CN108334842A - A method of identification pantograph-catenary current collection arcing size - Google Patents
A method of identification pantograph-catenary current collection arcing size Download PDFInfo
- Publication number
- CN108334842A CN108334842A CN201810107109.3A CN201810107109A CN108334842A CN 108334842 A CN108334842 A CN 108334842A CN 201810107109 A CN201810107109 A CN 201810107109A CN 108334842 A CN108334842 A CN 108334842A
- Authority
- CN
- China
- Prior art keywords
- arcing
- model
- unit number
- size
- layer
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Abstract
The present invention proposes that a kind of method of identification pantograph-catenary current collection arcing size, step include：Data set picture is obtained, and data set picture is generated into training set and test set；Arcing model is established, arcing model is convolutional neural networks model；Arcing model is trained using training set picture, adjustment and Optimized model parameter, the arcing model optimized；The picture of test set is identified using the arcing model of optimization.Method proposed by the present invention classifies to pantograph-catenary current collection arcing picture using arcing model, judge arcing size, compared to existing method, parameter amount and calculation amount when greatly can reduce trained, save the consumption of resource and time, higher accuracy also can guarantee by the intensification of network simultaneously so that be achieved to pantograph-catenary current collection arcing size progress automatic identification and adapt to practical application in industry.
Description
Technical field
The present invention relates to high speed railway vehicle mounted contact net condition monitoring technical field more particularly to a kind of identification bow nets
By the method for stream arcing size.
Background technology
With the construction and development present situation of high-speed railway, contact net arcing has been increasingly becoming restriction traction power supply contact net can
By the key factor of property, no matter consider from power supply reliability and the service life of contact net equipment, the detection of contact net arcing with it is pre-
The theoretical research of anti-measure is all imperative.
According to Electric contact theory it is found that pantograph-contact net current collecting system, which is typical sliding, is in electrical contact model, in train
During high-speed cruising, the generation of the bad contact phenomena of bow net is inevitable.Pantograph-OCS system vibration, slide plate or contact line are different
The factors such as object, contact line defect such as hard spot will likely all cause the generation of the bad contact phenomena of bow net, and Pantograph-OCS system is bad
Contact is usually associated with strong electric discharge phenomena between pantograph and contact net --- the generation of bow net arcing.The size of arcing is carried out
Identification can react the severity of strong electric discharge phenomena between pantograph and contact net to a certain extent, for especially severe
Phenomenon preferentially reports, preferential to solve.
Existing classics convolutional neural networks have AlexNet, VGGNet, Google InceptionNet, ResNet ---
This in four network arranged according to the sequencing of appearance, depth and complexity are also progressive successively.When these existing models solve
When industrial existing picture Question Classification, although accuracy performance is well, often since parameter amount is too big, model mistake
Cause the resource of consumption more in complexity, the trained time is longer, this causes these classical convolutional neural networks can not be in reality
It is used in the industry of border.
Invention content
To solve the above-mentioned problems, the present invention proposes a kind of method of identification pantograph-catenary current collection arcing size, used combustion
Parameter amount and calculation amount when arc model greatly can reduce trained, save the consumption of resource and time, while passing through network
Intensification also can guarantee higher accuracy, can adapt to actual industrial utilization.
Specifically, a kind of method of identification pantograph-catenary current collection arcing size, which is characterized in that include the following steps：
S1：Data set picture is obtained, selected part picture falls into 5 types according to arcing size in the data set picture, will be every
Pictures are tagged according to its classification, training set generated, using remaining picture in the data set as test set；
S2：Arcing model is established, the arcing model is convolutional neural networks model；
S3：The arcing model is trained using the training set picture, adjustment and Optimized model parameter are optimized
Arcing model；
S4：The picture of the test set is identified using the arcing model of the optimization.
Preferably, data set picture described in the S1 steps is that vehicle-mounted contact net running state detecting device is transported in train
Row claps the pantograph-catenary current collection arcing picture taken on the way.
Preferably, the utilization of the S3 steps training set picture is trained the arcing model, adjusts and excellent
Change the model parameter, the arcing model optimized is specially：It is in training process that the image data in the training set is defeated
Enter arcing model and carry out propagated forward, obtain the output valve of "current" model, by the output valve and training set that reduce "current" model
The penalty values of picture concrete class optimize the arcing model, and then adjust the parameter in arcing model, have trained
Finish the arcing model optimized.
Preferably, using the arcing model of the optimization tool is identified to the picture of the test set in the S4 steps
Body is the input test collection picture in the arcing model of optimization, carries out propagated forward, is divided using trained model parameter
Class is as a result, realize the identification to pantograph-catenary current collection arcing size.
Preferably, the arcing size falls into 5 types specially：It is smaller, it is small, in, it is larger, greatly.
Preferably, the arcing model shares 11 convolutional layers, 3 pond layers, and 2 local acknowledgements normalize layer, and 3 complete
Articulamentum, sequence are followed successively by：First convolutional layer（201）, the second convolutional layer（202）, First partial response normalization layer
（203）, the first pond layer（204）, third convolutional layer（205）, Volume Four lamination（206）, the second local acknowledgement normalize layer
（207）, the second pond layer（208）, the 5th convolutional layer（209）, the 6th convolutional layer（210）, the 7th convolutional layer（211）, volume eight
Lamination（212）, the 9th convolutional layer（213）, the tenth convolutional layer（214）, third pond layer（215）, the first full articulamentum（216）、
Second full articulamentum（217）, the full articulamentum of third（218）.
Preferably, first convolutional layer（201）2 size specifications used are 7*7, and the convolution kernel that step-length is 2 carries out
Feature extraction, characteristic pattern unit number are converted into 112*112*2 by the 224*224*1 of input picture specification；
Second convolutional layer（202）2 size specifications used are 7*7, and the convolution kernel that step-length is 2 carries out feature extraction, special
Sign figure unit number is converted into 56*56*2 by 112*112*2；
The third convolutional layer（205）4 size specifications used are 3*3, and the convolution kernel that step-length is 1 carries out feature extraction, special
Sign figure unit number is converted into 27*27*4 by 27*27*2；
The Volume Four lamination（206）8 size specifications used are 3*3, and the convolution kernel that step-length is 1 carries out feature extraction, special
Sign figure unit number is converted into 27*27*8 by 27*27*4；
5th convolutional layer（209）16 size specifications used are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction,
Characteristic pattern unit number is converted into 13*13*16 by 13*13*8；
6th convolutional layer（210）32 size specifications used are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction,
Characteristic pattern unit number is converted into 13*13*32 by 13*13*16；
7th convolutional layer（211）32 size specifications used are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction,
Characteristic pattern unit number is converted into 13*13*32 by 13*13*32；
8th convolutional layer（212）16 size specifications used are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction,
Characteristic pattern unit number is converted into 13*13*16 by 13*13*32；
9th convolutional layer（213）8 size specifications used are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction, special
Sign figure unit number is converted into 13*13*8 by 13*13*16；
Tenth convolutional layer（214）4 size specifications used are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction, special
Sign figure unit number is converted into 13*13*4 by 13*13*8；
Preferably, the First partial response normalization layer（203）, second local acknowledgement normalize layer（207）It is used to
Competition mechanism is created to the activity of local neuron, the bigger value that wherein responds made becomes relatively large, and inhibits other
The smaller neuron of feedback, enhances the generalization ability of model.
Preferably, first pond layer（204）The characteristic pattern of input is compressed, characteristic pattern unit number is by 56*56*
2 are converted into 27*27*2；
Second pond layer（208）The characteristic pattern of input is compressed, characteristic pattern unit number is converted into 13* by 27*27*8
13*8；
Third pond layer（215）The characteristic pattern of input is compressed, characteristic pattern unit number is converted into 6*6* by 13*13*4
4。
Preferably, the described first full articulamentum（216）Input feature vector figure unit number be 6*6*4, export characteristic pattern unit
Number is 64；
The second full articulamentum（217）Input feature vector figure unit number be 64, output characteristic pattern unit number be 16；
The full articulamentum of third（218）Input feature vector figure unit number be 16, output characteristic pattern unit number be 5.
The beneficial effects of the present invention are：
（1）Method proposed by the present invention can carry out automatic identification to pantograph-catenary current collection arcing size, can using the identification data
The severity of strong electric discharge phenomena is assessed between pantograph and contact net, related operation personnel can according to assessment result and
When take measures, the operation security of effective guarantee high-speed railway.
（2）Method proposed by the present invention classifies to pantograph-catenary current collection arcing picture using arcing model, judges that arcing is big
Small, compared to existing method, parameter amount and calculation amount when greatly can reduce trained save the consumption of resource and time,
Higher accuracy also can guarantee by the intensification of network simultaneously so that automatic identification is carried out to pantograph-catenary current collection arcing size and is able to
It realizes and adapts to practical application in industry.
Description of the drawings
Fig. 1 is a kind of step flow chart of the method for identification pantograph-catenary current collection arcing size proposed by the present invention.
Fig. 2 is arcing model structure proposed by the present invention.
Fig. 3 is the AlexNet model structures of the prior art.
Fig. 4 be using it is proposed by the present invention it is a kind of identification pantograph-catenary current collection arcing size method judge showing for arcing size
Example.
Specific implementation mode
For a clearer understanding of the technical characteristics, objects and effects of the present invention, now control illustrates this hair
Bright specific implementation mode.
Fig. 1 is a kind of method and step flow chart of identification pantograph-catenary current collection arcing size proposed by the present invention, including walks as follows
Suddenly：
S1：Data set picture is obtained, selected part picture falls into 5 types according to arcing size in data set picture, and every is schemed
Piece is tagged according to its classification, training set is generated, using remaining picture in data set as test set；
S2：Arcing model is established, arcing model is convolutional neural networks model；
S3：Arcing model is trained using training set picture, adjustment and Optimized model parameter, the arcing mould optimized
Type；
S4：The picture of test set is identified using the arcing model of optimization.
As a kind of preferred embodiment, data set picture described in the S1 steps is vehicle-mounted contact net condition monitoring
Device claps the pantograph-catenary current collection arcing picture taken in train operation on the way.
As a kind of preferred embodiment, arcing size falls into 5 types specially：It is smaller, it is small, in, it is larger, greatly.
As a kind of preferred embodiment, being trained to arcing model using training set picture for S3 steps is adjusted and excellent
Change model parameter, the arcing model optimized is specially：The image data in training set is inputted into arcing mould in training process
Type carries out propagated forward, obtains the output valve of "current" model, practical by the output valve and training set picture that reduce "current" model
The penalty values of classification optimize arcing model, and then adjust the parameter in arcing model, and training, which finishes, to be optimized
Arcing model.
As a kind of preferred embodiment, tool is identified to the picture of test set in the arcing model using optimization of S4 steps
Body is the input test collection picture in the arcing model of optimization, carries out propagated forward, is divided using trained model parameter
Class is as a result, realize the identification to pantograph-catenary current collection arcing size.
As a kind of preferred embodiment, arcing model structure proposed by the present invention is as shown in Figure 2.Arcing model shares 11
Convolutional layer, 3 pond layers, 2 local acknowledgements normalize layer, 3 full articulamentums, and sequence is followed successively by：First convolutional layer 201,
Second convolutional layer 202, First partial response normalization layer 203, the first pond layer 204, third convolutional layer 205, Volume Four lamination
206, the second local acknowledgement normalization layer 207, the second pond layer 208, the 5th convolutional layer 209, the 6th convolutional layer 210, volume seven
Lamination 211, the 8th convolutional layer 212, the 9th convolutional layer 213, the tenth convolutional layer 214, third pond layer 215, the first full articulamentum
216, the full articulamentum of the second full articulamentum 217, third 218.
Wherein, the effect of convolutional layer (Conv) is：Feature extraction is carried out with it.
Pond layer（MaxPooling）Effect be：The characteristic pattern of input is compressed, on the one hand characteristic pattern is made to become smaller,
Simplify network calculations complexity；On the one hand Feature Compression is carried out, main feature is extracted.
Local acknowledgement normalizes layer（Local Response Norm, LRN）Effect be：Activity to local neuron
Competition mechanism is created, the bigger value that wherein responds made becomes relatively large, and other is inhibited to feed back smaller neuron, increases
The generalization ability of strong model.
Full articulamentum（FC）Effect be：Play the role of in entire convolutional neural networks " grader ".
As a kind of preferred embodiment, 2 size specifications that the first convolutional layer 201 uses are 7*7, the convolution that step-length is 2
Core carries out feature extraction, and characteristic pattern unit number is converted into 112*112*2 by 224*224*1.Wherein, 224*224*1 is to input
The specification of image（The port number of the high * images of the wide * images of image）, input picture is using gray level image, channel 1.
2 size specifications that second convolutional layer 202 uses are 7*7, and the convolution kernel that step-length is 2 carries out feature extraction, feature
Figure unit number is converted into 56*56*2 by 112*112*2.
4 size specifications that third convolutional layer 205 uses are 3*3, and the convolution kernel that step-length is 1 carries out feature extraction, feature
Figure unit number is converted into 27*27*4 by 27*27*2.
8 size specifications that Volume Four lamination 206 uses are 3*3, and the convolution kernel that step-length is 1 carries out feature extraction, feature
Figure unit number is converted into 27*27*8 by 27*27*4.
16 size specifications that 5th convolutional layer 209 uses are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction, special
Sign figure unit number is converted into 13*13*16 by 13*13*8.
32 size specifications that 6th convolutional layer 210 uses are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction, special
Sign figure unit number is converted into 13*13*32 by 13*13*16.
32 size specifications that 7th convolutional layer 211 uses are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction, special
Sign figure unit number is converted into 13*13*32 by 13*13*32.
16 size specifications that 8th convolutional layer 212 uses are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction, special
Sign figure unit number is converted into 13*13*16 by 13*13*32.
8 size specifications that 9th convolutional layer 213 uses are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction, feature
Figure unit number is converted into 13*13*8 by 13*13*16.
4 size specifications that tenth convolutional layer 214 uses are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction, feature
Figure unit number is converted into 13*13*4 by 13*13*8.
As a kind of preferred embodiment, First partial response normalization layer 203, the second local acknowledgement normalization layer 207 are equal
For creating competition mechanism to the activity of local neuron, the bigger value that wherein responds made becomes relatively large, and inhibits
Other feed back smaller neuron, enhance the generalization ability of model.
As a kind of preferred embodiment, the first pond layer 204, the second pond layer 208, third pond layer 215 are maximum
The specification of pond layer, pond core is 3*3, step-length 2.The setting of maximum pond layer avoids the blurring effect in average pond
Fruit, and make the size of step-length ratio Chi Huahe small, overlapping and covering are had between the output of such pond layer, improves feature
It is rich.
The characteristic pattern of first 204 pairs of pond layer input compresses, and characteristic pattern unit number is converted into 27*27* by 56*56*2
2。
The characteristic pattern of second 208 pairs of pond layer input compresses, and characteristic pattern unit number is converted into 13*13* by 27*27*8
8。
The characteristic pattern of 215 pairs of input of third pond layer compresses, and characteristic pattern unit number is converted into 6*6*4 by 13*13*4.
As a kind of preferred embodiment, the input feature vector figure unit number of the first full articulamentum 216 is 6*6*4, exports feature
Figure unit number is 64.
The input feature vector figure unit number of second full articulamentum 217 is 64, and output characteristic pattern unit number is 16.
The input feature vector figure unit number of the full articulamentum of third 218 is 16, and output characteristic pattern unit number is 5, that is, represents arcing
Classification total number, a total of 5 class, respectively representing arcing size is：It is smaller, it is small, in, it is larger, greatly.
The arcing model that the embodiment of the present invention proposes, compared to existing Alexnet models（As shown in Figure 3）, Neng Gou great
Parameter amount and calculation amount when big reduction is trained.For example, when a gray-scale map inputs, the parameter amount of Alexnet MODEL Cs onv1
For（11*11+1）* 96=11712, arcing MODEL C onv1 parameter amounts proposed by the present invention are：（7*7+1）*2=100.As a result, originally
The arcing model that invention proposes saves the consumption of resource and time, while also can guarantee by the intensification of neural network higher
Accuracy.
The method that the embodiment of the present invention proposes can carry out automatic identification to pantograph-catenary current collection arcing size, utilize the identification number
According to the severity of strong electric discharge phenomena can be assessed between pantograph and contact net, related operation personnel can be according to assessment
As a result it takes timely measure, the operation security of effective guarantee high-speed railway；The method that the embodiment of the present invention proposes uses arcing mould
Type classifies to pantograph-catenary current collection arcing picture, judges arcing size, compared to existing method, when greatly can reduce trained
Parameter amount and calculation amount save the consumption of resource and time, while also can guarantee higher accuracy by the intensification of network,
So that being achieved to pantograph-catenary current collection arcing size progress automatic identification and adapting to practical application in industry.
It should be noted that for each embodiment of the method above-mentioned, for simple description, therefore it is all expressed as to a system
The combination of actions of row, but those skilled in the art should understand that, the application is not limited by the described action sequence, because
For according to the application, certain some step can be performed in other orders or simultaneously.Secondly, those skilled in the art also should
Know, embodiment described in this description belongs to preferred embodiment, involved action and unit not necessarily this Shen
It please be necessary.
In the above-described embodiments, it all emphasizes particularly on different fields to the description of each embodiment, is not described in some embodiment
Part, may refer to the associated description of other embodiment.
One of ordinary skill in the art will appreciate that realizing all or part of flow in above-described embodiment method, being can be with
Relevant hardware is instructed to complete by computer program, the program can be stored in computer read/write memory medium
In, the program is when being executed, it may include such as the flow of the embodiment of above-mentioned each method.Wherein, the storage medium can be magnetic
Dish, CD, ROM, RAM etc..
The above disclosure is only the preferred embodiments of the present invention, cannot limit the right model of the present invention with this certainly
It encloses, therefore equivalent changes made in accordance with the claims of the present invention, is still within the scope of the present invention.
Claims (10)
1. a kind of method of identification pantograph-catenary current collection arcing size, which is characterized in that include the following steps：
S1：Data set picture is obtained, selected part picture falls into 5 types according to arcing size in the data set picture, will be every
Pictures are tagged according to its classification, training set generated, using remaining picture in the data set as test set；
S2：Arcing model is established, the arcing model is convolutional neural networks model；
S3：The arcing model is trained using the training set picture, adjustment and Optimized model parameter are optimized
Arcing model；
S4：The picture of the test set is identified using the arcing model of the optimization.
2. a kind of method of identification pantograph-catenary current collection arcing size as described in claim 1, which is characterized in that in the S1 steps
The data set picture is that vehicle-mounted contact net running state detecting device claps the pantograph-catenary current collection arcing figure taken in train operation on the way
Piece.
3. a kind of method of identification pantograph-catenary current collection arcing size as described in claim 1, which is characterized in that the S3 steps
The arcing model is trained using the training set picture, adjusts and optimize the model parameter, the combustion optimized
Arc model is specially：The image data input arcing model in the training set is subjected to propagated forward in training process, is obtained
The output valve of "current" model, by reducing the output valve of "current" model and the penalty values of training set picture concrete class come to described
Arcing model optimizes, and then adjusts the parameter in arcing model, and training finishes the arcing model optimized.
4. a kind of method of identification pantograph-catenary current collection arcing size as described in claim 1, which is characterized in that the S4 steps
The picture of the test set is identified using the arcing model of the optimization specifically, being inputted in the arcing model of optimization
Test set picture carries out propagated forward, obtains classification results using trained model parameter, realizes big to pantograph-catenary current collection arcing
Small identification.
5. a kind of method of identification pantograph-catenary current collection arcing size as described in claim 1, which is characterized in that the arcing size
It falls into 5 types specially：It is smaller, it is small, in, it is larger, greatly.
6. a kind of method of identification pantograph-catenary current collection arcing size as described in claim 1, which is characterized in that the arcing model
11 convolutional layers, 3 pond layers are shared, 2 local acknowledgements normalize layer, 3 full articulamentums, and sequence is followed successively by：The first volume
Lamination（201）, the second convolutional layer（202）, First partial response normalization layer（203）, the first pond layer（204）, third convolution
Layer（205）, Volume Four lamination（206）, the second local acknowledgement normalize layer（207）, the second pond layer（208）, the 5th convolutional layer
（209）, the 6th convolutional layer（210）, the 7th convolutional layer（211）, the 8th convolutional layer（212）, the 9th convolutional layer（213）, volume ten
Lamination（214）, third pond layer（215）, the first full articulamentum（216）, the second full articulamentum（217）, the full articulamentum of third
（218）.
7. a kind of method of identification pantograph-catenary current collection arcing size as claimed in claim 6, which is characterized in that first convolution
Layer（201）2 size specifications used are 7*7, and the convolution kernel that step-length is 2 carries out feature extraction, and characteristic pattern unit number is by inputting
The 224*224*1 of picture specification is converted into 112*112*2；
Second convolutional layer（202）2 size specifications used are 7*7, and the convolution kernel that step-length is 2 carries out feature extraction, special
Sign figure unit number is converted into 56*56*2 by 112*112*2；
The third convolutional layer（205）4 size specifications used are 3*3, and the convolution kernel that step-length is 1 carries out feature extraction, special
Sign figure unit number is converted into 27*27*4 by 27*27*2；
The Volume Four lamination（206）8 size specifications used are 3*3, and the convolution kernel that step-length is 1 carries out feature extraction, special
Sign figure unit number is converted into 27*27*8 by 27*27*4；
5th convolutional layer（209）16 size specifications used are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction,
Characteristic pattern unit number is converted into 13*13*16 by 13*13*8；
6th convolutional layer（210）32 size specifications used are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction,
Characteristic pattern unit number is converted into 13*13*32 by 13*13*16；
7th convolutional layer（211）32 size specifications used are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction,
Characteristic pattern unit number is converted into 13*13*32 by 13*13*32；
8th convolutional layer（212）16 size specifications used are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction,
Characteristic pattern unit number is converted into 13*13*16 by 13*13*32；
9th convolutional layer（213）8 size specifications used are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction, special
Sign figure unit number is converted into 13*13*8 by 13*13*16；
Tenth convolutional layer（214）4 size specifications used are 2*2, and the convolution kernel that step-length is 1 carries out feature extraction, special
Sign figure unit number is converted into 13*13*4 by 13*13*8.
8. a kind of method of identification pantograph-catenary current collection arcing size as claimed in claim 6, which is characterized in that the First partial
Response normalization layer（203）, second local acknowledgement normalize layer（207）The activity establishment being used to local neuron is competing
Mechanism is striven, the bigger value that wherein responds made becomes relatively large, and other is inhibited to feed back smaller neuron, enhances model
Generalization ability.
9. a kind of method of identification pantograph-catenary current collection arcing size as claimed in claim 6, which is characterized in that first pond
Layer（204）The characteristic pattern of input is compressed, characteristic pattern unit number is converted into 27*27*2 by 56*56*2；
Second pond layer（208）The characteristic pattern of input is compressed, characteristic pattern unit number is converted into 13* by 27*27*8
13*8；
Third pond layer（215）The characteristic pattern of input is compressed, characteristic pattern unit number is converted into 6*6* by 13*13*4
4。
10. a kind of method of identification pantograph-catenary current collection arcing size as claimed in claim 6, which is characterized in that described first is complete
Articulamentum（216）Input feature vector figure unit number be 6*6*4, output characteristic pattern unit number be 64；
The second full articulamentum（217）Input feature vector figure unit number be 64, output characteristic pattern unit number be 16；
The full articulamentum of third（218）Input feature vector figure unit number be 16, output characteristic pattern unit number be 5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810107109.3A CN108334842A (en) | 2018-02-02 | 2018-02-02 | A method of identification pantograph-catenary current collection arcing size |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810107109.3A CN108334842A (en) | 2018-02-02 | 2018-02-02 | A method of identification pantograph-catenary current collection arcing size |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108334842A true CN108334842A (en) | 2018-07-27 |
Family
ID=62927852
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810107109.3A Pending CN108334842A (en) | 2018-02-02 | 2018-02-02 | A method of identification pantograph-catenary current collection arcing size |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108334842A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109886302A (en) * | 2019-01-21 | 2019-06-14 | 河北新兴铸管有限公司 | Caliber judgment method and terminal device based on machine learning |
| CN110378897A (en) * | 2019-07-25 | 2019-10-25 | 中车青岛四方机车车辆股份有限公司 | A kind of pantograph running state real-time monitoring method and device based on video |
| CN110763958A (en) * | 2019-09-23 | 2020-02-07 | 华为技术有限公司 | Direct current arc detection method, device, equipment, system and storage medium |
| CN112766195A (en) * | 2021-01-26 | 2021-05-07 | 西南交通大学 | Electrified railway bow net arcing visual detection method |
| KR102276634B1 (en) * | 2020-09-15 | 2021-07-13 | 엠아이엠테크 주식회사 | System for detecting abnormality of pantograph on electric train installed on vehicle and method for processing thereof |
| CN113255507A (en) * | 2021-05-20 | 2021-08-13 | 株洲中车时代电气股份有限公司 | Identification method for bow net arcing and related device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110222724A1 (en) * | 2010-03-15 | 2011-09-15 | Nec Laboratories America, Inc. | Systems and methods for determining personal characteristics |
| CN105469420A (en) * | 2016-01-25 | 2016-04-06 | 成都国铁电气设备有限公司 | Overhead line-pantograph arcing defect identification method and system thereof |
| US20170249339A1 (en) * | 2016-02-25 | 2017-08-31 | Shutterstock, Inc. | Selected image subset based search |
| CN107167726A (en) * | 2017-05-12 | 2017-09-15 | 清华大学 | A kind of circuit breaker internal puncture electric arc modeling method |
-
2018
- 2018-02-02 CN CN201810107109.3A patent/CN108334842A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110222724A1 (en) * | 2010-03-15 | 2011-09-15 | Nec Laboratories America, Inc. | Systems and methods for determining personal characteristics |
| CN105469420A (en) * | 2016-01-25 | 2016-04-06 | 成都国铁电气设备有限公司 | Overhead line-pantograph arcing defect identification method and system thereof |
| US20170249339A1 (en) * | 2016-02-25 | 2017-08-31 | Shutterstock, Inc. | Selected image subset based search |
| CN107167726A (en) * | 2017-05-12 | 2017-09-15 | 清华大学 | A kind of circuit breaker internal puncture electric arc modeling method |
Non-Patent Citations (2)
| Title |
|---|
| ALEX KRIZHEVSKY等: ""ImageNet Classification with Deep Convolutional Neural Networks"", 《COMMUNICATIONS OF THE ACM》 * |
| 杨恒: ""基于机器视觉的高速列车弓网动态性能检测方法研究"", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109886302A (en) * | 2019-01-21 | 2019-06-14 | 河北新兴铸管有限公司 | Caliber judgment method and terminal device based on machine learning |
| CN110378897A (en) * | 2019-07-25 | 2019-10-25 | 中车青岛四方机车车辆股份有限公司 | A kind of pantograph running state real-time monitoring method and device based on video |
| CN110763958A (en) * | 2019-09-23 | 2020-02-07 | 华为技术有限公司 | Direct current arc detection method, device, equipment, system and storage medium |
| KR102276634B1 (en) * | 2020-09-15 | 2021-07-13 | 엠아이엠테크 주식회사 | System for detecting abnormality of pantograph on electric train installed on vehicle and method for processing thereof |
| CN112766195A (en) * | 2021-01-26 | 2021-05-07 | 西南交通大学 | Electrified railway bow net arcing visual detection method |
| CN112766195B (en) * | 2021-01-26 | 2022-03-29 | 西南交通大学 | Electrified railway bow net arcing visual detection method |
| CN113255507A (en) * | 2021-05-20 | 2021-08-13 | 株洲中车时代电气股份有限公司 | Identification method for bow net arcing and related device |
| CN113255507B (en) * | 2021-05-20 | 2022-04-29 | 株洲中车时代电气股份有限公司 | Identification method, system and storage medium for bow net arcing |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108334842A (en) | A method of identification pantograph-catenary current collection arcing size | |
| KR102419136B1 (en) | Image processing apparatus and method using multiple-channel feature map | |
| US20180053091A1 (en) | System and method for model compression of neural networks for use in embedded platforms | |
| CN108665063B (en) | Bidirectional parallel processing convolution acceleration system for BNN hardware accelerator | |
| CN107679465A (en) | A kind of pedestrian's weight identification data generation and extending method based on generation network | |
| CN109635791B (en) | Video evidence obtaining method based on deep learning | |
| CN107229942A (en) | A kind of convolutional neural networks rapid classification method based on multiple graders | |
| CN103886623A (en) | Image compression method and equipment, and system | |
| CN110436294A (en) | A kind of battery truck enters elevator detection method | |
| CN107770525A (en) | A kind of method and device of Image Coding | |
| WO2024012574A1 (en) | Image coding method and apparatus, image decoding method and apparatus, readable medium, and electronic device | |
| CN112489164A (en) | Image coloring method based on improved depth separable convolutional neural network | |
| CN116310785A (en) | Unmanned aerial vehicle image pavement disease detection method based on YOLO v4 | |
| CN110796623A (en) | Infrared image rain removing method and device based on progressive residual error network | |
| CN114882278A (en) | Tire pattern classification method and device based on attention mechanism and transfer learning | |
| CN109508639B (en) | Road scene semantic segmentation method based on multi-scale porous convolutional neural network | |
| CN110852272B (en) | Pedestrian detection method | |
| CN109086819A (en) | Caffemodel model compression method, system, equipment and medium | |
| CN116993737A (en) | Lightweight fracture segmentation method based on convolutional neural network | |
| CN116523875A (en) | Insulator defect detection method based on FPGA pretreatment and improved YOLOv5 | |
| CN116977256A (en) | Training method, device, equipment and storage medium for defect detection model | |
| CN114821346A (en) | Radar image intelligent identification method and system based on embedded platform | |
| CN114461538A (en) | Cloud computing application memory management method based on real-time content prediction and historical resource occupation | |
| CN114386578A (en) | Convolution neural network method implemented on Haisi non-NPU hardware | |
| CN113989601A (en) | Feature fusion network, sample selection method, target detection method and device |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180727 |