnosis model structure.
Fig. 5 is a static RGB diagram of a sample.
Fig. 6 is a dynamic RGB diagram of a sample.
FIG. 7 is a comparison of the results of the dynamic and static models.
FIG. 8 is a comparison of diagnostic accuracy for different algorithms.
FIG. 9 shows the diagnostic accuracy of different models with different samples.
Detailed Description
The invention is described in further detail below with reference to the figures and specific embodiments.
1. Feature data selection and construction
The extraction of the prominent influencing factors is the basis for realizing the fault diagnosis of the power transformer by applying artificial intelligence. And analyzing the relation between the main fault and the characteristic quantity of the transformer, sequencing and analyzing according to the influence weight, and extracting the salient factors. The DGA data for a transformer fault has the following characteristics:
(1) In the event of medium-low temperature failure, the main component of the gas dissolved in the oil is CH 4 And C 2 H 4 ，C 2 H 6 And H 2 The content is relatively small.
(2) In the event of high temperature failure, the main component of the dissolved gas in the oil is CH 4 And C 2 H 2 。
(3) In the event of discharge failure, the main component of the dissolved gas in the oil is H 2 And C 2 H 2 ，CH 4 And C 2 H 4 ，C 2 H 6 The content is relatively small.
When the transformer has discharge fault and overheat fault, H 2 And CH 4 There is a large difference in concentration, therefore, if H is added 2 And CH 4 The concentration of the discharge fault and the overheating fault are respectively taken as x-axis and y-axis parameters of the Euclidean plane, so that the distribution of the data points of the discharge fault and the overheating fault on the plane is also obviously different.
A three primary color light mode (RGB color model), also called RGB color model or Red-Green-Blue color model, is an additive color model, which adds the color lights of the three primary colors of Red (Red), green (Green) and Blue (Blue) in different proportions to synthesize and generate various color lights. For a picture in RGB format, the input to CNN is a three-dimensional neuron (a matrix for each color channel in RGB), as shown in fig. 1.
Since RGB combines 3 colors of different intensities to represent almost completely the colors that can be perceived by humans. Therefore, when the concentration ratios of the 3 kinds of gases are expressed in RGB format, it is considered that different subclass failures have different "colors" and that subclass failure diagnosis can be performed by distinguishing the "colors".
As shown in fig. 2, it can be observed that the data points of the discharge fault are relatively concentrated in the lower right region of the plane, and the data points of the overheating fault are relatively concentrated in the upper left region of the plane, according to the apparent color difference of the image. Meanwhile, under the same type of fault, the concentrations of the remaining 3 gases may also have certain differences according to the corresponding RGB colors due to specific subclass faults (low-temperature overheating, medium-high temperature overheating, high-temperature overheating, low-energy discharge, high-energy discharge, partial discharge).
When the spatial topology of the distribution network is considered, namely the adjacency matrix is introduced, the neural network can better identify the correlation property between the data. Therefore, for a certain data set, if the relative position of each piece of data in the whole data set is considered, more reference characteristic quantities are introduced equivalently, so that the information contained in the data is more comprehensive, the neural network can be used for extracting the information more effectively, and a better fitting effect is achieved. Therefore, the invention utilizes DGA data of transformer fault to construct, and configures a corresponding spatial information map containing spatial information for each fault data, thereby supplementing more characteristic information and realizing the improvement of the accuracy of transformer fault diagnosis. The characteristic data block is constructed in the following way:
(1) Dividing existing transformer fault DGA data into a training set and a test set in proportion, and setting the number of data contained in the training set and the test set as [ M, N ] respectively;
(2) Creating a zero array of 100 by 3 (i.e., an "RGB map" with 100 length and width and 3 layers), each element being 0;
(3) The first data a in the training set is taken, and the spatial information map is filled with the remaining M-1 DGA data. The filling rule is as follows: for M-1 data, with H for each data 2 Concentration x 0 As x-axis coordinate, CH 4 Concentration y 0 As y-axis coordinate, C 2 H 6 、C 2 H 4 、C 2 H 2 The concentrations are filled in the spatial information map as parameters "R, G, B", respectively. Finally, an "RGB map" containing no more than M-1 non-zero pixels can be obtained (since there may be data with the same x and y parameters, in this case, multiple maps can be constructed specifically for this point).
Take an array in a Python environment as an example: for a piece of DGA data B, H 2 、CH 4 、C 2 H 2 、C 2 H 4 And C 2 H 6 The concentration ratios are respectively [0.095 0.073.621 0.13.081]Then, in the spatial information map, the layer 0 parameter of the 10 th column and the 9 th row is 0.621, the layer 1 parameter is 0.13, and the layer 2 parameter is 0.081, as shown in fig. 3.
(4) And (4) repeating the steps (2) to (3) for the rest data in the training set, and finally obtaining M different spatial information graphs in total without repetition.
(5) Taking one piece of data in the training set and a spatial information graph corresponding to the piece of data (namely, the spatial information graph does not contain nonzero pixel points corresponding to the piece of data and has other M-1 nonzero pixel points) as a feature data block, and counting M different feature data blocks. Thereby, a feature data block of the training data set is obtained.
(6) In the manner in step (2) and step (3), M training data are filled into a spatial information map, which is called a "reference map".
(7) If one data in the test set and the reference image are taken as one characteristic data block, the total number of the characteristic data blocks is N. Thereby, a feature data block of the test data set is obtained. The structural form of the characteristic data block of the test data is consistent with the structural form of the characteristic data block of a piece of DGA data to be diagnosed, which is acquired in actual use.
2. Transformer fault diagnosis model structure based on CNN convolutional network
Considering the construction of the characteristic data block, the structure of the transformer fault diagnosis model based on the CNN convolutional network is shown in FIG. 4.
The model is composed of four parts in total:
one layer is a convolution layer part with an input interface size of 100 × 3, and the part is used for carrying out feature identification and feature extraction on the spatial information graph.
The two layers are flat layers, on the premise of ensuring the model performance and avoiding huge calculation overhead, the output data of the convolution layer part is a two-dimensional array, the ranks of the two-dimensional array are not 1, and therefore the two-dimensional array is processed into a one-dimensional array by using the flat layers.
The three layers are all-link layers, and the first layer is also a gathering layer for gathering the output data of the flat layer and the fault DGA data in sequence as the input of the all-link layer part. And the rest layers of the full connection layer are used for carrying out data analysis and feature extraction.
The four layers are output layers, the input of each output layer is the output of the last layer of the full connection layer, the output dimensionality of each output layer is equal to the fault category of the transformer, and the output layers are subjected to normalization processing through a Softmax function.
3. Transformer fault diagnosis model operation mode based on CNN convolutional network
The core idea of the operation of the transformer fault diagnosis model based on the CNN convolutional network is an image identification technology. By inputting the spatial information map as a part of feature data into the convolutional layer, because the data in the spatial information map has obvious distribution difference on spatial position and color, and the CNN convolutional network can identify data features on different dimensions, after data fitting is performed, it can be considered that the CNN convolutional network has accurately found a high-dimensional curved surface of the partitioned data in the spatial information map through learning, and the output of the CNN convolutional network is the boundary curved surface of the spatial information map or the high-dimensional representation thereof. By inputting the curved surface or the high-dimensional representation thereof and the corresponding fault DGA data into the full-link network for further fitting, the full-link network can be trained to classify the fault DGA data according to the boundary curved surface, and correct classified data can be obtained finally.
In the training process, for each fault DGA data, the corresponding spatial information image does not contain the pixel points of the fault DGA data, the difference of the spatial information images is very small dynamic images from the perspective of a training set corpus, and the difference of only two pixel points exists between every two images under the condition that repetition is not considered. However, this operation is very desirable, and if all the training data are "drawn" directly on the spatial information map, each training data corresponds to the same map, as shown in fig. 5. The input of the CNN convolutional network is changed into static state in the training process, and when fault diagnosis is carried out, because pixel points are newly added in the spatial information graph and the previous training tends to be static, the generalization capability of the convolutional layer to the spatial information graph is insufficient, the output of the convolutional layer has large deviation, and finally the model prediction precision is reduced.
Instead, a moving picture construction is used, as shown in fig. 6. Sufficient information is provided by other data in the spatial information map, only one or two pixel points are lost, the influence on the contained information quantity is small, and the slight dynamic change of the spatial information map can stimulate a convolution network to identify the difference of the image during training, so that sample data can be divided from multiple angles, a more reasonable boundary curved surface is finally divided in the spatial information map, data expansion is effectively performed on the basis of the existing data set, and the model is continuously learned and continuously updated by using new fault DGA data in long-term application.
4. Example analysis and Experimental results
(1) Sample data selection
The training and testing samples selected by the method mainly come from actual transformer fault experimental data and transformer fault data given by various documents. The DGA analysis result of the oil-immersed transformer containing 835 cases in total when the oil-immersed transformer fails, and the tested gas comprises H 2 ，CH 4 ，C 2 H 4 ，C 2 H 6 ，C 2 H 2 The test concentration of the five types of gases is in ppm unit, the proportion of the five types of gases is used as input, and the proportion of the training set to the test set is 4:1. The selected attributes are all from the operation combination of DGA analysis gas, and the data are all applied to the judgment of relevant transformer fault diagnosis. The integration into the input space vector H can be represented as a matrix 835 x 5:
H＝[H 2 ,CH 4 ,C 2 H 4 ,C 2 H 6 ,C 2 H 2 ] 835×5
(2) Transformer fault state coding
The fault diagnosis states of the transformer are totally divided into 6 types, and the coding conditions are shown in table 1:
TABLE 1 Transformer Fault status coding
(3) Model structure comparison of dynamic and static spatial information diagram
4 groups of training data are adopted to carry out testing under a dynamic and static network space information graph model respectively, and the experimental result is shown in figure 7.
From the results of fig. 7, it can be seen that the diagnosis accuracy of the dynamic spatial information map of the sample is improved to different degrees in comparison with the diagnosis accuracy of the static spatial information map under 4 types of training sample data, and it can be seen that the dynamic training model can enhance the basic feature recognition of the sample, and has better difference recognition and classification effects.
Comparison of different types of diagnostic results:
the fault sample data used in the present document are respectively tested and evaluated in a single DNN network, a simple CNN network, an SVM combination optimization network and an FI-CNN network algorithm, and the test result is shown in fig. 8.
According to the test results, the fault accuracy of the DNN, CNN, DCAE-SVM and FI-CNN network algorithms is increased in sequence, and the fault accuracy of the FI-CNN network is obviously higher than that of other algorithms, because the algorithms realize extraction of the prominent key factors, and a high-dimensional curve for dividing the data categories is found through learning of the spatial information map in the training process, the fault data are classified more strictly.
In order to further study the diagnosis efficiency of each algorithm under different numbers of training samples, the sample data in the training set is divided, so that the samples in the training set are 167, 334, 501 and 668 respectively, the sample capacity of the test set is 167, and the diagnosis accuracy of each algorithm under different sample numbers is calculated. The experimental results are shown in FIG. 9.
According to the test results, the fault diagnosis rate of each algorithm is improved to different degrees along with the increase of the number of the samples in the training set, and the FI-CNN algorithm has the most obvious improvement of the diagnosis accuracy rate, which is related to the unique property of the algorithm. When the number of samples is continuously increased, the filling of a space information graph of each sample on a two-dimensional coordinate is more complete, the feature information is increased, and the division of a data type high-dimensional feature curve is more accurate, so that the features of DGA data under different conditions are easier to extract, and the fault diagnosis accuracy is improved.
Claims (4)
1. A transformer fault diagnosis method based on FI-CNN is characterized by comprising the following steps:
step 1: analyzing the relation between the main faults and the characteristic quantity of the transformer, sorting and analyzing according to the influence weight, and extracting the content of various dissolved gases in the transformer oil;
step 2: selecting 3 gases from the dissolved gases in the transformer oil, wherein the concentration ratios of the 3 gases are expressed in an RGB (Red Green blue) form, constructing by using DGA (differential global positioning System) data of transformer faults, configuring a corresponding spatial information map for each fault data, supplementing more characteristic information and completing the construction of a characteristic data block;
and step 3: considering the structure of the characteristic data block, constructing a transformer fault diagnosis model based on a CNN convolutional network;
and 4, step 4: and training the transformer fault diagnosis model based on the CNN convolutional network to realize transformer fault diagnosis.
2. The FI-CNN-based transformer fault diagnosis method of claim 1, wherein the characteristic data block is constructed in a specific manner as follows:
step 2.1: dividing existing transformer fault DGA data into a training set and a test set in proportion, and setting the number of data contained in the training set and the test set to be M and N respectively;
step 2.2: creating a zero array of 100 × 3, namely an RGB graph with the length and width of 100 and the number of layers of 3, wherein each element is 0;
step 2.3: taking the first data A in the training set, and filling the spatial information map with the remaining M-1 DGA data; the filling rule is as follows: for M-1 data, with H for each data 2 Concentration x 0 X-axis coordinate as Euclidean plane, CH 4 Concentration y 0 As y-axis coordinate, C 2 H 6 、C 2 H 4 、C 2 H 2 The concentrations are respectively used as R, G, B parameters and are filled in a spatial information graph to obtain an RGB graph containing not more than M-1 nonzero pixel points;
step 2.4: repeating the steps 2.2-2.3 for the residual data in the training set, and finally obtaining a total of M different spatial information graphs without repetition;
step 2.5: taking one data in the training set and a corresponding spatial information graph as a feature data block, wherein the total number of the feature data blocks is M; thus, obtaining a characteristic data block of the training data set;
step 2.6: filling M training data into a spatial information graph to obtain a reference graph in the mode of the steps 2.2-2.3;
step 2.7: taking one data in the test set and the reference image as a characteristic data block, wherein N different characteristic data blocks are counted; thereby, a feature data block of the test data set is obtained.
3. The FI-CNN-based transformer fault diagnosis method of claim 2, wherein the transformer fault diagnosis model comprises four layers:
one layer is a convolution layer part, the size of an input interface of the convolution layer part is 100 × 3, and the part is used for carrying out feature identification and feature extraction on the spatial information graph;
the second layer is a flat layer and is used for processing the two-dimensional array output by the convolution layer part into a one-dimensional array;
the three layers are full connection layers, the first layer is also a gathering layer and is used for gathering output data of the flat layer and fault DGA data in sequence as input of the full connection layer part; the rest layers of the full connection layer are used for data analysis and feature extraction; the four layers are output layers, the input of the output layers is the output of the last layer of the full connection layer, the output dimensionality of the output layers is equal to the fault type of the transformer, and normalization processing is carried out through a Softmax function.
4. The FI-CNN-based transformer fault diagnosis method of claim 3, wherein the transformer fault diagnosis model operates in the following manner:
inputting the spatial information graph into a convolutional layer as a part of characteristic data, after data fitting is carried out, accurately finding a high-dimensional curved surface of the partitioned data in the spatial information graph through learning by a CNN convolutional network, wherein the output of the CNN convolutional network is the boundary curved surface of the spatial information graph or high-dimensional representation thereof;
and simultaneously inputting the curved surface or the high-dimensional representation thereof and the corresponding fault DGA data into a full-link network for further fitting, training the full-link network to classify the fault DGA data according to the boundary curved surface, and finally obtaining correct classified data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210703453.5A CN115201640A (en) | 2022-06-21 | 2022-06-21 | FI-CNN-based transformer fault diagnosis method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210703453.5A CN115201640A (en) | 2022-06-21 | 2022-06-21 | FI-CNN-based transformer fault diagnosis method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115201640A true CN115201640A (en) | 2022-10-18 |
Family
ID=83575562
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210703453.5A Pending CN115201640A (en) | 2022-06-21 | 2022-06-21 | FI-CNN-based transformer fault diagnosis method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115201640A (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104299035A (en) * | 2014-09-29 | 2015-01-21 | 国家电网公司 | Method for diagnosing fault of transformer on basis of clustering algorithm and neural network |
| CN111337768A (en) * | 2020-03-02 | 2020-06-26 | 武汉大学 | Deep parallel fault diagnosis method and system for dissolved gas in transformer oil |
| CN112085084A (en) * | 2020-08-24 | 2020-12-15 | 宁波大学 | Transformer fault diagnosis method based on multi-feature fusion common vector |
| CN112163619A (en) * | 2020-09-27 | 2021-01-01 | 北华大学 | Transformer fault diagnosis method based on two-dimensional tensor |
| US20210397158A1 (en) * | 2020-06-23 | 2021-12-23 | Horiba Stec, Co., Ltd. | Computing system with discriminative classifier for determining similarity of a monitored gas delivery process |
| CN114119765A (en) * | 2021-10-11 | 2022-03-01 | 湖北忠和冶金机械制造股份有限公司 | Automatic sole gluing method based on shoe shape characteristic curve extraction |
-
2022
- 2022-06-21 CN CN202210703453.5A patent/CN115201640A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104299035A (en) * | 2014-09-29 | 2015-01-21 | 国家电网公司 | Method for diagnosing fault of transformer on basis of clustering algorithm and neural network |
| CN111337768A (en) * | 2020-03-02 | 2020-06-26 | 武汉大学 | Deep parallel fault diagnosis method and system for dissolved gas in transformer oil |
| US20210397158A1 (en) * | 2020-06-23 | 2021-12-23 | Horiba Stec, Co., Ltd. | Computing system with discriminative classifier for determining similarity of a monitored gas delivery process |
| CN112085084A (en) * | 2020-08-24 | 2020-12-15 | 宁波大学 | Transformer fault diagnosis method based on multi-feature fusion common vector |
| CN112163619A (en) * | 2020-09-27 | 2021-01-01 | 北华大学 | Transformer fault diagnosis method based on two-dimensional tensor |
| CN114119765A (en) * | 2021-10-11 | 2022-03-01 | 湖北忠和冶金机械制造股份有限公司 | Automatic sole gluing method based on shoe shape characteristic curve extraction |
Non-Patent Citations (2)
| Title |
|---|
| NAN LIN 等: "Transformer fault diagnosis model based on FI-CNN method", 《INTERNATIONAL CONFERENCE ON INTELLIGENT TRAFFIC SYSTEMS AND SMART CITY (ITSSC 2021)》, 14 March 2022 (2022-03-14), pages 1 - 10, XP060152803, DOI: 10.1117/12.2628003 * |
| 朱双静: "基于深度学习的换流变压器局部放电模式识别方法研究", 《中国优秀硕士学位论文全文数据库工程科技Ⅱ辑 》, no. 1, 15 January 2020 (2020-01-15), pages 042 - 2760 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Dai et al. | Dissolved gas analysis of insulating oil for power transformer fault diagnosis with deep belief network | |
| CN114219147A (en) | Power distribution station fault prediction method based on federal learning | |
| CN116451123A (en) | Transformer early fault diagnosis method based on ISSA-MKELM | |
| Li et al. | Transformer fault diagnosis based on improved deep coupled dense convolutional neural network | |
| Huang et al. | Research on transformer fault diagnosis method based on GWO optimized hybrid kernel extreme learning machine | |
| Han et al. | Research on quality problems management of electric power equipment based on knowledge–data fusion method | |
| CN117421571A (en) | Topology real-time identification method and system based on power distribution network | |
| Zhang et al. | Encoding time series as images: A robust and transferable framework for power system DIM identification combining rules and VGGNet | |
| CN114266396A (en) | Transient stability discrimination method based on intelligent screening of power grid characteristics | |
| CN116756939B (en) | Transformer fault diagnosis method based on ANOVA-IWOA-SVM hybrid model | |
| CN113887623A (en) | IFCM-BB-based transformer fault diagnosis method | |
| CN115201640A (en) | FI-CNN-based transformer fault diagnosis method | |
| CN117332330A (en) | Transformer fault evaluation method and system based on multi-source data graph fusion model | |
| CN114595788B (en) | Transformer fault diagnosis method, device and equipment | |
| CN112085083B (en) | Transformer fault diagnosis method based on similarity analysis strategy | |
| Carmona et al. | Genetic lateral tuning for subgroup discovery with fuzzy rules using the algorithm NMEEF-SD | |
| CN115828185A (en) | Fault diagnosis method for oil immersed transformer | |
| CN114565051A (en) | Test method of product classification model based on neuron influence degree | |
| CN114157552A (en) | Distribution network fault detection method based on twin timing diagram network | |
| Peng et al. | Transformer DGA fault diagnosis method based on stacked sparse auto-encoders | |
| Zhang et al. | Fault diagnosis of power transformer based on SSA—MDS pretreatment | |
| Li et al. | Research on transformer fault diagnosis based on online sequential extreme learning machine | |
| Lin et al. | Transformer fault diagnosis model based on FI-CNN method | |
| Preetha et al. | Diagnosis of Transformer Faults Using the Multi-Class Adaboost Algorithm | |
| Jiang et al. | Research on Transformer Fault Diagnosis with Support Vector Machine Optimized by Slime Mold Algorithm Based on Principal Component Analysis |
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 |