CN108573277A - A kind of pantograph carbon slide surface disease automatic recognition system and method - Google Patents
A kind of pantograph carbon slide surface disease automatic recognition system and method Download PDFInfo
- Publication number
- CN108573277A CN108573277A CN201810200743.1A CN201810200743A CN108573277A CN 108573277 A CN108573277 A CN 108573277A CN 201810200743 A CN201810200743 A CN 201810200743A CN 108573277 A CN108573277 A CN 108573277A
- Authority
- CN
- China
- Prior art keywords
- network
- image
- parameter
- layer
- module
- 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
- 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
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/08—Detecting or categorising vehicles
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention provides a kind of pantograph carbon slide surface disease automatic recognition system and methods, including：Acquisition module；Processing module；Modeling module；Training module；Transfer module.The present invention provides a kind of pantograph carbon slide surface disease automatic recognition system and methods, the autonomous learning to image data and feature extraction by network eliminate complicated flows such as a series of images pretreatment image enhancing edge detection feature extraction target identification of traditional means；The influence of the variations such as network makes model that can not be illuminated by the light the identification of disease advanced features the non-linear fusion of low-level features, displacement, scale has better robustness；Model is completed once training, you can is directly used in image recognition, and training manually carries out complex operations without relying on using process to image and model, has higher intellectually and automatically degree.
Description
Technical field
The present invention relates to rail traffic vehicles Diagnosis Technique fields.More particularly, to a kind of pantograph carbon
Sled surface disease automatic recognition system and method.
Background technology
Urban track traffic goes on a journey to urban transportation and has important role to the guiding of urban development.With city rail
The fast development of road traffic, to the safety of vehicle part, more stringent requirements are proposed.Pantograph is powered as municipal rail train
The important component of system, because of its complicated working environment, strong mechanics and electrical effect always exist higher failure rate.Cause
This, the timely detection to bow failure, to ensureing that the normal safe operation of train is of great significance.There is serious event in pantograph
When barrier, it may occur that component breakage lacks, and hinders normally to be flowed between bow net, or even cause to be lost to contact line, to row
Vehicle safety causes significant damage with line facility.
The existing detection for pantograph carbon slide surface abrasion is more to be carried out by way of manually stepping on roof, although
The comprehensive identification to various faults type may be implemented, but have higher want to the technical merit of testing staff and working attitude
It asks, and there is also larger interference to driving for the maintenance model climbed to the top of a mountain of parking.Exploitation driving interference is smaller, identification function is various,
The high pantograph pan surface disease recognition method of accuracy rate is realized and carries out abrasion of pantograph pan situation in way efficiently inspection
Survey, to do not influence normal vehicle operation in the case of, ensure train part reliably with the operation security of vehicle.
Contact can be classified as according to its working method for pantograph pan surface abrasion automatic detection at present
Two kinds of detection and non-contact detection；For non-contact detection, then it can be divided into laser detection according to the difference of detection signal,
Ultrasound examination and image detection these three.Mutually compared with contact measurement, non-contact detection survey is less to the interference of driving, and
Mechanics between bow net, substantially without influence, further improves the efficiency and accuracy of detection with electrical effect.And non-contact
In formula detection means, often there is very high precision using the detection of laser, but function is more single；Ultrasound examination has
Wider function, but accuracy of detection is more insufficient.
Invention content
At least one of in order to solve the above-mentioned technical problem, one aspect of the present invention provides a kind of pantograph carbon cunning
Plate surface disease automatic recognition system, including：
Acquisition module acquires pantograph pan surface defect image；
Processing module, the processing module include：
Operation is normalized to the image data of acquisition in normalizing unit；
Image after normalization is converted to data set by converting unit；
Modeling module builds network model, the parameter of network model is arranged；
Training module trains the network model, the network model generation to preserve according to the data by data set
The file for the parameter and value information that training is completed；
Module is transferred, the file is called.
Preferably, described normalize includes：
Size normalization, color mode normalization, global characteristics standardization and image go mean value.
Preferably, the modeling module includes：
Network vision layer dispensing unit, the convolutional layer in vision layer, pond layer, Quan Lian by configuring convolutional neural networks
Layer is connect, realizes the extraction and identification of the data characteristics of data set；
Network hyper parameter dispensing unit, the hyper parameter of Configuration network model in solver.prototxt files.
Preferably, the network vision layer dispensing unit further configures in the vision layers of convolutional neural networks
Dropout layers.
Preferably, the modeling module further comprises：
Network structure dispensing unit, the parameter of Configuration network structure.
Another aspect of the present invention provides a kind of pantograph carbon slide surface disease automatic identifying method, including：
Acquisition module acquires pantograph pan surface defect image；
Operation is normalized to the image data of acquisition in the normalizing unit of processing module；
Image after normalization is converted to data set by the converting unit of processing module；
Modeling module builds network model, and the parameter of network model is arranged；
Training module trains the network model, the network model generation to preserve instruction according to the data by data set
Practice the file of the parameter and value information completed；
It transfers module and calls the file.
Preferably, the modeling module builds network model, and the parameter that network model is arranged includes：
Convolutional layer, pond layer, Quan Lian in vision layer of the network vision layer dispensing unit by configuring convolutional neural networks
Layer is connect, realizes the extraction and identification of the data characteristics of data set；
The hyper parameter of network hyper parameter dispensing unit Configuration network model in solver.prototxt files.
Preferably, the network vision layer dispensing unit further configures in the vision layers of convolutional neural networks
Dropout layers.
Preferably, the modeling module builds network model, and the parameter that network model is arranged further comprises：
Network structure dispensing unit, the parameter of Configuration network structure.
Beneficial effects of the present invention are as follows：
The present invention provides a kind of pantograph carbon slide surface disease automatic recognition system and methods, by network to image
The autonomous learning of data and feature extraction eliminate a series of images pretreatment-image enhancement-edge detection-of traditional means
Complicated flow such as feature extraction-target identification；Network makes model to disease advanced features the non-linear fusion of low-level features
Identification can not be illuminated by the light, displacement, the variations such as scale influence, there is better robustness；Model is completed once training, i.e.,
Image recognition is can be directly used for, and training manually carries out complex operations, tool without relying on using process to image and model
There is higher intellectually and automatically degree.
Description of the drawings
Specific embodiments of the present invention will be described in further detail below in conjunction with the accompanying drawings.
Fig. 1 shows a kind of pantograph carbon slide surface disease automatic identification that the embodiment of first aspect present invention provides
System structure diagram.
Fig. 2 shows the pantograph three-dimensional coordinate schematic diagrames in the embodiment of first aspect present invention.
Fig. 3 shows image data global characteristics standardization schematic diagram in the embodiment of first aspect present invention.
Fig. 4 shows that data go mean value principle and effect diagram in the embodiment of first aspect present invention.
Fig. 5 shows convolutional layer schematic diagram of connecting in the embodiment of first aspect present invention.
Fig. 6 shows dropout layers in the embodiment of first aspect present invention of algorithm principle figure.
Fig. 7 shows a kind of pantograph carbon slide surface disease provided in the embodiment using second aspect of the present invention certainly
Dynamic recognition methods flow diagram.
Fig. 8 shows one of the idiographic flow schematic diagram of S400 in Fig. 7.
Fig. 9 shows two of the idiographic flow schematic diagram of S400 in Fig. 7.
Figure 10 shows that the method provided in the embodiment of second aspect of the present invention is accurate to the test of pantograph image recognition
True rate and error change figure.
Figure 11 shows the algorithm model of the method provided in the embodiment using second aspect of the present invention completion to by electricity
Bend the top-5 result figures of sled surface disease geo-radar image identification.
Specific implementation mode
In order to illustrate more clearly of the present invention, the present invention is done further with reference to preferred embodiments and drawings
It is bright.Similar component is indicated with identical reference numeral in attached drawing.It will be appreciated by those skilled in the art that institute is specific below
The content of description is illustrative and be not restrictive, and should not be limited the scope of the invention with this.
The various sectional views that embodiment is disclosed according to the present invention are shown in the accompanying drawings.These figures are not drawn to scale
, wherein for the purpose of clear expression, some details are magnified, and some details may be omitted.It is shown in the drawings
Various regions, the shape of layer and the relative size between them, position relationship are merely exemplary, in practice may be due to system
It makes tolerance or technology restriction and is deviated, and those skilled in the art may be additionally designed as required with not similar shape
Shape, size, the regions/layers of relative position.
The high frequency time of rail traffic, big load use, and are proposed to the Stability and dependability of train pantograph higher
It is required that.The automatic fault detection without driving interference may be implemented in pantograph pan surface Defect inspection based on image.However mesh
Preceding image detecting technique need to carry out image cumbersome image preprocessing-image enhancement-edge detection-feature extraction-target
The flows such as identification, not only process is complicated, it is also necessary to and various knowledge accumulation such as a large amount of mathematics, signal filtering, image procossings are right
Related technical personnel have higher requirement.The present invention is proposed using deep learning theory, to pantograph carbon slide surface defect
Image carries out automatic identification and classification.Mutually compared with traditional image detecting method, deep learning only needs to build network mould appropriate
Type, and original image is simply pre-processed, autonomous learning and feature extraction can be carried out to it using network, is realized complete
Automatic image recognition；Model is once completed to train, you can is directly used in image recognition.Training is not necessarily to dependence with using process
Complex operations manually are carried out to image and model.Network makes model to disease advanced features the non-linear fusion of low-level features
Identification can not be illuminated by the light, displacement, the variations such as scale influence, have better robustness and higher automation with
Intelligence degree.
In the first aspect of the present invention, shown in Fig. 1, a kind of pantograph carbon slide surface disease is provided and is known automatically
Other system, including：Acquisition module 100 acquires pantograph pan surface defect image；Processing module, the processing module 200 are wrapped
It includes：Operation is normalized to the image data of acquisition in normalizing unit 201；Converting unit 202 turns the image after normalization
It is changed to data set；Modeling module 300 builds network model, the parameter of network model is arranged；Training module 400, according to by number
The network model, the network model is trained to generate the text of the parameter and value information of preserving training completion according to the data of collection
Part；Module 500 is transferred, the file is called.
Wherein, the acquisition of pantograph image is completed by the industrial camera in tunnel with area array cameras, acquires direction
High speed candid photograph is carried out to vehicle pantograph by the induction to carrying out vehicle for the oblique shooting in 30-60 degree angle above pantograph.In addition,
Using hand-held DSLR cameras and no reverse phase machine, the failure pantograph picture that collection vehicle section is changed supplements training data.
Optionally, described normalize includes：Size normalization, color mode normalization, global characteristics standardization and figure
As going mean value.
Wherein, the size of image is normalized to：
Original rectangle diagram picture is subjected to resize, is converted to the identical square-shaped image of size, i.e., to the length of image
Degree direction carries out certain compression.As shown in Fig. 2, definition This train is bound for XXX be pantograph x axis, perpendicular to pantograph slide
Lath mounting plane is z-axis, is y-axis with x-z-plane vertical-right direction.By image as can be seen that sled surface wears away trace
It is in one-dimensional distribution that mark, which shows as position along y-axis, and depth is distributed along z-axis, and the directions x are without significant difference in same bow item.To image
It is main to influence the information in the directions y when the operation that depth-width ratio is stretched, and the abrasion depth on the directions z is had no significant effect.Mill
The significance degree of consumption can be by Z'(y) reflection, and have：
As can be seen that stretched operation can make the increase in image so that defect is easier identified.Simultaneously as not having
Have and change the ratio that abrasion depth accounts for slide plate thickness, without changing position distribution of the defect along y-axis, therefore resize operations are not yet
Too big negative effect can be caused to the identification classification of disease, instead so that original disease shows to become apparent, be more conducive to
The identification of system.
In addition, color mode is normalized to：
In image capture device used, except Industrial Optics camera and area array cameras acquire as gray level image in tunnel,
Remaining is RGB color image.It is unwise to color data since sled surface defect is there is only the deformation characteristics on geometric scale
Feel, and can be seen that the pantograph image under different color channels through overtesting, the identification of slide plate defect is not present bright
Aobvious difference, therefore image is converted to gray level image using matlab.
Further, for the convolutional neural networks used in this system in calculating process, data are distributed in the model of [0,1]
In enclosing, and data distribution therefore need to carry out global characteristics standard in [0,255] range to gray level image in normal grayscale image
Change, so that the numberical range of its each pixel is zoomed to network operations and be used.Its process can be by as shown in Figure 3.
In addition, before carrying out network training, after subtracting mean value to training data, then other operations are carried out, it on the one hand can be with
Influence of the similar background that do not reject to feature recognition is reduced, the feature of different type disease on the other hand can be protruded, be convenient for
The classification of feature.As shown in figure 4, data go the normalized process of mean value to be one moves to coordinate original by data distribution center
The process of point, to reduce network iterative process very long after weight initialization so that network can be in shorter time
Reach convergence, the classification of complete paired data.
Further, the modeling module includes：Network vision layer dispensing unit, by configuring regarding for convolutional neural networks
Feel convolutional layer, pond layer, the full articulamentum in layer, realizes the extraction and identification of the data characteristics of data set；Network hyper parameter is matched
Unit is set, the hyper parameter of Configuration network model in solver.prototxt files.Network adjusts this three layers by backpropagation
In weights, realize extraction and identification to data characteristics.
The frame of convolutional layer is as shown in table 1：
1 convolution story frame structure of table
Wherein, the configuration of convolutional layer includes：
I, essential information
Part main definitions network layer name, the connection relation of channel type, top layer and bottom, to determine that network layer connects
Relationship flows to define network structure with data transfer.
Ii, parameter regulation rate
This part is defining the learning rates of partial structurtes and its rate of decay.On the basis of basic learning rate, lead to
It crosses and defines this layer of learning rate multiplying power, realize the definition of local learning rate.Due to having two parts parameter in every layer of transmittance process, that is, weigh
Value and biasing, therefore two groups of learning rate definition can be at most set.Two groups of parameters are respectively the multiplying power of weights and biasing, if only one
Group, then two multiplying powers are identical.
Iii, weighting parameter and initialization
The part is to define the convolution kernel size of convolutional layer, the parameters such as sliding step and Filling power, at the same to weights into
Row initialization, therefore this part is the most crucial part of convolutional layer setting, determines the operation and right when data pass through convolutional layer
The extraction strategy of data characteristics, to further affect extraction and data dimension of subsequent image information etc..
When carrying out initial training, the weights of each convolutional layer and biasing are required to through one value of initializing set, with into
Row data transfer, error back propagation will also carry out the adjustment of parameter on the basis of initiation parameter, therefore, at the beginning of parameter
The mode of beginningization has a great impact to the training effect of network with convergence rate.
Configuration for pond layer, pond layer multi-configuration carry out down-sampling after convolutional layer, to the characteristic pattern after convolution,
To reduce operand and data dimension.Pond layer structure is relatively simple, and there is only the definition to Chi Huahe.
Configuration for full articulamentum, in convolutional neural networks, full articulamentum is usually configured as Feature Mapping structure
It is last at all layers, the two dimensional character figure that convolutional layer exports is mapped as one-dimensional characteristic vector, further map to needed for
Output dimension.Identical as convolutional layer, the configuration of full articulamentum equally includes parameter regulation rate and weighting parameter and its initial
Change.
Network hyper parameter is configured, is existed to the configuration of deep learning model hyper parameter under CAFFE frames
Configured in solver.prototxt files, be mainly used for indicating in network training process training network used and
The place path of test network for verifying and assessing；Iterations of the network in forward direction and back-propagation process are set；
Setting evaluation is spaced with display；The storage of Configuration network snapshot.Incorporated by reference to table 2 to joining to super in solver.prototxt files
Several definition.
Definition in 2 solver.prototxt files of table to hyper parameter
The learning rate of model is constantly adjusted by learning rate decaying strategy, and to avoid training, later stage learning rate is excessive to lead
The gradient of cause is exploded.
Present aspect provides a kind of pantograph carbon slide surface disease automatic recognition systems, by network to image data
Autonomous learning and feature extraction, a series of images pretreatment-image enhancement-edge detection-feature for eliminating traditional means carry
Take-complicated flow such as target identification；Network makes identification of the model to disease advanced features to the non-linear fusion of low-level features
Can not be illuminated by the light, displacement, the variations such as scale influence, there is better robustness；Model is completed once training, you can directly
For image recognition, and training manually carries out complex operations without relying on using process to image and model, has higher
Intellectually and automatically degree.
Preferably, the network vision layer dispensing unit further configures in the vision layers of convolutional neural networks
Dropout layers.
Dropout layers refer to the weight for closing at random in the training process part hidden layer node, so that it is not involved in epicycle and change
In generation, calculates, but still retains its weights in a network, reduces the simultaneous adaptation between neuron node, to a certain extent
Enhance the generalization ability of network.Its principle is as shown in Figure 5.
When the amount of training data of network is smaller and the network number of plies is more, it is easy to happen the over-fitting of network so that
Network judging nicety rate on training set is higher, and larger in the upper error of verification collection, even if such network training effect is preferable,
Also it is difficult to be promoted the use of.To reduce the over-fitting of network, dropout layers are arranged in a network, by closing network at random
Connection, reduces the complexity of network, to reduce the over-fitting of network.
In addition to the network layer essential information of standard configuration, network parameter only includes this parameter of dropout ratios,
It selects part hidden layer node at random in equivalent layer according to this ratio, its weight is set to 0 temporarily it is made to be not involved in epicycle and change
Generation.The layer is chiefly used in after full articulamentum, and is only involved in the operation of network training process.
Preferably, the modeling module further comprises：Network structure dispensing unit 303, the parameter of Configuration network structure.
Setting integrally-built to network is carried out referring especially to the basic thought of existing AlexNet and VGGNet, i.e., right
In the setting up procedure of network structure, convolutional layer size is successively decreased from large to small；Large scale list is substituted using small scale series connection convolutional layer
One convolutional layer reduces the number of parameters of network under the premise of ensureing constant to the receptive field of image, reduces operation pressure, together
Shi Zengjia nonlinear activation numbers make network have better effect.
When convolutional layer carries out feature extraction to the 2-D data of input, the change in size of output data meets rule such as
Under：
Wherein k.size is convolution kernel size, and I.size and O.size is respectively to input to be with output figure size, stride
Convolution kernel sliding step, it is Boundary filling to be usually arranged as 1, pad, is usually arranged as 0.
Therefore it is K for convolution kernel sizesM convolution of series connection, equivalent convolution kernel size KbIt can be by indicating as follows：
kb=m (ks-1)+1 (3)
Based on above formula, as shown in fig. 6, the convolution kernel arranged in series of two 3*3 sizes, receptive field range are equivalent to one
The convolution kernel of 5*5.
Since convolutional neural networks are completed to work to the Feature extraction and recognition of image by convolutional layer and pond layer, it is not necessarily to
It is manually extracted, therefore the extraction of the training and characteristics of image to network is directly slided using the pantograph carbon that processing module obtains
Plate surface disease geo-radar image data set executes modeling module, is arranged with network structure to the parameter in model, carries out network training.
After training, the * .caffemodel files ultimately generated are the model file of algorithm, remain all of training completion
Parameter and value information, can be by calling this document implementation model to classify the identification of pantograph carbon slide surface disease geo-radar image.
As shown in Figure 10, this method is finally 90.625% or so to the recognition accuracy of image, error 0.68 or so, compared with current
AlexNet network models, accuracy rate improve 12.5%, and error reduces 20.9% by 0.86.It is instructed as shown in figure 11 to utilize
Damage Types shown in picture are identified in the network model perfected, and export top-5 probability, and picture used is subordinate to class respectively
Other 0, classification 1 and classification 3, i.e., respectively normal, excessive wear, kerve, right side are prediction result of the network to the picture, wherein
Left side of the digital represents prediction classification, and right side is the probability of the category.It can be seen that for selected test image, this method can
Judged with carrying out accurately identification to it.
In the second aspect of the present invention, in conjunction with Fig. 7, a kind of pantograph carbon slide surface disease automatic identifying method is provided,
Including：
S100：Acquisition module acquires pantograph pan surface defect image；
S200：Operation is normalized to the image data of acquisition in the normalizing unit of processing module；
S300：Image after normalization is converted to data set by the converting unit of processing module；
S400：Modeling module builds network model, and the parameter of network model is arranged；
Specifically, please referring to Fig. 8, the modeling module builds network model, and the parameter that network model is arranged includes：
S411：Convolutional layer, pond layer in vision layer of the network vision layer dispensing unit by configuring convolutional neural networks,
Full articulamentum realizes the extraction and identification of the data characteristics of data set；
S412：The hyper parameter of network hyper parameter dispensing unit Configuration network model in solver.prototxt files.
S500：Training module trains the network model according to the data by data set, and the network model, which generates, to be protected
There is the file of the parameter and value information of training completion；
S600：It transfers module and calls the file.
Present aspect provides a kind of pantograph carbon slide surface disease automatic identifying method, by network to picture number
According to autonomous learning and feature extraction, eliminate a series of images pretreatment-image enhancement-edge detection-spy of traditional means
Levy complicated flows such as extraction-target identification；Network makes model to disease advanced features the non-linear fusion of low-level features
Identification can not be illuminated by the light, displacement, the variations such as scale influence, there is better robustness；Model is completed once training, you can
It is directly used in image recognition, and training manually carries out complex operations without relying on using process to image and model, has
Higher intellectually and automatically degree.
Further, in one preferred embodiment of present aspect, the modeling module builds network model, please refers to
Fig. 9 and parameter that network model is arranged includes：
S421：Convolutional layer, pond layer in vision layer of the network vision layer dispensing unit by configuring convolutional neural networks,
Full articulamentum realizes the extraction and identification of the data characteristics of data set；
S422：The hyper parameter of network hyper parameter dispensing unit Configuration network model in solver.prototxt files.
S423：The network vision layer dispensing unit further configures the Dropout in the vision layer of convolutional neural networks
Layer.
Further, the modeling module builds network model, and the parameter that network model is arranged further comprises：
S424：Network structure dispensing unit, the parameter of Configuration network structure.
The method can be simultaneously including S423 and S424 steps, can also be only including wherein any one step, this hair
It is bright to repeat no more.
When the amount of training data of network is smaller and the network number of plies is more, it is easy to happen the over-fitting of network so that
Network judging nicety rate on training set is higher, and larger in the upper error of verification collection, even if such network training effect is preferable,
Also it is difficult to be promoted the use of.To reduce the over-fitting of network, dropout layers are arranged in a network, by closing network at random
Connection, reduces the complexity of network, to reduce the over-fitting of network.
In addition to the network layer essential information of standard configuration, network parameter only includes this parameter of dropout ratios,
It selects part hidden layer node at random in equivalent layer according to this ratio, its weight is set to 0 temporarily it is made to be not involved in epicycle and change
Generation.The layer is chiefly used in after full articulamentum, and is only involved in the operation of network training process.
Setting integrally-built to network is carried out referring especially to the basic thought of existing AlexNet and VGGNet, i.e., right
In the setting up procedure of network structure, convolutional layer size is successively decreased from large to small；Large scale list is substituted using small scale series connection convolutional layer
One convolutional layer reduces the number of parameters of network under the premise of ensureing constant to the receptive field of image, reduces operation pressure, together
Shi Zengjia nonlinear activation numbers make network have better effect.
Belong to " first ", " second " etc. in description and claims of this specification and above-mentioned attached drawing are for distinguishing
Different objects, rather than for describing particular order.In addition, term " comprising " and " having " and their any deformations, meaning
Figure, which is to cover, non-exclusive includes.Such as contain process, method, system, product or the equipment of series of steps or unit
The step of being not limited to list or unit, but further include the steps that optionally not listing or unit, or optionally also
Include for these processes, method or the intrinsic gas step or unit of equipment.
Obviously, the above embodiment of the present invention be only to clearly illustrate example of the present invention, and not be pair
The restriction of embodiments of the present invention may be used also on the basis of the above description for those of ordinary skill in the art
To make other variations or changes in different ways, all embodiments can not be exhaustive here, it is every to belong to this hair
Row of the obvious changes or variations that bright technical solution is extended out still in protection scope of the present invention.
Claims (9)
1. a kind of pantograph carbon slide surface disease automatic recognition system, which is characterized in that including：
Acquisition module acquires pantograph pan surface defect image；
Processing module, the processing module include：
Operation is normalized to the image data of acquisition in normalizing unit；
Image after normalization is converted to data set by converting unit；
Modeling module builds network model, the parameter of network model is arranged；
Training module trains the network model, the network model generation to preserve training according to the data by data set
The parameter of completion and the file of value information；
Module is transferred, the file is called.
2. system according to claim 1, which is characterized in that the normalization includes：
Size normalization, color mode normalization, global characteristics standardization and image go mean value.
3. system according to claim 1, which is characterized in that the modeling module includes：
Network vision layer dispensing unit, the convolutional layer in vision layer, pond layer, full connection by configuring convolutional neural networks
Layer, realizes the extraction and identification of the data characteristics of data set；
Network hyper parameter dispensing unit, the hyper parameter of Configuration network model in solver.prototxt files.
4. system according to claim 3, which is characterized in that the network vision layer dispensing unit further configures convolution god
Dropout layers in vision layer through network.
5. system according to claim 3, which is characterized in that the modeling module further comprises：
Network structure dispensing unit, the parameter of Configuration network structure.
6. a kind of pantograph carbon slide surface disease automatic identifying method, which is characterized in that including：
Acquisition module acquires pantograph pan surface defect image；
Operation is normalized to the image data of acquisition in the normalizing unit of processing module；
Image after normalization is converted to data set by the converting unit of processing module；
Modeling module builds network model, and the parameter of network model is arranged；
Training module trains the network model according to the data by data set, and the network model generation, which is preserved, have been trained
At parameter and value information file；
It transfers module and calls the file.
7. method according to claim 1, which is characterized in that the modeling module builds network model, and network mould is arranged
The parameter of type includes：
Convolutional layer, pond layer, full articulamentum in vision layer of the network vision layer dispensing unit by configuring convolutional neural networks,
Realize the extraction and identification of the data characteristics of data set；
The hyper parameter of network hyper parameter dispensing unit Configuration network model in solver.prototxt files.
8. method according to claim 7, which is characterized in that the network vision layer dispensing unit further configures convolution god
Dropout layers in vision layer through network.
9. method according to claim 7, which is characterized in that the modeling module builds network model, and network mould is arranged
The parameter of type further comprises：
Network structure dispensing unit, the parameter of Configuration network structure.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810200743.1A CN108573277A (en) | 2018-03-12 | 2018-03-12 | A kind of pantograph carbon slide surface disease automatic recognition system and method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810200743.1A CN108573277A (en) | 2018-03-12 | 2018-03-12 | A kind of pantograph carbon slide surface disease automatic recognition system and method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108573277A true CN108573277A (en) | 2018-09-25 |
Family
ID=63576784
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810200743.1A Pending CN108573277A (en) | 2018-03-12 | 2018-03-12 | A kind of pantograph carbon slide surface disease automatic recognition system and method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108573277A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109658387A (en) * | 2018-11-27 | 2019-04-19 | 北京交通大学 | The detection method of the pantograph carbon slide defect of power train |
| CN109783929A (en) * | 2019-01-10 | 2019-05-21 | 北京交通大学 | Railcar pantograph carbon slide wears away evaluation method and life-span prediction method |
| CN110118788A (en) * | 2019-04-30 | 2019-08-13 | 东南大学 | Submerged structure Defect inspection contact visual sensing equipment and image processing method |
| CN111006592A (en) * | 2019-12-02 | 2020-04-14 | 西安宝利铁路器材有限公司 | Locomotive hook height monitoring device and method |
| CN111238393A (en) * | 2020-01-20 | 2020-06-05 | 成都铁安科技有限责任公司 | Pantograph carbon slide plate detecting system and its control method |
| CN111259762A (en) * | 2020-01-13 | 2020-06-09 | 西南交通大学 | Pantograph abnormity detection method |
| CN111563439A (en) * | 2020-04-28 | 2020-08-21 | 北京海益同展信息科技有限公司 | Aquatic organism disease detection method, device and equipment |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104850845A (en) * | 2015-05-30 | 2015-08-19 | 大连理工大学 | Traffic sign recognition method based on asymmetric convolution neural network |
| WO2015180397A1 (en) * | 2014-05-31 | 2015-12-03 | 华为技术有限公司 | Method and device for recognizing data category based on deep neural network |
| CN107437245A (en) * | 2017-06-26 | 2017-12-05 | 西南交通大学 | High-speed railway touching net method for diagnosing faults based on depth convolutional neural networks |
| CN107742130A (en) * | 2017-10-25 | 2018-02-27 | 西南交通大学 | High iron catenary based on deep learning supports device fastener failure diagnostic method |
-
2018
- 2018-03-12 CN CN201810200743.1A patent/CN108573277A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2015180397A1 (en) * | 2014-05-31 | 2015-12-03 | 华为技术有限公司 | Method and device for recognizing data category based on deep neural network |
| CN104850845A (en) * | 2015-05-30 | 2015-08-19 | 大连理工大学 | Traffic sign recognition method based on asymmetric convolution neural network |
| CN107437245A (en) * | 2017-06-26 | 2017-12-05 | 西南交通大学 | High-speed railway touching net method for diagnosing faults based on depth convolutional neural networks |
| CN107742130A (en) * | 2017-10-25 | 2018-02-27 | 西南交通大学 | High iron catenary based on deep learning supports device fastener failure diagnostic method |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109658387A (en) * | 2018-11-27 | 2019-04-19 | 北京交通大学 | The detection method of the pantograph carbon slide defect of power train |
| CN109658387B (en) * | 2018-11-27 | 2023-10-13 | 北京交通大学 | Method for detecting defects of pantograph carbon slide plate of electric train |
| CN109783929A (en) * | 2019-01-10 | 2019-05-21 | 北京交通大学 | Railcar pantograph carbon slide wears away evaluation method and life-span prediction method |
| CN109783929B (en) * | 2019-01-10 | 2021-02-19 | 北京交通大学 | Subway vehicle pantograph carbon slide plate abrasion estimation method and service life prediction method |
| CN110118788A (en) * | 2019-04-30 | 2019-08-13 | 东南大学 | Submerged structure Defect inspection contact visual sensing equipment and image processing method |
| CN111006592A (en) * | 2019-12-02 | 2020-04-14 | 西安宝利铁路器材有限公司 | Locomotive hook height monitoring device and method |
| CN111259762A (en) * | 2020-01-13 | 2020-06-09 | 西南交通大学 | Pantograph abnormity detection method |
| CN111259762B (en) * | 2020-01-13 | 2022-03-29 | 西南交通大学 | Pantograph abnormity detection method |
| CN111238393A (en) * | 2020-01-20 | 2020-06-05 | 成都铁安科技有限责任公司 | Pantograph carbon slide plate detecting system and its control method |
| CN111563439A (en) * | 2020-04-28 | 2020-08-21 | 北京海益同展信息科技有限公司 | Aquatic organism disease detection method, device and equipment |
| CN111563439B (en) * | 2020-04-28 | 2023-08-08 | 京东科技信息技术有限公司 | Aquatic organism disease detection method, device and equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108573277A (en) | A kind of pantograph carbon slide surface disease automatic recognition system and method | |
| CN111598861B (en) | Improved Faster R-CNN model-based non-uniform texture small defect detection method | |
| CN107576960B (en) | Target detection method and system for visual radar space-time information fusion | |
| CN110188705B (en) | Remote traffic sign detection and identification method suitable for vehicle-mounted system | |
| CN108154508B (en) | Method, apparatus, storage medium and the terminal device of product defects detection positioning | |
| CN106504248B (en) | Vehicle damage judging method based on computer vision | |
| CN110298387A (en) | Incorporate the deep neural network object detection method of Pixel-level attention mechanism | |
| CN104112269B (en) | A kind of solar battery laser groove parameter detection method and system based on machine vision | |
| CN110163187A (en) | Remote road traffic sign detection recognition methods based on F-RCNN | |
| CN108830188A (en) | Vehicle checking method based on deep learning | |
| CN106841216A (en) | Tunnel defect automatic identification equipment based on panoramic picture CNN | |
| CN109389599A (en) | A kind of defect inspection method and device based on deep learning | |
| CN109726627A (en) | A kind of detection method of neural network model training and common ground line | |
| CN112329747B (en) | Vehicle parameter detection method based on video identification and deep learning and related device | |
| CN105354568A (en) | Convolutional neural network based vehicle logo identification method | |
| CN110210350A (en) | A kind of quick parking space detection method based on deep learning | |
| CN110120036A (en) | A kind of multiple dimensioned tire X-ray defect detection method | |
| CN104166841A (en) | Rapid detection identification method for specified pedestrian or vehicle in video monitoring network | |
| Dib et al. | A review on negative road anomaly detection methods | |
| CN105260734A (en) | Commercial oil surface laser code recognition method with self modeling function | |
| CN102332089A (en) | Railway wagon brake shoe key going-out fault recognition method based on artificial neural network | |
| CN110163208A (en) | A kind of scene character detecting method and system based on deep learning | |
| CN117037119A (en) | Road target detection method and system based on improved YOLOv8 | |
| CN114648669A (en) | Motor train unit fault detection method and system based on domain-adaptive binocular parallax calculation | |
| CN115937518A (en) | Pavement disease identification method and system based on multi-source image fusion |
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 |
Application publication date: 20180925 |
|
| RJ01 | Rejection of invention patent application after publication |