CN109614488A - Distribution network live line work condition distinguishing method based on text classification and image recognition - Google Patents
Distribution network live line work condition distinguishing method based on text classification and image recognition Download PDFInfo
- Publication number
- CN109614488A CN109614488A CN201811470666.8A CN201811470666A CN109614488A CN 109614488 A CN109614488 A CN 109614488A CN 201811470666 A CN201811470666 A CN 201811470666A CN 109614488 A CN109614488 A CN 109614488A
- Authority
- CN
- China
- Prior art keywords
- model
- layer
- network
- condition
- preselected area
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 25
- 238000012549 training Methods 0.000 claims abstract description 32
- 238000010801 machine learning Methods 0.000 claims abstract description 22
- 238000012360 testing method Methods 0.000 claims abstract description 15
- 238000004519 manufacturing process Methods 0.000 claims abstract description 4
- 239000011159 matrix material Substances 0.000 claims description 25
- 238000013527 convolutional neural network Methods 0.000 claims description 22
- 238000005516 engineering process Methods 0.000 claims description 13
- 238000013135 deep learning Methods 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 5
- 239000000203 mixture Substances 0.000 claims description 5
- 238000000605 extraction Methods 0.000 claims description 4
- 230000005484 gravity Effects 0.000 claims description 4
- 230000008676 import Effects 0.000 claims description 4
- 230000000007 visual effect Effects 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 230000015572 biosynthetic process Effects 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 230000007547 defect Effects 0.000 claims description 3
- 238000007689 inspection Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 238000012546 transfer Methods 0.000 claims description 3
- 239000012212 insulator Substances 0.000 claims description 2
- 239000000284 extract Substances 0.000 description 4
- 238000001514 detection method Methods 0.000 description 3
- 230000004069 differentiation Effects 0.000 description 3
- 238000012423 maintenance Methods 0.000 description 3
- 230000009466 transformation Effects 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 230000007812 deficiency Effects 0.000 description 2
- 230000001151 other effect Effects 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000013139 quantization Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 230000019771 cognition Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000009472 formulation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000009413 insulation Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Economics (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Public Health (AREA)
- Water Supply & Treatment (AREA)
- General Health & Medical Sciences (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Image Analysis (AREA)
Abstract
The distribution network live line work condition distinguishing method based on text classification and image recognition that the invention discloses a kind of, first distribution line external condition data are exported from grid company production management system, it generates distribution line live working external condition and differentiates text database, again by taking pictures to distribution line scene, generate line facility image database, database is pre-processed, then Automated Text Categorization for Chinese Documents model and image recognition disaggregated model based on machine learning are built, database is divided into training set and test set two major classes, the training for having supervision is carried out to two models respectively using training set, trained model is tested using test set；Finally freshly harvested data are input in trained model, model is identified and is scored to each criterion feature, grading system is corresponded to according to general comment score value and differentiates whether the distribution line meets the requirement of livewire work condition, is staff to whether can be carried out livewire work and provide intelligently effective decision-making foundation.
Description
Technical field
The invention belongs to low and medium voltage distribution network line live-line work technical fields, and in particular to one kind based on text classification and
The distribution network live line work condition distinguishing method of image recognition.
Background technique
Livewire work is the important means and method of grid equipment detection, repair and maintenance and transformation, is to guarantee electric system
The important technique measure of reliable and stable operation.With the continuous improvement that distribution network reliability requires, distribution line electrification is made
Industry and uninterrupted operation method are also used widely and are promoted.New technology, new equipment, new material livewire work field application
Also the progress and development of live line tool, equipment, standard formulation etc. have effectively been pushed, and to livewire work technology from now on
It puts forward new requirements and developing direction.However, current major part grid company is still about the differentiation of livewire work condition
Pass through the artificial clear cognition gone to scene to check equipment, lack to current device the location of in electric network composition, intelligence
It is insufficient to change analysis means, artificial subjective factor influences more；None unified discrimination standard can carry out band to route
The internal and external factor of electric operation analyzes deficiency, can not quantify whether each condition can be carried out the influence of livewire work to route.
Summary of the invention
In view of the deficiency of the prior art, the present invention provides a kind of distribution based on text classification and image recognition
Livewire work condition distinguishing method, so that solving at present can not be intelligentized according to distribution line external condition and appointed condition number
Whether can be carried out the problem of livewire work judges according to distribution line.
The technical problems to be solved by the invention are achieved by the following technical programs:
A kind of distribution network live line work condition distinguishing method based on text classification and image recognition, comprising the following steps:
S1. distribution line external condition data are exported from grid company production management system, generates distribution line electrification and makees
Condition distinguishing text database in portion's out of trade；
S2. in distribution line collection in worksite picture, line facility condition data is formed, generates line facility image data
Library；
S3. text database and image data base are pre-processed, comprising: by route external condition and line facility item
Part respectively corresponds different score value formation condition score tables, and condition score table reflects the corresponding score value of every kind of condition, score value size
Reflect the specific gravity that every kind of condition accounts for；Text is indicated in the form of matrix or vector, image is split and is extracted
To the character representation with invariance；
S4. the Automated Text Categorization for Chinese Documents model based on machine learning and the image recognition classification based on machine learning are built
Model；
S5. pretreated text database and image data base are divided into training set and test set two major classes, utilize instruction
Practice collection data respectively to the Automated Text Categorization for Chinese Documents model put up based on machine learning and based on the image of machine learning
Identification disaggregated model carries out the training for having supervision, the accuracy rate then identified using the trained model of test set data test,
Model is set to reach 90% or more to the accuracy rate of the data identification in test set by adjusting parameter；
S6. grading system is divided to distribution line, will newly acquires data and imports in trained model, model identification electrification
Operation criterion feature simultaneously scores, and corresponds to whether grading system judgement meets livewire work item under this condition according to general comment score value
Part requirement.
The route external condition data of the step S1 include: power supply area, electric network composition, N-n inspection, landform, user
Access and power distribution automation are horizontal.
The line facility condition data of the step S2 includes: overhead line, rod-type, cut-off equipment, transformer equipment, insulation are set
Standby and fitting.
The step S4 builds the concrete operations of the Automated Text Categorization for Chinese Documents model based on machine learning are as follows: with word for singly
Position carries out text representation and forms term vector, then term vector is spliced according to the sequence that word occurs in sentence, is formed and is represented
The matrix of sentence is then fed into the convolutional neural networks model based on deep learning technology, on the basis of term vector, is realized
Sentence characteristics automatically extract and learn, and finally realize the automatic classification of defect text.Wherein, described to be based on deep learning technology
Convolutional neural networks model be one four layers of convolutional neural networks model, concrete form are as follows:
First layer is input layer, and input layer is the corresponding phrase matrix W ∈ R of a non-classified external conditions×n, W representative
The corresponding phrase of one non-classified external condition, R represent the matrix of phrase conversion, and every a line of matrix represents every in phrase
The corresponding vector of a word, line number s are the word number of phrase, and columns n is the dimension of vector；
The second layer is one-dimensional convolutional layer, uses that columns is identical as W, line number is the convolution matrix window I ∈ R of hh×n, with input
Each h row n column matrix block of layer matrix W successively carries out convolution algorithm from top to bottom, and wherein each convolution window can be from input
Matrix R in extract a characteristic pattern feature, referred to as text feature；
Third layer is pond layer, using the method in maximum pond, the characteristic pattern vector for taking each convolution window convolution to obtain
In maximum element as characteristic value, to extract the corresponding characteristic value of each convolution window, and all characteristic values are successively spelled
The one-dimensional vector for constituting pond layer is connect, the vector of sentence global characteristics is as represented；
4th layer is output layer, and output layer is connect entirely with pond layer, is input with the one-dimensional vector of pond layer, by activation
Function output, along with lose layer removal partial data prevent over-fitting, finally using softmax classifier to one-dimensional vector into
Row classification, and export final classification results.
The step S4 builds the concrete operations of the image recognition disaggregated model based on machine learning are as follows: Xian Yi is international large-scale
The Classification and Identification of visual object and the database of detection challenge match are that template establishes image recognition database, to store by pre-
Image data that treated is then fed into the convolutional neural networks model based on deep learning technology, in image preprocessing
On the basis of, it realizes automatically extracting and learning for characteristics of image, finally realizes the marking classification of line facility condition in image data.
Wherein, the convolutional neural networks model based on deep learning technology includes three bulky components network models:
First piece is pre-training front network, using ResNet50 as pre-training model, first ResNet50 network mould
Then model parameter in type not comprising full articulamentum defines the network structure of ResNet50, reloads Model Weight to locally
Parameter finally changes the structure of the last one full articulamentum into the network structure of definition, starts to train with lower learning rate,
Obtain the good front network model of pre-training；
Second piece is preselected area network, and preselected area network is to export rectangle mesh using image in training set as input
The set of preselected area is marked, each preselected area has a score, this score judges whether selected region is target
Region；In order to generate rectangular target preselected area, by the way that in pre-training front network, the last one is shared behind convolutional layer
A small sliding window is added, this sliding window is connected to entirely in the spatial window of input convolution Feature Mapping, each cunning
Dynamic window is mapped on a low-dimensional vector, this vector, which is exported, returns layer and preselected area classification layer, pre-selection to preselected area
Region returns the codes co-ordinates that layer finally exports preselected area, and preselected area classification layer finally exports the score of preselected area, leads to
It crosses score and judges whether the preselected area is target region, then being sent to down for genuine rectangular target preselected area set
Classification and Identification is carried out in primary network station；
Third block is fast area convolutional neural networks, and fast area convolutional neural networks are total with preselected area network
The sharing feature layer for enjoying the initialization of pre-training front network carries out convolutional network feature extraction to image in pre-training front network
Later, rectangular target preselected area is exported through preselected area network, generates rectangular target preselected area convolution characteristic pattern, take out square
Corresponding depth characteristic on shape target preselected area convolution characteristic pattern, will be in channel with a rectangular target preselected area pond layer
Whole features be unified into same size, generate the characteristic pattern of a fixed dimension, finally obtain by two full connection features layers
To feature vector, feature vector completes line facility in image via two multi task models in respective full articulamentum again
Identification and frame choosing；Described two multi task models are identification disaggregated model and pre-selected zone based on flexible maximum value transfer function
Domain window regression model.
The invention has the following advantages: the present invention proposes to differentiate the impact factor whether distribution is capable of livewire work,
Reasonable work quantity evaluation table is proposed to determine the influence specific gravity of each impact factor, builds the Chinese text based on machine learning certainly
Dynamic disaggregated model carries out quantization marking come the external condition to route to be assessed, builds the image recognition based on machine learning point
Class model carries out quantization marking come the interior condition to route to be assessed, provides intelligence for whether distribution meets livewire work condition
The method of discrimination of energyization, and model can accurately differentiate the section of livewire work condition.To which solution at present can not be intelligent
Asked according to distribution line external condition and appointed condition data what whether distribution line can be carried out that livewire work judges
Topic.
Detailed description of the invention
Fig. 1 is the process of the distribution network live line work condition distinguishing method of the invention based on text classification and image recognition
Figure；
Specific embodiment
The present invention is described in further detail with reference to the accompanying drawing.
As shown in Figure 1, a kind of distribution network live line work condition based on text classification and image recognition proposed by the present invention is sentenced
Other method, using deep learning technology, using text classification and image recognition model to the route for needing livewire work in distribution
Differentiation marking is carried out, judges whether distribution line is able to carry out livewire work and provides intelligence effective decision for livewire work personnel
Foundation.Specific step is as follows:
S1. distribution line external condition data are exported, distribution line live working external condition is generated and differentiates text data
Library:
It can be used in the distribution line that distribution line live working condition judges from the export of grid company production management system
External condition data generate distribution line live working external condition and differentiate text database.Route external condition data include:
Power supply area, electric network composition, N-n inspection, landform, user's access and power distribution automation are horizontal.
S2. picture is acquired in distribution line device context, forms line facility condition data, generate line facility image number
According to library: taking pictures at distribution line scene, acquisition, which is able to reflect distribution line real-time status and can be used in distribution line electrification, to be made
The distribution line appointed condition data of industry condition judgement, generate line facility image database.Line facility condition data includes:
Overhead line (cable), cut-offs equipment, transformer equipment, insulator arrangement and fitting at rod-type.
S3. text database and image data base are pre-processed:
Route external condition and line facility condition are respectively corresponded to different score value formation condition score tables, condition score value
Table reflects the corresponding score value of every kind of condition, and score value size reflects specific gravity shared by every kind of condition.
Route external condition and the corresponding score value of line facility condition are as shown in table 1 below.
1 condition score value of table corresponds to table
In being pre-processed to text database and image data base, using hidden Markov model, in text
Segmentation, subordinate sentence, participle, removal stop words etc. extract, then convert text to computer and can recognize and the form of processing,
Text is indicated in the form of matrix or vector；Image data base is pre-processed, excavates electric power widget first
Prior shape statistical nature, then utilize local invariant feature, image is split and is extracted will obtain having it is constant
The character representation of property.
S4. the Automated Text Categorization for Chinese Documents model based on machine learning and the image recognition classification based on machine learning are built
Model:
S41. the Automated Text Categorization for Chinese Documents model based on machine learning is built:
Carry out text representation as unit of word and form term vector, then the sequence that term vector is occurred in sentence according to word into
Row splicing, forms and represents the matrix of sentence, be then fed into the convolutional neural networks model based on deep learning technology, word to
On the basis of amount, automatically extracting and learning for sentence characteristics is realized, finally realize the automatic classification of defect text.
Wherein, the convolutional neural networks model that the convolutional neural networks model based on deep learning technology is one four layers,
Concrete form is as follows:
First layer is input layer.Input layer is the corresponding phrase matrix W ∈ R of a non-classified external conditions×n, W representative
The corresponding phrase of one non-classified external condition, R represent the matrix of phrase conversion, and every a line of matrix represents every in phrase
The corresponding vector of a word, line number s, that is, phrase word number, columns n, that is, vector dimension.
The second layer is one-dimensional convolutional layer.Use columns (for n) identical as W, the convolution matrix window I ∈ R that line number is hh×n,
Convolution algorithm is successively carried out from top to bottom with each h row n column matrix block of input layer matrix W, wherein each convolution window energy
A characteristic pattern feature, referred to as text feature are extracted from the matrix R of input.
Third layer is pond layer.Using the method in maximum pond, the characteristic pattern vector for taking each convolution window convolution to obtain
In maximum element as characteristic value, to extract the corresponding characteristic value of each convolution window, and all characteristic values are successively spelled
The one-dimensional vector for constituting pond layer is connect, the vector of sentence global characteristics is as represented.
4th layer is output layer.Output layer is connect entirely with pond layer, is input with the one-dimensional vector of pond layer, by activation
Function output, along with lose layer removal partial data prevent over-fitting, finally using softmax classifier to one-dimensional vector into
Row classification, and export final classification results.
S42. the image recognition disaggregated model based on machine learning is built:
The Classification and Identification of the international large-scale visual object of Xian Yi and detection challenge match (Pattern Analysis,
Statistical modelling and Computational Learning Visual Object Classes, PASCAL
VOC database) is that template establishes image recognition database, to store by pretreated image data, is then fed into
In convolutional neural networks model based on deep learning technology, on the basis of image preprocessing, the automatic of characteristics of image is realized
It extracts and learns, finally realize the marking classification of line facility condition in image data.
Wherein, the convolutional neural networks model based on deep learning technology includes three bulky components network models altogether:
First piece is pre-training front network model.In the learning process of deep learning, since computing resource is limited or
Person's training set is smaller, preferably more stable as a result, by being finely adjusted some trained network models, then in order to obtain
It is imported in whole network identification model again.Using ResNet50 as pre-training model.First in ResNet50 network model not
Then model parameter comprising full articulamentum defines the network structure of ResNet50, reloads Model Weight parameter and arrive to locally
In the network structure of definition, the structure of the last one full articulamentum is finally changed, starts to train with lower learning rate, is obtained pre-
Trained front network model.
Second piece is preselected area network (Regional Proposal Network, RPN).The core concept of RPN network
It is to export the set of rectangular target preselected area (Region Of Interest, ROI) using image in training set as input,
Each preselected area has a score, this score judges whether selected region is target region.In order to generate
Rectangular target preselected area, by adding a small sliding window behind the last one shared convolutional layer in pre-training front network
Mouthful, this sliding window be connected to full input convolution Feature Mapping spatial window on, each sliding window be mapped to one it is low
On dimensional vector, this vector, which is exported, returns layer (reg) and preselected area classification to two full articulamentums at the same level-preselected area
Layer (cls), the codes co-ordinates of reg layers of last output preselected area, the score of cls layers of last output preselected area pass through score
Judge whether the preselected area is target region, then being sent to next stage net for genuine rectangular target preselected area set
Classification and Identification is carried out in network.
Third block is fast area convolutional neural networks (Fast R-CNN).Fast R-CNN network is total with RPN network
Enjoy the sharing feature layer of pre-training front network initialization.Convolutional network feature extraction is carried out to image in pre-training front network
Later, ROI is exported through RPN network, generates ROI convolution characteristic pattern, take out corresponding depth characteristic on ROI convolution characteristic pattern, used
Whole features in channel are unified into same size by one pond ROI layer, are generated the characteristic pattern of a fixed dimension, are most passed through afterwards
It crosses two full connection features layers and obtains feature vector, feature vector is again via two multitask moulds in respective full articulamentum
Type-is based on the identification disaggregated model of flexible maximum value transfer function (softmax) and preselected area window returns (BBox) model
To complete the identification of power equipment and frame choosing in image.
S5. pretreated text database and image data base are divided into training set and test set two major classes, utilize instruction
Practice collection data respectively to the Automated Text Categorization for Chinese Documents model put up based on machine learning and based on the image of machine learning
Identification disaggregated model carries out the training for having supervision, then utilizes the accuracy rate of the trained model identification of test set data test.
So that two models is reached 90% or more to the accuracy rate of the data identification in test set by adjusting parameter, obtains an optimal knowledge
The Automated Text Categorization for Chinese Documents model and image recognition disaggregated model based on machine learning of other effect.
S6. grading system is divided to distribution line, will newly acquires data and imports in trained model, model identification electrification
Operation criterion feature simultaneously scores, and corresponds to whether grading system differentiation meets livewire work item under this condition according to general comment score value
Part requirement.
Assuming that certain route has n if appropriate for livewire work condition, the score value of some condition is X, then whole route is
The average of no suitable electrification operating condition are as follows:
Wherein L is the total score of the no suitable electrification operating condition of certain route, and i indicates the one of Rule of judgment of route,
Value range is from 1 to n.
After obtaining general comment score value, according to the corresponding grade of the score value of following table and its explanation, to differentiate the distribution line
Whether livewire work condition requirement is met.
2 score value of table and grading system classification chart
| General comment score value | 70 points and less | 70-80 points | 80-90 points | 90-100 points |
| Grading system | Interruption maintenance route | Route to be rebuilt | Quasi- not dead line | Not dead line |
Grading system:
(1) not dead line: full line overall score meets the requirement of not dead line；
(2) quasi- not dead line: full line has equipment component that need to can reach not dead line by small range transformation
It is required that；
(3) route to be rebuilt: full line can be only achieved the requirement of uninterrupted operation route by large range of transformation；
(4) interruption maintenance route: full line needs full cut-off that the requirement that can be only achieved uninterrupted operation route is transformed.
When full line overall score meets the requirement of each rank not dead line, but full line has a few devices not meet not stop
Electric job requirements, then drop level-one for circuit grade.
The present invention is under Windows10 operating system, CPU model Intel Core i7, GPU model
Building for network model is completed in the configuration of NVIDIAGeForce GTX 970,4G independence video memory.Using Tensorflow framework
Realize convolutional neural networks model.Partial data in database is imported model as training set first to be trained, wait train
After completion, imports test set data and tested, the knowledge of textual classification model is considered using error rate and severe deviations rate
Other effect, the recognition effect of image recognition model is considered using accuracy rate and recall rate.The result shows that the error rate of model and
Severe deviations rate is respectively 2.86% and 0.80%, and accuracy rate and recall rate are respectively 92% and 86%, illustrates mould of the invention
Whether type largely can accurately differentiate the section of livewire work condition, be staff to can be carried out livewire work and provide
The effective decision-making foundation of intelligence.
Claims (7)
1. the distribution network live line work condition distinguishing method based on text classification and image recognition, which is characterized in that including following step
It is rapid:
S1. distribution line external condition data are exported from grid company production management system, generated outside distribution line live working
Portion's condition distinguishing text database；
S2. in distribution line collection in worksite picture, line facility condition data is formed, generates line facility image database；
S3. text database and image data base are pre-processed, comprising: by route external condition and line facility condition point
Different score value formation condition score tables is not corresponded to, and condition score table reflects the corresponding score value of every kind of condition, the reflection of score value size
Specific gravity shared by every kind of condition；Text is indicated in the form of matrix or vector, image is split and extraction obtains
Character representation with invariance；
S4. the Automated Text Categorization for Chinese Documents model based on machine learning and the image recognition classification mould based on machine learning are built
Type；
S5. pretreated text database and image data base are divided into training set and test set two major classes, utilize training set
Data are respectively to the Automated Text Categorization for Chinese Documents model put up based on machine learning and based on the image recognition of machine learning
Disaggregated model carries out the training for having supervision, then using the accuracy rate of the trained model identification of test set data test, passes through
Adjusting parameter makes model reach 90% or more to the accuracy rate of the data identification in test set；
S6. grading system is divided to distribution line, will newly acquires data and import in trained model, model identifies livewire work
Criterion feature simultaneously scores, and corresponds to grading system according to general comment score value and differentiates that whether meeting livewire work condition under this condition wants
It asks.
2. the method according to claim 1, wherein the route external condition data of the step S1 include: confession
Electric region, electric network composition, N-n inspection, landform, user's access and power distribution automation are horizontal.
3. the method according to claim 1, wherein the line facility condition data of the step S2 includes: frame
Ceases to be busy, rod-type cut-off equipment, transformer equipment, insulator arrangement and fitting.
4. the method according to claim 1, wherein the step S4 builds the Chinese text based on machine learning
The concrete operations of automatic disaggregated model are as follows: carry out text representation as unit of word and form term vector, then term vector is existed according to word
The sequence occurred in sentence is spliced, and the matrix for representing sentence is formed, and is then fed into the convolution mind based on deep learning technology
Through in network model, on the basis of term vector, realizes automatically extracting and learning for sentence characteristics, finally realize defect text
Automatic classification.
5. the method according to claim 1, wherein the step S4 builds the image recognition based on machine learning
The concrete operations of disaggregated model are as follows: the Classification and Identification of the international large-scale visual object of Xian Yi is template with the database for detecting challenge match
Image recognition database is established, to store by pretreated image data, is then fed into based on deep learning technology
In convolutional neural networks model, on the basis of image preprocessing, realizes automatically extracting and learning for characteristics of image, finally realize
The marking classification of line facility condition in image data.
6. according to the method described in claim 4, it is characterized in that, the convolutional neural networks model based on text classification is
One four layers of convolutional neural networks model, concrete form are as follows:
First layer is input layer, and input layer is the corresponding phrase matrix W ∈ R of a non-classified external conditions×n, W represents one
The corresponding phrase of non-classified external condition, R represent the matrix of phrase conversion, and every a line of matrix represents each word in phrase
Corresponding vector, line number s are the word number of phrase, and columns n is the dimension of vector；
The second layer is one-dimensional convolutional layer, uses that columns is identical as W, line number is the convolution matrix window I ∈ R of hh×n, with input layer square
Each h row n column matrix block of battle array W successively carries out convolution algorithm from top to bottom, and wherein each convolution window can be from the square of input
A characteristic pattern feature, referred to as text feature are extracted in battle array R；
Third layer is pond layer, using the method in maximum pond, in the characteristic pattern vector for taking each convolution window convolution to obtain most
All characteristic values to extract the corresponding characteristic value of each convolution window, and are successively spliced structure as characteristic value by big element
Cheng Chihua layers of one-dimensional vector as represents the vector of sentence global characteristics；
4th layer is output layer, and output layer is connect entirely with pond layer, is input with the one-dimensional vector of pond layer, by activation primitive
Output is prevented over-fitting along with layer removal partial data is lost, is finally divided using softmax classifier one-dimensional vector
Class, and export final classification results.
7. according to the method described in claim 5, it is characterized in that, the convolutional neural networks model packet based on image recognition
Containing three bulky components network models:
First piece is pre-training front network, using ResNet50 as pre-training model, first in ResNet50 network model
Then model parameter not comprising full articulamentum defines the network structure of ResNet50, reloads Model Weight parameter to locally
Into the network structure of definition, the structure of the last one full articulamentum is finally changed, starts to train with lower learning rate, obtain
The good front network model of pre-training；
Second piece is preselected area network, and preselected area network is using image in training set as input, and output rectangular target is pre-
The set of favored area, each preselected area have a score, this score is come where judging whether selected region is target
Region；In order to generate rectangular target preselected area, by being added behind the last one shared convolutional layer in pre-training front network
One small sliding window, this sliding window are connected to entirely in the spatial window of input convolution Feature Mapping, each sliding window
Mouth is mapped on a low-dimensional vector, this vector, which is exported, returns layer and preselected area classification layer, preselected area to preselected area
The codes co-ordinates that layer finally exports preselected area are returned, preselected area classification layer finally exports the score of preselected area, by
Point judge whether the preselected area is target region, then being sent to next stage for genuine rectangular target preselected area set
Classification and Identification is carried out in network；
Third block is fast area convolutional neural networks, and fast area convolutional neural networks are pre- with preselected area network share
Training front network initialization sharing feature layer, pre-training front network to image carry out convolutional network feature extraction it
Afterwards, rectangular target preselected area is exported through preselected area network, generates rectangular target preselected area convolution characteristic pattern, take out rectangle
Corresponding depth characteristic on target preselected area convolution characteristic pattern, will be in channel with a rectangular target preselected area pond layer
Whole features are unified into same size, generate the characteristic pattern of a fixed dimension, finally obtain by two full connection features layers
Feature vector, feature vector complete line facility in image via two multi task models in respective full articulamentum again
Identification and frame choosing；Described two multi task models are identification disaggregated model and preselected area based on flexible maximum value transfer function
Window regression model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811470666.8A CN109614488B (en) | 2018-12-04 | 2018-12-04 | Text classification and image recognition-based distribution network live working condition judgment method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811470666.8A CN109614488B (en) | 2018-12-04 | 2018-12-04 | Text classification and image recognition-based distribution network live working condition judgment method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109614488A true CN109614488A (en) | 2019-04-12 |
| CN109614488B CN109614488B (en) | 2022-12-02 |
Family
ID=66006906
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811470666.8A Active CN109614488B (en) | 2018-12-04 | 2018-12-04 | Text classification and image recognition-based distribution network live working condition judgment method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109614488B (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110490105A (en) * | 2019-08-06 | 2019-11-22 | 南京大国科技有限公司 | Distribute-electricity transformer district acceptance method, device and computer storage medium based on image recognition |
| CN110908901A (en) * | 2019-11-11 | 2020-03-24 | 福建天晴数码有限公司 | Automatic verification method and system for image recognition capability |
| CN111026870A (en) * | 2019-12-11 | 2020-04-17 | 华北电力大学 | ICT system fault analysis method integrating text classification and image recognition |
| WO2020220370A1 (en) * | 2019-05-01 | 2020-11-05 | Microsoft Technology Licensing, Llc | Method and system of utilizing unsupervised learning to improve text to content suggestions |
| CN112348066A (en) * | 2020-10-28 | 2021-02-09 | 国网浙江省电力有限公司绍兴供电公司 | Line uninterrupted power rating evaluation method based on gray clustering algorithm |
| CN112784652A (en) * | 2019-11-11 | 2021-05-11 | 中强光电股份有限公司 | Image recognition method and device |
| CN113742508A (en) * | 2021-07-30 | 2021-12-03 | 国网河南省电力公司信息通信公司 | Graphic data mining method for monitoring mass information on line by power equipment |
| CN113761904A (en) * | 2020-06-05 | 2021-12-07 | 阿里巴巴集团控股有限公司 | Training method and device of text recognition model, electronic equipment and storage medium |
| US11455466B2 (en) | 2019-05-01 | 2022-09-27 | Microsoft Technology Licensing, Llc | Method and system of utilizing unsupervised learning to improve text to content suggestions |
| CN116416247A (en) * | 2023-06-08 | 2023-07-11 | 常州微亿智造科技有限公司 | Pre-training-based defect detection method and device |
| CN116701303A (en) * | 2023-07-06 | 2023-09-05 | 浙江档科信息技术有限公司 | Electronic file classification method, system and readable storage medium based on deep learning |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160283814A1 (en) * | 2015-03-25 | 2016-09-29 | Alibaba Group Holding Limited | Method and apparatus for generating text line classifier |
| CN106569059A (en) * | 2016-11-01 | 2017-04-19 | 广西电网有限责任公司电力科学研究院 | Production service system with function of converting test data to structured storage |
| CN107038480A (en) * | 2017-05-12 | 2017-08-11 | 东华大学 | A kind of text sentiment classification method based on convolutional neural networks |
| CN107392901A (en) * | 2017-07-24 | 2017-11-24 | 国网山东省电力公司信息通信公司 | A kind of method for transmission line part intelligence automatic identification |
| CN107491435A (en) * | 2017-08-14 | 2017-12-19 | 深圳狗尾草智能科技有限公司 | Method and device based on Computer Automatic Recognition user feeling |
| CN107808132A (en) * | 2017-10-23 | 2018-03-16 | 重庆邮电大学 | A kind of scene image classification method for merging topic model |
| CN107918772A (en) * | 2017-12-10 | 2018-04-17 | 北京工业大学 | Method for tracking target based on compressive sensing theory and gcForest |
| CN108563731A (en) * | 2018-04-08 | 2018-09-21 | 北京奇艺世纪科技有限公司 | A kind of sensibility classification method and device |
| CN108898138A (en) * | 2018-05-30 | 2018-11-27 | 西安理工大学 | Scene text recognition methods based on deep learning |
-
2018
- 2018-12-04 CN CN201811470666.8A patent/CN109614488B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160283814A1 (en) * | 2015-03-25 | 2016-09-29 | Alibaba Group Holding Limited | Method and apparatus for generating text line classifier |
| CN106569059A (en) * | 2016-11-01 | 2017-04-19 | 广西电网有限责任公司电力科学研究院 | Production service system with function of converting test data to structured storage |
| CN107038480A (en) * | 2017-05-12 | 2017-08-11 | 东华大学 | A kind of text sentiment classification method based on convolutional neural networks |
| CN107392901A (en) * | 2017-07-24 | 2017-11-24 | 国网山东省电力公司信息通信公司 | A kind of method for transmission line part intelligence automatic identification |
| CN107491435A (en) * | 2017-08-14 | 2017-12-19 | 深圳狗尾草智能科技有限公司 | Method and device based on Computer Automatic Recognition user feeling |
| CN107808132A (en) * | 2017-10-23 | 2018-03-16 | 重庆邮电大学 | A kind of scene image classification method for merging topic model |
| CN107918772A (en) * | 2017-12-10 | 2018-04-17 | 北京工业大学 | Method for tracking target based on compressive sensing theory and gcForest |
| CN108563731A (en) * | 2018-04-08 | 2018-09-21 | 北京奇艺世纪科技有限公司 | A kind of sensibility classification method and device |
| CN108898138A (en) * | 2018-05-30 | 2018-11-27 | 西安理工大学 | Scene text recognition methods based on deep learning |
Non-Patent Citations (2)
| Title |
|---|
| MILOŠ ORAVEC ET AL.: "Feature extraction and classification by machine learning methods for biometric recognition of face and iris", 《PROCEEDINGS ELMAR-2014》 * |
| 徐文达: "浅析配网10kV带电作业的项目分类", 《科技与企业》 * |
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020220370A1 (en) * | 2019-05-01 | 2020-11-05 | Microsoft Technology Licensing, Llc | Method and system of utilizing unsupervised learning to improve text to content suggestions |
| US11429787B2 (en) | 2019-05-01 | 2022-08-30 | Microsoft Technology Licensing, Llc | Method and system of utilizing unsupervised learning to improve text to content suggestions |
| US11455466B2 (en) | 2019-05-01 | 2022-09-27 | Microsoft Technology Licensing, Llc | Method and system of utilizing unsupervised learning to improve text to content suggestions |
| CN110490105A (en) * | 2019-08-06 | 2019-11-22 | 南京大国科技有限公司 | Distribute-electricity transformer district acceptance method, device and computer storage medium based on image recognition |
| CN110908901B (en) * | 2019-11-11 | 2023-05-02 | 福建天晴数码有限公司 | Automatic verification method and system for image recognition capability |
| CN110908901A (en) * | 2019-11-11 | 2020-03-24 | 福建天晴数码有限公司 | Automatic verification method and system for image recognition capability |
| CN112784652A (en) * | 2019-11-11 | 2021-05-11 | 中强光电股份有限公司 | Image recognition method and device |
| CN112784652B (en) * | 2019-11-11 | 2024-08-13 | 中强光电股份有限公司 | Image recognition method and device |
| CN111026870A (en) * | 2019-12-11 | 2020-04-17 | 华北电力大学 | ICT system fault analysis method integrating text classification and image recognition |
| CN113761904A (en) * | 2020-06-05 | 2021-12-07 | 阿里巴巴集团控股有限公司 | Training method and device of text recognition model, electronic equipment and storage medium |
| CN112348066A (en) * | 2020-10-28 | 2021-02-09 | 国网浙江省电力有限公司绍兴供电公司 | Line uninterrupted power rating evaluation method based on gray clustering algorithm |
| CN113742508B (en) * | 2021-07-30 | 2023-09-08 | 国网河南省电力公司信息通信公司 | Graphic data mining method for monitoring mass information of power equipment on line |
| CN113742508A (en) * | 2021-07-30 | 2021-12-03 | 国网河南省电力公司信息通信公司 | Graphic data mining method for monitoring mass information on line by power equipment |
| CN116416247A (en) * | 2023-06-08 | 2023-07-11 | 常州微亿智造科技有限公司 | Pre-training-based defect detection method and device |
| CN116701303A (en) * | 2023-07-06 | 2023-09-05 | 浙江档科信息技术有限公司 | Electronic file classification method, system and readable storage medium based on deep learning |
| CN116701303B (en) * | 2023-07-06 | 2024-03-12 | 浙江档科信息技术有限公司 | Electronic file classification method, system and readable storage medium based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109614488B (en) | 2022-12-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109614488A (en) | Distribution network live line work condition distinguishing method based on text classification and image recognition | |
| CN110598736B (en) | Power equipment infrared image fault positioning, identifying and predicting method | |
| CN114092832B (en) | High-resolution remote sensing image classification method based on parallel hybrid convolutional network | |
| CN111325347A (en) | Automatic danger early warning description generation method based on interpretable visual reasoning model | |
| CN107818302A (en) | Non-rigid multi-scale object detection method based on convolutional neural network | |
| CN106504233A (en) | Image electric power widget recognition methodss and system are patrolled and examined based on the unmanned plane of Faster R CNN | |
| CN110322564B (en) | Three-dimensional model construction method suitable for VR/AR transformer substation operation environment | |
| CN110321891A (en) | A kind of big infusion medical fluid foreign matter object detection method of combined depth neural network and clustering algorithm | |
| CN111444939A (en) | Small-scale equipment component detection method based on weak supervision cooperative learning in open scene of power field | |
| CN108121702A (en) | Mathematics subjective item reads and appraises method and system | |
| CN114387499A (en) | Island coastal wetland waterfowl identification method, distribution query system and medium | |
| CN111402224A (en) | Target identification method for power equipment | |
| CN112883801A (en) | Accurate identification method for household distribution box system diagram subgraph in CAD distribution system diagram | |
| CN111209864A (en) | Target identification method for power equipment | |
| CN108170823B (en) | Hand-drawn interactive three-dimensional model retrieval method based on high-level semantic attribute understanding | |
| CN109583499B (en) | Power transmission line background object classification system based on unsupervised SDAE network | |
| CN114283285A (en) | Cross consistency self-training remote sensing image semantic segmentation network training method and device | |
| CN109840905A (en) | Power equipment rusty stain detection method and system | |
| CN116070523A (en) | Simulation construction method for substation defect detection image sample | |
| CN105354591B (en) | Three-dimensional outdoor scene semantic segmentation system based on high-order classification correlation priori knowledge | |
| CN111507249A (en) | Transformer substation nest identification method based on target detection | |
| Feng et al. | Gabor-YOLONet: A lightweight and efficient detection network for low-voltage power lines from unmanned aerial vehicle images | |
| CN113469938A (en) | Pipe gallery video analysis method and system based on embedded front-end processing server | |
| CN113191320A (en) | Food safety index monitoring method based on deep learning | |
| Tian et al. | Feature fusion–based preprocessing for steel plate surface defect recognition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |