CN110246132A - Rail vehicle bolt looseness detection method and system - Google Patents
Rail vehicle bolt looseness detection method and system Download PDFInfo
- Publication number
- CN110246132A CN110246132A CN201910545803.8A CN201910545803A CN110246132A CN 110246132 A CN110246132 A CN 110246132A CN 201910545803 A CN201910545803 A CN 201910545803A CN 110246132 A CN110246132 A CN 110246132A
- Authority
- CN
- China
- Prior art keywords
- network
- critical component
- layer
- image
- bolt
- 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
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M17/00—Testing of vehicles
- G01M17/08—Railway vehicles
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30108—Industrial image inspection
- G06T2207/30164—Workpiece; Machine component
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of rail vehicle bolt looseness detection method and system, the detection method includes: S1, multiple rail vehicle critical component images under different illumination conditions at acquisition bolt；S2, by every image cutting-out to being sized；Every image is manually marked；S3, foundation and training critical component position network；Layer is cut out in S4, setting；S5, foundation and training bolt looseness detect network；S6, image measurement.The system comprises multispectral timesharing light source, image acquiring device, image processing module, critical component positioning network establishment and training module, cut out a layer setup module, bolt looseness detection network establishment and training module and storage and test module.The present invention can carry out bolt status monitoring to critical component, provide the more specific location information of loose bolts, and detection speed is fast, can be realized real-time online detection, not only improve detection efficiency, additionally it is possible to improve the accuracy rate of detection.
Description
Technical field
The invention belongs to railway fault detection technique fields, are related to rail vehicle bolt looseness detection technique, specifically,
It is related to a kind of rail vehicle bolt looseness detection method and system.
Background technique
Bolt is the conventional fasteners of the rail vehicles such as EMU, urban rail and subway, when driving due to by burn into
The influence of the factors such as vibration and impact, it is easy to cause the deformation of bolt, loosening, be broken or fall off, so as to cause equipment fault,
Even cause serious accident.Therefore, the big event of always vehicle daily test is checked the loosening of bolt.
For guarantee rail vehicle safe operation, staff need daily to bolt carry out status checkout, inspection item its
In one be exactly to check whether bolt loosens.Under normal conditions, just there are several hundred even more bolts above a section compartment, one day
There is multiple row train again in the middle, the detection to bolt looseness is an extremely tedious job.The method that tradition checks bolt is behaved
Work inspection method, there are mainly two types of mode: a kind of mode is regardless of whether bolt has loosened, and staff can screw up with a wrench
Bolt, it is ensured that bolt is in the state tightened；Another way be bolt after tightening for the first time respectively on screw bolt and nut
The state tightened with wire tag, it is whether scribing line on screw bolt and nut is aligned that when inspection, which need to only check, is twisted if being misaligned
Tightly to its alignment.First way heavy workload, does not have specific aim at inefficiency, and the second way is easy to make staff
Visual fatigue is generated after detecting a large amount of bolts, leads to false retrieval, missing inspection, the bullet train especially opened now will more work
Personnel can be completed rapidly and accurately the detection of bolt.Further, since most trains are put in storage at night, night personnel are easy tired
Labor, repair quality depend on the sense of responsibility of staff, and details dimly is difficult to find again under the influence of the factors such as illumination, fatigue,
It is easy to cause potential security risk.In order to reduce workload, working efficiency is improved.Currently, more commonly used bolt detection side
Method is based primarily upon analysis of vibration signal and image recognition.
Bolt detection method based on analysis of vibration signal, when obtaining vibration signal, mode that there are mainly two types of: one
Kind is using ODS (full name: Operational Deflection Shape) mathematical model, and analysis wheel passes through the vibration in certain section
Dynamic signal, artificial observation filter out the abnormal conditions of ODS value, and then judge to loosen section, but which is not able to achieve online reality
When detect；Another way is to hammer into shape constantly to tap by percussion to be fixed structure, while remained shock signal is acquired and being analyzed,
But there are slow problems for which.In addition, although above two mode improves work efficiency to a certain extent,
Working efficiency is not high, however it remains ineffective problem.
Bolt looseness detection method based on image recognition depends on locking mark and template image.Such as: Publication No.
The Chinese patent application of CN 108469336A discloses a kind of bolt looseness detection method based on image procossing, including following
Step: on the bolt tightened, installation direction marker；Using structural edge line relatively-stationary near detected bolt as ginseng
Examine line；Camera is directed at bolt, bearing mark object and reference line；Absorb the first time image of bolt tight condition；Pass through meter
Calculation machine handles the first time image of intake, identifies bolt-center point, bearing mark structure, reference line, then identify
Angle point is marked, vertical line is then made, makees emergent ray from bolt-center point, calculated in the counterclockwise direction since vertical line is with zero degree
Angle of the ray relative to vertical line；It is iteratively repeated the angle that above-mentioned steps obtain n-th；By the angle obtained every time and for the first time
Angle compares, and when deviation is larger, will test out the bolt loosened and tightens.Since driving conditions environment is complicated, locking mark may
Can be blocked or fallen off by greasy dirt covering, foreign matter, and the shape color of greasy dirt, foreign matter is different, be inverted template matching score very it is low very
Can not extremely it judge.Thus, it relies on existing single-spectral images acquisition device and normal image recognition methods is difficult to meet practical need
It asks, there is a problem of that Detection accuracy is low.
Summary of the invention
The above problems such as the present invention is difficult for on-line checking existing for existing bolt detection method, Detection accuracy is low,
Provide a kind of rail vehicle bolt looseness detection method and system, can real-time online detection bolt whether loosen, detection effect
Rate is high, using the self study detection algorithm of depth network structure, can effectively improve the accuracy rate of detection.
In order to achieve the above object, the present invention provides a kind of rail vehicle bolt looseness detection method, the steps include:
S1, image is obtained
Multiple rail vehicle critical component images at bolt are obtained under different illumination conditions；
S2, image procossing
Clipping image
By every image cutting-out to being sized；
Mark image
Every image is manually marked, mark image is obtained and marks file；
S3, foundation and training critical component position network
Using convolutional neural networks ZFNet as input terminal, full articulamentum is output end, according to being sequentially connected with convolutional Neural
Network, Pooling layers of ROI and full articulamentum are suggested in network ZFNet, region, establish critical component positioning network, mark is schemed
As be divided into training set and verifying collection, by training set be input to critical component positioning network learn, and by verifying collect into
Row verifying, the critical component after being trained position network；
Layer is cut out in S4, setting
The rail vehicle critical component image information that network output is positioned according to critical component, will be comprising closing by cutting out layer
The rectangular area of key member is cut out from original track vehicle critical component image to be come, and critical component rectangular region image is obtained；
S5, foundation and training bolt looseness detect network
Using residual error network as input terminal, full articulamentum is output end, and network and ROI are suggested in region interconnected
Pooling layers are plugged in residual error network, and region suggests that network is connect with residual error network the second last layer convolutional layer, ROI
Pooling layers connect with the last first layer convolutional layer of residual error network, and residual error network the last layer convolutional layer is connect with full articulamentum,
It establishes training bolt looseness and detects network, critical component rectangular region image is divided into training set and verifying collects, by training set
It is input to bolt looseness detection network to be learnt, and collects the bolt looseness detection net carried out after verification is trained by verifying
Network；
S6, image measurement
The rail vehicle critical component image at bolt is obtained in real time as test image, and test image is input to training
Critical component afterwards positions network, and the output information of network is positioned according to critical component, is cut out by cutting out layer comprising key portion
The bolt looseness detection network that the rectangular area of part obtains critical component rectangular region image and is input to after training, according to
The output result of bolt looseness detection network judges whether bolt loosens.
Further, in step S2, before clipping image, by the method for image enhancement to the image of acquisition at
Reason.
Further, by cutting out layer that the rectangular area comprising critical component is crucial from original track vehicle in step S4
It is cut out in image of component before coming, whether which is completely confirmed, the specific steps are that:
Equivalent length of side ratio D is defined, equivalent length of side ratio D is indicated are as follows:
In formula, H is the height of input picture, and W is the width of input picture, (xmin,ymin) it is that critical component positions network output
The left upper apex coordinate of rectangular area, (xmax,ymax) it is the bottom right vertex coordinate that critical component positions that network exports rectangular area；
Definition decides whether that the threshold value cut out is d, and as D > d, the critical component is imperfect, and critical component is imperfect
Picture reject, as D≤d, the critical component is complete, obtains the complete picture of critical component and obtains for being cut out by cutting out layer
Obtain critical component rectangular region image.
Preferably, in step S2, every image is manually marked using labelImg annotation tool, marks out every
The minimum circumscribed rectangle frame coordinate of each critical component of image simultaneously generates corresponding xml mark file.
Preferably, in step S3, will mark image according to the quantitative proportion of 4:1 or 3:1 or 5:1 be divided into training set and
Verifying collection calls one-time authentication collection to be verified when being verified according to setting interval time, positions network in key position
Deconditioning, the key position after being trained positioned net before occurring fitting and when training set loss curve has been restrained
Network.
Preferably, in step S3, mark image obtains characteristic pattern after convolutional neural networks, and characteristic pattern is admitted to region and builds
Discuss network, by region suggest network in full convolution network and non-linear layer relu operation after, on characteristic pattern each 3 ×
3 region can be changed to the feature vector of one 512 dimension, and this feature vector is sent into region and suggests that the full articulamentum in network carries out
The classification of candidate frame and the recurrence of candidate frame position, according to the location information of candidate frame, after convolutional neural networks processing
To characteristic pattern on by ROI Pooling layer Pooling layers of Pooling, ROI of progress ROI export result enter key
Full articulamentum in positioning parts network carries out the classification of best candidate frame and the recurrence of best candidate frame position, critical component
The rail vehicle critical component image for positioning network final output includes a left side for one or more completely rectangle frames of critical components
Upper apex coordinate and bottom right vertex coordinate, critical component title and confidence level.
Preferably, in step S3, convolutional neural networks are made of the series connection of one or more groups of convolution units.
Preferably, every group of convolution unit is connected in series to form by a convolutional layer conv or several convolutional layers conv
Convolution layer unit and the non-linear layer unit being connected in series to form by a non-linear layer relu or several non-linear layers relu
Composition, convolution layer unit the last layer convolutional layer conv are connect with the first layer non-linear layer relu of non-linear layer unit；First
Group convolution unit non-linear layer unit connect with second group of convolution layer unit, the convolution layer unit of third group convolution unit and
Second group of non-linear layer unit connection, and so on, convolution layer unit and the second last group of last group of convolution unit
The connection of non-linear layer unit.
Preferably, every group of convolution unit is connected in series to form by a convolutional layer conv or several convolutional layers conv
Convolution layer unit, by non-linear layer unit that a non-linear layer relu or several non-linear layers relu are connected in series to form with
And the pond layer being connected in series to form by a maximum pond layer maxpooling or several maximums pond layer maxpooling
Unit composition, convolution layer unit the last layer convolutional layer conv are connect with the first layer non-linear layer relu of non-linear layer unit,
The last layer non-linear layer relu of non-linear layer unit and the first layer maximum pond layer maxpooling of pond layer unit connect
It connects；The pond layer unit of first group of convolution unit is connect with second group of convolution layer unit, the convolutional layer of third group convolution unit
Unit is connect with second group of pond layer unit, and so on, the convolution layer unit and the second last of last group of convolution unit
The pond layer unit connection of group.
Preferably, in step S4, when establishing bolt looseness detection network, using labelImg annotation tool to every key
Component rectangular region image is manually marked, and the minimum of each bolt in every critical component rectangular region image is marked out
Boundary rectangle frame coordinate simultaneously generates corresponding xml mark file.
Preferably, in step S4, by the critical component rectangular region image marked according to the number of 4:1 or 3:1 or 5:1
Measuring ratio cut partition is training set and verifying collection, when being verified, calls one-time authentication collection to be tested according to setting interval time
Card, deconditioning before fitting occurred in bolt looseness detection network and when training set loss curve has been restrained, obtains
Bolt looseness after training detects network.
Preferably, in step S4, the output of the second last layer convolutional layer of residual error network is that network and ROI are suggested in region
The size dimension of Pooling layers of shared part, Pooling layers of ROI output characteristic pattern meets the last layer in former residual error network
The input of convolutional layer, residual error network the last layer convolutional layer meet average pond layer Average Pooling, obtain 2048 Wei Te
Sign carries out the classification of candidate frame and the recurrence of candidate frame position, bolt into the full articulamentum in bolt looseness detection network
Loosen the rectangle frame position, confidence level and the label for indicating bolt state that detection network final output includes loose bolts；Institute
Stating mark bolt state includes normal and loosening two states.
Preferably, the residual error network uses 18 layers, 34 layers, 50 layers, 101 layers, 152 layers of residual error network.
In order to achieve the above object, the present invention provides a kind of rail vehicle bolt looseness detection systems, comprising:
Multispectral timesharing light source, for generating the light of different-waveband；
Image acquiring device, for obtaining the rail vehicle critical component image under different illumination conditions and obtaining real-time rail
Road vehicle critical component image；
Image processing module, for by rail vehicle critical component image cutting-out to being sized, and to the rail after cutting out
Road vehicle critical component image is labeled, and obtains corresponding mark image；
Critical component positions network establishment and training module, for building critical component positioning network, and will mark image
It is divided into training set and verifying collection, training set is input to the critical component positioning network built and is learnt, and passes through verifying collection
It is verified, the critical component after being trained positions network；
A layer setup module is cut out, layer is cut out for being arranged, the output information of network is positioned according to critical component, by cutting out
Rectangular area comprising critical component is cut out by layer from original track vehicle critical component image to be come, and critical component rectangle is obtained
Area image；
Bolt looseness detects network establishment and training module, for building bolt looseness detection network, and by critical component
Rectangular region image is divided into training set and verifying collection, training set is input to the critical component rectangular region image network built into
Row study, and verified by verifying collection, the bolt looseness after being trained detects network；
Storage and test module, for store training after critical component positioning network and bolt looseness detection network and
It cuts out layer, and network is positioned by critical component, cuts out layer and bolt looseness detects network and carries out spiral shell to the image obtained in real time
Bolt loosens test, and outputs test result.
Preferably, the multispectral timesharing light source is made of multiple groups LED point light source.
Preferably, LED point light source uses three kinds of colors of red, green, blue, by the LED point light source of three kinds of colors of red, green, blue,
Red, green, blue three coloured light is provided, or recalls other color of light in addition to red, green, blue three coloured light.
Preferably, the LED point light source is using red, green, blue and white four kinds of colors, wherein by red, green, blue with
And the LED point light source of white four kinds of colors, panchromatic light or red, green, blue three coloured light are provided, or recall except red, green, blue three coloured light
Other outer color of light.
Compared with prior art, the advantages and positive effects of the present invention are:
(1) detection method obtains image under different illumination conditions, can extract image in different illumination conditions
Under characteristic information, suitable for the project situation that rail vehicle detection background is complicated, detection target is numerous.
(2) working experience and sense of responsibility of the detection method independent of service personnel, be greatly saved the time at
Sheet and human input detect cascade formula network frame using critical component positioning network+bolt looseness, initiative by spiral shell
Bolt loosens test problems and is converted into target identification problem twice, and carries out self study training using to above-mentioned two network, can
Study is not only able to efficiently locate the crucial portion in rail vehicle image to effective feature representation from data well
Part, additionally it is possible to bolt status monitoring be carried out to critical component, provide the more specific location information of loose bolts, detection speed is fast, energy
It enough realizes real-time online detection, not only improves detection efficiency, also improve the accuracy rate that can be improved detection.
(3) layer is cut out in detection method setting, critical component is cut out from original image to be sent into bolt looseness net
Network, compared to non-cascaded formula structure (i.e. directly detecting bolt on whole image), the crucial portion that the present invention is cut out by cutting out layer
In part rectangular region image, elemental area shared by bolt is big, and bolt region information is comprehensive, and when positioning bolt is accurate, therefore detects
Precision is high.Opposite, it is directly detected for bolt on whole image, the elemental area as shared by bolt is too small, in convolution knot
Pond layer in structure can further reduce bolt, and bolt region has lost many information, cause to position spiral shell on whole figure
Just not enough precisely, detection accuracy is relatively low for bolt.
(4) detection method is independent of image template, for rail vehicle cross vehicle direction it is different, because of speed unevenness
The inclined engineering background of certain angle occurs for even critical component, and the method for the present invention remains to be applicable in.
(5) method of detection method critical component positioning Web vector graphic deep learning training can be crucial in positioning
Possess enough robustness when position to guarantee that bolt looseness detects the recall rate that network detects bolt looseness.
(6) detection system of the present invention is illuminated using the light that multispectral timesharing light source provides different-waveband, is obtained different
Image under the conditions of band of light photograph, can obtain the characteristic information under different illumination conditions, be suitable for rail vehicle object to be measured
The project situation of the complexity of background locating for numerous and target.
(7) detection system of the present invention detects cascade formula network frame using critical component positioning network+bolt looseness,
Initiative converts bolt looseness test problems to target identification problem twice, and detection speed is fast, can be realized real-time online
Detection, detection efficiency is high, and using the automatic learning training of depth network structure, can be good at learning from data to effective
Feature representation, improve the accuracy rate of detection.
Detailed description of the invention
Fig. 1 is the flow chart of rail vehicle of embodiment of the present invention bolt looseness detection method；
Fig. 2 a is the critical component image that the critical component that the embodiment of the present invention obtains is axle box cover；
Fig. 2 b is the critical component image that the critical component that the embodiment of the present invention obtains is height adjusting valve；
Fig. 2 c is the critical component image that the critical component that the embodiment of the present invention obtains is bogie frame；
Fig. 3 is the structure chart that critical component of the embodiment of the present invention positions network；
Fig. 4 is the structure chart that bolt looseness of the embodiment of the present invention detects network；
Fig. 5 is the structure chart that critical component of the embodiment of the present invention positions network and bolt looseness detects cascade；
Fig. 6 is critical component of embodiment of the present invention positioning result schematic diagram；
Fig. 7 is bolt looseness of embodiment of the present invention testing result schematic diagram；
Fig. 8 a-d is that bolt is normally bolt looseness schematic diagram；
Fig. 9 a-e is that treated schemes using rotation, fuzzy, contrast normalization, the wrong image enchancing method cut, translate
As schematic diagram；
Figure 10 is one embodiment of the invention convolutional neural networks structure chart；
Figure 11 is the structure chart of neural network described in the embodiment of the present invention；
Figure 12 a-12b is buildingblock figure；
Figure 13 is the structural schematic diagram of the multispectral timesharing light source of the embodiment of the present invention.
Specific embodiment
In the following, the present invention is specifically described by illustrative embodiment.It should be appreciated, however, that not into one
In the case where step narration, element, structure and features in an embodiment can also be advantageously incorporated into other embodiments
In.
Referring to Fig. 1, one embodiment of the invention provides a kind of rail vehicle bolt looseness detection method, the steps include:
S1, image is obtained
Multiple rail vehicle critical component images at bolt are obtained under different illumination conditions.A-2c referring to fig. 2 is obtained
Image include axle box cover (referring to fig. 2 a), height adjusting valve (referring to fig. 2 b), (c) etc. many places contain bogie frame referring to fig. 2
There is the rail vehicle critical component image of bolt.
S2, image procossing
S21, clipping image
By every image cutting-out to being sized.
S22, mark image
Every image is manually marked, mark image is obtained and marks file.Specifically, it is marked using labelImg
Tool manually marks every image, marks out the minimum circumscribed rectangle frame coordinate of each critical component of every image simultaneously
Generate corresponding xml mark file, that is, VOC2007 data set format.
S3, foundation and training critical component position network
Using convolutional neural networks, (referred to as: for ZFNet as input terminal, full articulamentum is output end, is rolled up according to being sequentially connected with
Product neural network, region suggest network (English: Region Proposal Network, abbreviation: RPN), Pooling layers of ROI
With full articulamentum, critical component positioning network is established, mark image is divided into training set and verifying collects, training set is input to
Critical component positioning network is learnt, and is verified by verifying collection, and the critical component after being trained positions network.
Specifically, when training critical component positioning network, mark image is divided into training set according to the quantitative proportion of 4:1
Collect with verifying, when being verified, call one-time authentication collection to be verified according to setting interval time, positions net in key position
Deconditioning, the key position after being trained positioned before fitting occurred in network and when training set loss curve has been restrained
Network.
Specifically, referring to Fig. 3, Fig. 5, critical component positions the process that network positions critical component in mark image
Are as follows: mark image obtains characteristic pattern after convolutional neural networks, and characteristic pattern is admitted to region and suggests network, suggests net by region
After full convolution network and non-linear layer relu operation in network, each 3 × 3 region can be changed to one on characteristic pattern
The feature vector of 512 dimensions, this feature vector are sent into region and suggest that the full articulamentum in network carries out the classification and time of candidate frame
Select the recurrence of frame position.According to the location information of candidate frame, pass through ROI on the characteristic pattern obtained after convolutional neural networks processing
The result of Pooling layers of Pooling layers of Pooling, ROI of progress ROI output enters connecting in critical component positioning network entirely
It connects layer and carries out the classification of best candidate frame and the recurrence of best candidate frame position.The rail of critical component positioning network final output
Road vehicle critical component image includes the left upper apex coordinate and bottom right vertex of the rectangle frame of one or more complete critical components
Coordinate, critical component title and confidence level (referring to Fig. 6).
Layer is cut out in S4, setting
The rail vehicle critical component image information that network output is positioned according to critical component, will be comprising closing by cutting out layer
The rectangular area of key member is cut out from original track vehicle critical component image to be come, and critical component rectangular region image is obtained.
S5, foundation and training bolt looseness detect network
With residual error network (referred to as: RestNet) for input terminal, full articulamentum is output end, and the region of sequential connection is built
It discusses network and Pooling layers of ROI is plugged in residual error network, network and residual error network the second last layer convolutional layer are suggested in region
Connection, Pooling layers of ROI connect with the last first layer convolutional layer of residual error network, residual error network the last layer convolutional layer and Quan Lian
Layer connection is connect, training bolt looseness is established and detects network.Critical component rectangular region image is divided into training set and verifying collection,
Training set is input to bolt looseness detection network to learn, and collects the bolt pine carried out after verification is trained by verifying
Dynamic detection network.
Specifically, when establishing bolt looseness detection network, using labelImg annotation tool to every critical component rectangle
Area image is manually marked, and the minimum circumscribed rectangle of each bolt in every critical component rectangular region image is marked out
Frame coordinate simultaneously generates corresponding xml mark file, that is, VOC2007 data set format.
Specifically, identical as critical component positioning network training process, when training bolt looseness detects network, it will mark
Critical component rectangular region image according to the quantitative proportion of 4:1 be divided into training set and verifying collect, when being verified, according to
Setting interval time call one-time authentication collection verified, bolt looseness detection network occurred fitting before and training set
Deconditioning when loss curve has been restrained, the bolt looseness after being trained detect network.
Specifically, referring to fig. 4, Fig. 5, bolt looseness detect network to spiral shell in the critical component rectangular region image marked
Bolt loosens the process detected are as follows: the output of the second last layer convolutional layer of residual error network is that network and ROI are suggested in region
The size dimension of Pooling layers of shared part, Pooling layers of ROI output characteristic pattern meets the last layer in former residual error network
The input of convolutional layer, residual error network the last layer convolutional layer meet average pond layer Average Pooling, obtain 2048 Wei Te
Sign carries out the classification of candidate frame and the recurrence of candidate frame position, bolt into the full articulamentum in bolt looseness detection network
The image for loosening detection network final output includes the rectangle frame position of loose bolts, confidence level and the mark for indicating bolt state
It signs (referring to Fig. 7)；The mark bolt state includes normal and loosening two states.
S6, image measurement
The rail vehicle critical component image at bolt is obtained in real time as test image, and test image is input to training
Critical component afterwards positions network, and the output information of network is positioned according to critical component, is cut out by cutting out layer comprising key portion
The bolt looseness detection network that the rectangular area of part obtains critical component rectangular region image and is input to after training, according to
The output result of bolt looseness detection network judges whether bolt loosens.
Currently, bolt looseness detection depends on locking mark.When shooting visual angle is fixed, can on bolt mark level
Locking mark is defined as bolt looseness when 15 degree or more dislocation occur for locking mark relative horizontal position.With bogie frame
For, when locking mark marks on bolt and bogie frame respectively, when two sections of locking marks are not arranged on the same straight line,
Think that bolt looseness occurs.Normal bolt is as shown in Fig. 8 a, 8c, and loose bolts are as shown in Fig. 8 b, 8d.Above-mentioned detection method into
When row image measurement, rectangle frame position, confidence level and the mark of the loose bolts of the output of network are detected according to bolt looseness
The label of bolt state determines that the position of bolt looseness, bolt are loose bolts or normal bolt.
In a preferred real mode of above-mentioned detection method, it is contemplated that vehicle direction is crossed under Practical Project background, in order to increase
Add sample size, in step s 2, before clipping image, the image of acquisition is handled by the method for image enhancement.Specifically
, preferably, being handled using the image enchancing method that Random Level is overturn the image of acquisition, to increase image
Quantity.The above-mentioned detection method of the present invention is not limited to using the image enchancing method of Random Level overturning to the image of acquisition
Reason can also be cut using rotation (referring to Fig. 9 a), fuzzy (referring to Fig. 9 b), contrast normalization (referring to Fig. 9 c), mistake (referring to figure
9d), image enchancing methods such as (referring to Fig. 9 e) are translated to be handled.
In a preferred embodiment of above-mentioned detection method, the critical component rectangular area come is cut out due to cutting out layer
Image is sent into bolt looseness detection network and carries out bolt state (normal or loosening) detection.For the ease of storing and reading, track
Vehicle is generally stored with the small diagram form of fixed size, but there are individual critical components completely to present on a small figure,
These critical components do not have testing conditions due to sufficiently complete, need to be rejected, be not sent into bolt looseness detect network into
Row bolt state (normal or loosening) detection.In order to confirm whether critical component is complete, in step s 4, will be wrapped by cutting out layer
Whether the rectangular area containing critical component is complete to the critical component before being cut out in original track vehicle critical component image
It is whole to be confirmed, the specific steps are that:
Equivalent length of side ratio D is defined, equivalent length of side ratio D is indicated are as follows:
In formula, H is the height of input picture, and W is the width of input picture, (xmin,ymin) it is that critical component positions network output
The left upper apex coordinate of rectangular area, (xmax,ymax) it is the bottom right vertex coordinate that critical component positions that network exports rectangular area；
Definition decides whether that the threshold value cut out is d, and as D > d, the critical component is imperfect, and critical component is imperfect
Picture reject, as D≤d, the critical component is complete, obtains the complete picture of critical component and obtains for being cut out by cutting out layer
Obtain critical component rectangular region image.
In the step S3 of the above-mentioned detection method of the present embodiment, in step S3, convolutional neural networks are by one or more groups of convolution
Composition is connected in series in unit.In one in embodiment, every group of convolution unit include by a convolutional layer conv or several
The convolution layer unit and gone here and there by a non-linear layer relu or several non-linear layers relu that convolutional layer conv is connected in series to form
The non-linear layer unit that connection connection is formed, convolution layer unit the last layer convolutional layer conv and the first layer of non-linear layer unit are non-
Linear layer relu connection；The non-linear layer unit of first group of convolution unit is connect with second group of convolution layer unit, third group volume
The convolution layer unit of product unit is connect with second group of non-linear layer unit, and so on, the convolution of last group of convolution unit
Layer unit is connect with the non-linear layer unit of the second last group.In another embodiment, every group of convolution unit include by
Convolution layer unit that one convolutional layer conv or several convolutional layers conv are connected in series to form, by a non-linear layer relu or
Non-linear layer unit that several non-linear layers relu is connected in series to form and by a maximum pond layer maxpooling or
The pond layer unit that several maximums pond layer maxpooling is connected in series to form, convolution layer unit the last layer convolutional layer
Conv is connect with the first layer non-linear layer relu of non-linear layer unit, the last layer non-linear layer relu of non-linear layer unit
It is connect with the first layer maximum pond layer maxpooling of pond layer unit；The pond layer unit and second of first group of convolution unit
The convolution layer unit connection of group, the convolution layer unit of third group convolution unit is connect with second group of pond layer unit, with such
It pushes away, the convolution layer unit of last group of convolution unit is connect with the pond layer unit of the second last group.
Specifically, the quantity of convolution unit can be the multiple groups such as one group or two groups, four groups, six groups, eight groups.Particular number
It can be depending on the data volume of sample image.Similarly every group of convolution unit includes at least convolution layer unit and non-linear layer list
Member can be equipped with pond layer unit, can also not have pond layer unit, depending on actual conditions.Further, convolution
The number of plies of convolutional layer in layer unit, maximum pond in the number of plies of non-linear layer and pond layer unit unit in non-linear layer unit
The number of plies of layer, can be one layer, is also possible to several layers, depending on actual conditions.
As a preferred embodiment of the present embodiment, referring to Figure 10, convolutional neural networks ZFNet is by 5 groups of convolution units
Composition is connected in series.Convolution layer unit that every group of convolution unit is connected in series to form by several convolutional layers conv, by several
The non-linear layer unit and connected by several maximum pond layer maxpooling that a non-linear layer relu is connected in series to form
The pond layer unit formed is connected, is indicated are as follows:
In formula, ZFNet is expressed as convolutional neural networks, and conv indicates that convolutional layer, relu indicate non-linear layer,
Maxpooling indicates maximum pond layer, indicates full articulamentum, softmax presentation class device.
As a preferred embodiment of the present embodiment, the residual error network uses 18 layers, 34 layers, 50 layers, 101 layers, 152
Layer residual error network, the web results of above-mentioned five kinds of residual error networks are referring to Figure 11.Wherein, it is related to two kinds of building block, joins
See Figure 12 a, 12b.Both structures are directed to ResNet18/34 (Figure 12 a) and ResNet50/101/152 (Figure 12 b) respectively, and one
As total be referred to as one " building block ".Wherein Figure 12 b is also known as " bottleneck design ", purpose
It is the number in order to reduce parameter, 256 dimension channel are dropped to 64 dimensions by the convolution of first 1*1, then finally by 1*1
Convolution is restored, number of parameters on the whole: 1*1*256*64+3*3*64*64+1*1*64*256=69632.If do not used
It is exactly the convolution of two 3*3*256, number of parameters: 3*3*256*256*2=1179648, difference if bottleneck
16.94 times.For conventional ResNet, can be used in 34 layers or less network, for bottleneck design's
ResNet is commonly used in deeper as 101 in network, it is therefore an objective to reduce and calculate and parameter amount.
In the present embodiment, bolt looseness detection network preferably uses 101 layers of residual error network.Specifically, firstly, input
Then the convolution of 7*7*64 is 3 layers by 3+4+23+3=33 building block, each block, so there is 33x 3
=99 layers, finally there is fc layers a (for classifying), so 1+99+1=101 layers, there are 101 layer networks really.It is worth noting that,
101 layers of residual error network only refer to convolution or full articulamentum, and non-linear layer and pond layer do not count.
Another embodiment of the present invention provides a kind of rail vehicle bolt looseness detection methods, examine with described in above-described embodiment
Unlike survey method, in the present embodiment, when training critical component positions network, image will be marked according to the quantitative proportion of 3:1
It is divided into training set and verifying collection, when being verified, calls one-time authentication collection to be verified according to setting interval time, is closing
Deconditioning before fitting occurred in key spots localization network and when training set loss curve has been restrained, after being trained
Key position positions network.If the standard image data collection obtained is sufficiently large, theoretically verifying collects bigger, verification result
It is more accurate.In consideration of it, the result that critical component positions network verification is more accurate in the present embodiment, examined compared to above-described embodiment
Survey method, the positioning of critical component positioning network is more accurate in this implementation.
Detection method described in the present embodiment, when training bolt looseness detection network, the critical component rectangle region that will mark
Area image is divided into training set according to the quantitative proportion of 3:1 and verifying collects, and when being verified, calls according to setting interval time
One-time authentication collection is verified, and before fitting occurred in bolt looseness detection network and training set loss curve has restrained it
Border deconditioning, the bolt looseness after being trained detect network.The result that bolt looseness detects network verification is more accurate, compares
Bolt looseness detection network is more accurate to the detection of bolt looseness in above-described embodiment detection method, this implementation.
Further embodiment of this invention provides a kind of rail vehicle bolt looseness detection method, examines with described in above-described embodiment
Unlike survey method, in the present embodiment, when training critical component positions network, image will be marked according to the quantitative proportion of 5:1
It is divided into training set and verifying collection, when being verified, calls one-time authentication collection to be verified according to setting interval time, is closing
Deconditioning before fitting occurred in key spots localization network and when training set loss curve has been restrained, after being trained
Key position positions network.If the standard image data collection obtained is sufficiently large, theoretically verifying collects bigger, verification result
It is more accurate.In consideration of it, the accuracy that bolt looseness detects the result of network verification is relatively poor, compared to upper in the present embodiment
Two embodiment detection methods are stated, the accuracy that bolt looseness detection network detects bolt looseness in this implementation is relatively poor,
But it compared with prior art, positions more accurate.
Detection method described in the present embodiment, when training bolt looseness detection network, the critical component rectangle region that will mark
Area image is divided into training set according to the quantitative proportion of 5:1 and verifying collects, and when being verified, calls according to setting interval time
One-time authentication collection is verified, and before fitting occurred in bolt looseness detection network and training set loss curve has restrained it
Border deconditioning, the bolt looseness after being trained detect network.The result that bolt looseness detects network verification is more accurate, compares
The positioning accuracy of critical component positioning network is relatively poor in above-mentioned two embodiment detection method, this implementation, but with it is existing
There is technology to compare, bolt looseness detection is more accurate.
Detection method described in the above embodiment of the present invention, independent of the working experience and sense of responsibility of service personnel, significantly
Time cost and human input are saved；Cascade formula network frame is detected using critical component positioning network+bolt looseness,
Initiative converts bolt looseness test problems to target identification problem twice, and detection speed is fast, can be realized real-time online
Detection, detection efficiency is high, and using the automatic learning training of depth network structure, can be good at learning from data to effective
Feature representation, improve the accuracy rate of detection.
One embodiment of the invention provides a kind of rail vehicle bolt looseness detection system, comprising:
Multispectral timesharing light source, for generating the light of different-waveband；
Image acquiring device, for obtaining rail vehicle critical component image；
Image processing module, for by rail vehicle critical component image cutting-out to being sized, and to the rail after cutting out
Road vehicle critical component image is labeled, and obtains corresponding mark image；
Critical component positions network establishment and training module, for building critical component positioning network, and will mark image
It is divided into training set and verifying collection, training set is input to the critical component positioning network built and is learnt, and passes through verifying collection
It is verified, the critical component after being trained positions network；
A layer setup module is cut out, layer is cut out for being arranged, the output information of network is positioned according to critical component, by cutting out
Rectangular area comprising critical component is cut out by layer from original track vehicle critical component image to be come, and critical component rectangle is obtained
Area image；
Bolt looseness detects network establishment and training module, for building bolt looseness detection network, and by critical component
Rectangular region image is divided into training set and verifying collection, training set is input to the critical component rectangular region image network built into
Row study, and verified by verifying collection, the bolt looseness after being trained detects network.
Storage and test module, for store training after critical component positioning network and bolt looseness detection network and
It cuts out layer, and network is positioned by critical component, cuts out layer and bolt looseness detects network and carries out spiral shell to the image obtained in real time
Bolt loosens test, and outputs test result.
In said detecting system, the multispectral timesharing light source is made of multiple groups LED point light source.In a kind of preferred implementation side
LED point light source uses three kinds of colors of red, green, blue in formula, by the LED point light source of three kinds of colors of red, green, blue, provide it is red, green,
Blue three color light, or recall other color of light in addition to red, green, blue three coloured light.In another preferred embodiment, the LED
Point light source is using red, green, blue and white four kinds of colors, wherein passes through red, green, blue and the LED point light of white four kinds of colors
Source provides panchromatic light or red, green, blue three coloured light, or recalls other color of light in addition to red, green, blue three coloured light.For example, with reference to
Figure 13, multispectral timesharing light source include being located on edge circumference and symmetrically arranged red point source 1, green point light source 2, blue
Color dot light source 3 and white pointolite array (including one or more lamp beads) 4 and it is located in the middle high-power white point light
Source 5, the camera 6 for obtaining rail vehicle critical component image is set on the outside of high-power white point light source, and is arranged symmetrically, phase
The quantity of machine includes but is not limited to binocular camera or more mesh cameras.
Said detecting system of the present invention when detecting bolt looseness, provides the light of different-waveband using multispectral timesharing light source
It is illuminated, obtains the image under different-waveband illumination condition, the characteristic information under different illumination conditions can be obtained, be suitable for
Rail vehicle object to be measured is numerous and target locating for background complexity project situation.Network+spiral shell is positioned using critical component simultaneously
Bolt loosens detection cascade formula network frame, initiative to convert target identification twice for bolt looseness test problems and ask
Topic, detection speed is fast, can be realized real-time online detection, and detection efficiency is high, and is instructed using the automatic study of depth network structure
Practice, can be good at the study from data and improve the accuracy rate of detection to effective feature representation.
Above-described embodiment is used to explain the present invention, rather than limits the invention, in spirit and right of the invention
It is required that protection scope in, to any modifications and changes for making of the present invention, both fall within protection scope of the present invention.
Claims (17)
1. a kind of rail vehicle bolt looseness detection method, which is characterized in that the steps include:
S1, image is obtained
Multiple rail vehicle critical component images at bolt are obtained under different illumination conditions；
S2, image procossing
Clipping image
By every image cutting-out to being sized；
Mark image
Every image is manually marked, mark image is obtained and marks file；
S3, foundation and training critical component position network
Using convolutional neural networks ZFNet as input terminal, full articulamentum is output end, according to being sequentially connected with convolutional neural networks
Network, Pooling layers of ROI and full articulamentum are suggested in ZFNet, region, establish critical component positioning network, mark image is drawn
It is divided into training set and verifying collection, training set is input to critical component positioning network and is learnt, and is tested by verifying collection
Card, the critical component after being trained position network；
Layer is cut out in S4, setting
The rail vehicle critical component image information that network output is positioned according to critical component, will include key portion by cutting out layer
The rectangular area of part is cut out from original track vehicle critical component image to be come, and critical component rectangular region image is obtained；
S5, foundation and training bolt looseness detect network
Using residual error network as input terminal, full articulamentum is output end, and network and ROI Pooling are suggested in the region of sequential connection
Layer is plugged in residual error network, and region suggests that network is connect with residual error network the second last layer convolutional layer, Pooling layers of ROI and
The last first layer convolutional layer connection of residual error network, residual error network the last layer convolutional layer are connect with full articulamentum, establish training spiral shell
Bolt loosens detection network, and critical component rectangular region image is divided into training set and verifying collects, training set is input to bolt
It loosens detection network to be learnt, and the bolt looseness after progress verification is trained is collected by verifying and detects network；
S6, image measurement
The rail vehicle critical component image at bolt is obtained in real time as test image, after test image is input to training
Critical component positions network, and the output information of network is positioned according to critical component, is cut out by cutting out layer comprising critical component
The bolt looseness detection network that rectangular area obtains critical component rectangular region image and is input to after training, according to bolt
The output result for loosening detection network judges whether bolt loosens.
2. rail vehicle bolt looseness detection method as described in claim 1, which is characterized in that in step S2, clipping image
Before, it is handled by image of the method for image enhancement to acquisition.
3. rail vehicle bolt looseness detection method as claimed in claim 2, which is characterized in that in step S4, by cutting out
Layer by the rectangular area comprising critical component from be cut out in original track vehicle critical component image come before, to the critical component
Whether completely confirmed, the specific steps are that:
Equivalent length of side ratio D is defined, equivalent length of side ratio D is indicated are as follows:
In formula, H is the height of input picture, and W is the width of input picture, (xmin,ymin) it is that critical component positions network output rectangle
The upper left in region pinpoints coordinate, (xmax,ymax) it is the bottom right vertex coordinate that critical component positions that network exports rectangular area；
Definition decides whether that the threshold value cut out is d, and as D > d, the critical component is imperfect, by the incomplete figure of critical component
Piece is rejected, and as D≤d, the critical component is complete, is obtained the complete picture of critical component and is closed for being cut out by cutting out layer
Key member rectangular region image.
4. the rail vehicle bolt looseness detection method as described in claims 1 to 3 any one, which is characterized in that step S2
In, every image is manually marked using labelImg annotation tool, marks out each critical component of every image
Minimum circumscribed rectangle frame coordinate simultaneously generates corresponding xml mark file.
5. rail vehicle bolt looseness detection method as claimed in claim 4, which is characterized in that in step S3, mark is schemed
As being divided into training set and verifying collection according to the quantitative proportion of 4:1 or 3:1 or 5:1, when being verified, it is spaced according to setting
Time call one-time authentication collection verified, key position positioning network occurred fitting before and training set loss curve
Through deconditioning when restraining, the key position after being trained positions network.
6. rail vehicle bolt looseness detection method as claimed in claim 5, which is characterized in that in step S3, mark image
Characteristic pattern is obtained after convolutional neural networks ZFNet, characteristic pattern is admitted to region and suggests network, suggests in network by region
After full convolution network and non-linear layer relu operation, each 3 × 3 region can be changed to one 512 dimension on characteristic pattern
Feature vector, this feature vector are sent into region and suggest that the full articulamentum in network carries out classification and the candidate frame position of candidate frame
Recurrence, according to the location information of candidate frame, by ROI Pooling on the characteristic pattern that is obtained after convolutional neural networks processing
Layer carries out ROI Pooling, and the full articulamentum that the result of Pooling layers of ROI output enters in critical component positioning network carries out
The classification of best candidate frame and the recurrence of best candidate frame position, the rail vehicle that critical component positions network final output close
Key member image includes the left upper apex coordinate and bottom right vertex coordinate, key of the rectangle frame of one or more complete critical components
Component names and confidence level.
7. rail vehicle bolt looseness detection method as claimed in claim 6, which is characterized in that in step S3, convolutional Neural
Network ZFNet is made of the series connection of one or more groups of convolution units.
8. rail vehicle bolt looseness detection method as claimed in claim 7, which is characterized in that every group of convolution unit is by one
Convolution layer unit that a convolutional layer conv or several convolutional layers conv are connected in series to form and by a non-linear layer relu or
The non-linear layer unit composition that several non-linear layers relu is connected in series to form, convolution layer unit the last layer convolutional layer conv
It is connect with the first layer non-linear layer relu of non-linear layer unit；The non-linear layer unit of first group of convolution unit and second group
The connection of convolution layer unit, the convolution layer unit of third group convolution unit are connect with second group of non-linear layer unit, and so on,
The convolution layer unit of last group of convolution unit is connect with the non-linear layer unit of the second last group.
9. rail vehicle bolt looseness detection method as claimed in claim 7, which is characterized in that every group of convolution unit is by one
If the convolution layer unit that a convolutional layer conv or several convolutional layers conv are connected in series to form, by a non-linear layer relu or
If non-linear layer unit that dry non-linear layer relu is connected in series to form and by a maximum pond layer maxpooling or
The pond layer unit composition that dry maximum pond layer maxpooling is connected in series to form, convolution layer unit the last layer convolutional layer
Conv is connect with the first layer non-linear layer relu of non-linear layer unit, the last layer non-linear layer relu of non-linear layer unit
It is connect with the first layer maximum pond layer maxpooling of pond layer unit；The pond layer unit and second of first group of convolution unit
The convolution layer unit connection of group, the convolution layer unit of third group convolution unit is connect with second group of pond layer unit, with such
It pushes away, the convolution layer unit of last group of convolution unit is connect with the pond layer unit of the second last group.
10. the rail vehicle bolt looseness detection method as described in claims 1 to 3 any one, which is characterized in that step S4
In, when establishing bolt looseness detection network, every critical component rectangular region image is carried out using labelImg annotation tool
Artificial mark, marks out the minimum circumscribed rectangle frame coordinate of each bolt in every critical component rectangular region image and generates
Corresponding xml marks file.
11. rail vehicle bolt looseness detection method as claimed in claim 10, which is characterized in that in step S4, will mark
Good critical component rectangular region image is divided into training set according to the quantitative proportion of 4:1 or 3:1 or 5:1 and verifying collects,
It when being verified, calls one-time authentication collection to be verified according to setting interval time, occurred intending in bolt looseness detection network
Deconditioning, the bolt looseness after being trained detect network before and when training set loss curve has been restrained.
12. rail vehicle bolt looseness detection method as claimed in claim 11, which is characterized in that in step S4, residual error net
The output of the second last layer convolutional layer of network is that network and Pooling layers of ROI shared part, ROI Pooling are suggested in region
The size dimension of layer output characteristic pattern meets the input of the last layer convolutional layer in former residual error network, residual error network the last layer volume
Lamination meets average pond layer Average Pooling, obtains 2048 dimensional features, the full connection into bolt looseness detection network
Layer carries out the classification of candidate frame and the recurrence of candidate frame position, and it includes loose bolts that bolt looseness, which detects network final output,
Rectangle frame position, confidence level and the label for indicating bolt state；The mark bolt state includes normal and two kinds of shapes of loosening
State.
13. rail vehicle bolt looseness detection method as claimed in claim 12, which is characterized in that the residual error network uses
18 layers, 34 layers, 50 layers, 101 layers, 152 layers of residual error network.
14. a kind of rail vehicle bolt looseness detection system characterized by comprising
Multispectral timesharing light source, for generating the light of different-waveband；
Image acquiring device, for obtaining the rail vehicle critical component image under different illumination conditions and obtaining real-time track vehicle
Critical component image；
Image processing module, for by rail vehicle critical component image cutting-out to being sized, and to the railcar after cutting out
Critical component image is labeled, and obtains corresponding mark image；
Critical component positions network establishment and training module, is divided into for building critical component positioning network, and by mark image
Training set is input to the critical component positioning network built and learnt, and carried out by verifying collection by training set and verifying collection
Verifying, the critical component after being trained position network；
A layer setup module is cut out, layer is cut out for being arranged, the output information of network is positioned according to critical component, it will by cutting out layer
Rectangular area comprising critical component is cut out from original track vehicle critical component image to be come, and critical component rectangular area is obtained
Image；
Bolt looseness detects network establishment and training module, for building bolt looseness detection network, and by critical component rectangle
Area image is divided into training set and verifying collection, and training set is input to the critical component rectangular region image network built
It practises, and verified by verifying collection, the bolt looseness after being trained detects network；
Storage and test module, for storing the critical component positioning network after training and bolt looseness detection network and cutting out
Layer, and network positioned by critical component, layer and bolt looseness is cut out and detects network to carry out bolt to the image obtained in real time loose
Dynamic test, and output test result.
15. rail vehicle bolt looseness detection method as claimed in claim 14, which is characterized in that the multispectral timesharing light
Source is made of multiple groups LED point light source.
16. rail vehicle bolt looseness detection method as claimed in claim 15, which is characterized in that the LED point light source is adopted
Red, green, blue three coloured light is provided by the LED point light source of three kinds of colors of red, green, blue with three kinds of colors of red, green, blue, or is recalled
Other color of light in addition to red, green, blue three coloured light.
17. rail vehicle bolt looseness detection method as claimed in claim 15, which is characterized in that the LED point light source is adopted
With red, green, blue and white four kinds of colors, wherein by red, green, blue and the LED point light source of white four kinds of colors, provide
Panchromatic light or red, green, blue three coloured light, or recall other color of light in addition to red, green, blue three coloured light.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910545803.8A CN110246132A (en) | 2019-06-23 | 2019-06-23 | Rail vehicle bolt looseness detection method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910545803.8A CN110246132A (en) | 2019-06-23 | 2019-06-23 | Rail vehicle bolt looseness detection method and system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110246132A true CN110246132A (en) | 2019-09-17 |
Family
ID=67888921
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910545803.8A Pending CN110246132A (en) | 2019-06-23 | 2019-06-23 | Rail vehicle bolt looseness detection method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110246132A (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110766011A (en) * | 2019-12-26 | 2020-02-07 | 南京智莲森信息技术有限公司 | Contact net nut abnormity identification method based on deep multistage optimization |
| CN110781887A (en) * | 2019-10-25 | 2020-02-11 | 上海眼控科技股份有限公司 | License plate screw detection method and device and computer equipment |
| CN110782443A (en) * | 2019-10-23 | 2020-02-11 | 四川大学 | Railway track defect detection method and system |
| CN110826413A (en) * | 2019-10-10 | 2020-02-21 | 深圳供电局有限公司 | Transmission line damper fault detection method and system and storage medium |
| CN111091543A (en) * | 2019-12-12 | 2020-05-01 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon swing bolster spring loss fault target detection method |
| CN111159150A (en) * | 2019-12-19 | 2020-05-15 | 北京文安智能技术股份有限公司 | Data expansion method and device |
| CN111259886A (en) * | 2020-01-08 | 2020-06-09 | 上海眼控科技股份有限公司 | License plate screw detection method, electronic device, computer equipment and storage medium |
| CN111815623A (en) * | 2020-07-28 | 2020-10-23 | 南方电网数字电网研究院有限公司 | Power transmission line cotter pin missing identification method |
| CN112232222A (en) * | 2020-10-19 | 2021-01-15 | 哈尔滨市科佳通用机电股份有限公司 | Bullet train axle box end cover bolt loss fault detection method based on image processing |
| CN112819779A (en) * | 2021-01-29 | 2021-05-18 | 北京明略软件***有限公司 | Method, device and equipment for identifying bolt looseness |
| CN113236653A (en) * | 2021-04-21 | 2021-08-10 | 西南交通大学 | Bolt structure with loosening detection function and bolt loosening detection method |
| CN113298750A (en) * | 2020-09-29 | 2021-08-24 | 湖南长天自控工程有限公司 | Detection method for wheel falling of circular cooler |
| CN113469966A (en) * | 2021-06-25 | 2021-10-01 | 西南交通大学 | Train bolt looseness detection method based on anti-loosening line identification |
| CN113643252A (en) * | 2021-08-10 | 2021-11-12 | 深圳市道通科技股份有限公司 | Method for determining screw torque of vehicle and torque wrench |
| CN113962961A (en) * | 2021-10-22 | 2022-01-21 | 常州市新创智能科技有限公司 | Screw loosening detection method based on high-speed machine |
| CN114723714A (en) * | 2022-04-12 | 2022-07-08 | 西南交通大学 | Image recognition detection method for train bottom bolt |
| CN117593515A (en) * | 2024-01-17 | 2024-02-23 | 中数智科(杭州)科技有限公司 | Bolt loosening detection system and method for railway vehicle and storage medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN205146716U (en) * | 2015-05-11 | 2016-04-13 | 南京文采科技有限责任公司 | Shaker based on timesharing exposure and many spectral imaging |
| CN106600581A (en) * | 2016-12-02 | 2017-04-26 | 北京航空航天大学 | Train operation fault automatic detection system and method based on binocular stereoscopic vision |
| CN109344753A (en) * | 2018-09-21 | 2019-02-15 | 福州大学 | A kind of tiny fitting recognition methods of Aerial Images transmission line of electricity based on deep learning |
| CN109598290A (en) * | 2018-11-22 | 2019-04-09 | 上海交通大学 | A kind of image small target detecting method combined based on hierarchical detection |
| CN109767423A (en) * | 2018-12-11 | 2019-05-17 | 西南交通大学 | A kind of crack detection method of bituminous pavement image |
| CN109784204A (en) * | 2018-12-25 | 2019-05-21 | 江苏大学 | A kind of main carpopodium identification of stacking string class fruit for parallel robot and extracting method |
-
2019
- 2019-06-23 CN CN201910545803.8A patent/CN110246132A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN205146716U (en) * | 2015-05-11 | 2016-04-13 | 南京文采科技有限责任公司 | Shaker based on timesharing exposure and many spectral imaging |
| CN106600581A (en) * | 2016-12-02 | 2017-04-26 | 北京航空航天大学 | Train operation fault automatic detection system and method based on binocular stereoscopic vision |
| CN109344753A (en) * | 2018-09-21 | 2019-02-15 | 福州大学 | A kind of tiny fitting recognition methods of Aerial Images transmission line of electricity based on deep learning |
| CN109598290A (en) * | 2018-11-22 | 2019-04-09 | 上海交通大学 | A kind of image small target detecting method combined based on hierarchical detection |
| CN109767423A (en) * | 2018-12-11 | 2019-05-17 | 西南交通大学 | A kind of crack detection method of bituminous pavement image |
| CN109784204A (en) * | 2018-12-25 | 2019-05-21 | 江苏大学 | A kind of main carpopodium identification of stacking string class fruit for parallel robot and extracting method |
Non-Patent Citations (6)
| Title |
|---|
| GAHYUN SUH等: "Deep Faster R-CNN-based automated detection and localization of multiple types of damage", 《SENSORS AND SMART STRUCTURES TECHNOLOGIES FOR CIVIL, MECHANICAL, AND》 * |
| 张望等: "动车组运行状态智能检测装备", 《铁路计算机应用》 * |
| 李晓慧: "基于多光谱柑桔检测方法研究", 《中国优秀硕士学位论文全文数据库 (信息科技辑)》 * |
| ***等: "基于级联Faster R-CNN的高铁接触网支撑装置等电位线故障检测", 《铁道学报》 * |
| 段俊欢: "基于CIS的图像采集处理设备的硬件设计与实现", 《中国优秀硕士学位论文全文数据库 (信息科技辑)》 * |
| 王子昊: "深度学习在输电铁塔关键部件缺陷检测中的应用研究", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 * |
Cited By (25)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110826413A (en) * | 2019-10-10 | 2020-02-21 | 深圳供电局有限公司 | Transmission line damper fault detection method and system and storage medium |
| CN110826413B (en) * | 2019-10-10 | 2022-01-25 | 深圳供电局有限公司 | Transmission line damper fault detection method and system and storage medium |
| CN110782443A (en) * | 2019-10-23 | 2020-02-11 | 四川大学 | Railway track defect detection method and system |
| CN110781887A (en) * | 2019-10-25 | 2020-02-11 | 上海眼控科技股份有限公司 | License plate screw detection method and device and computer equipment |
| CN111091543A (en) * | 2019-12-12 | 2020-05-01 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon swing bolster spring loss fault target detection method |
| CN111091543B (en) * | 2019-12-12 | 2020-10-09 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon swing bolster spring loss fault target detection method |
| CN111159150A (en) * | 2019-12-19 | 2020-05-15 | 北京文安智能技术股份有限公司 | Data expansion method and device |
| CN110766011A (en) * | 2019-12-26 | 2020-02-07 | 南京智莲森信息技术有限公司 | Contact net nut abnormity identification method based on deep multistage optimization |
| CN111259886A (en) * | 2020-01-08 | 2020-06-09 | 上海眼控科技股份有限公司 | License plate screw detection method, electronic device, computer equipment and storage medium |
| CN111815623A (en) * | 2020-07-28 | 2020-10-23 | 南方电网数字电网研究院有限公司 | Power transmission line cotter pin missing identification method |
| CN111815623B (en) * | 2020-07-28 | 2024-02-23 | 南方电网数字电网研究院有限公司 | Power transmission line cotter pin missing identification method |
| CN113298750A (en) * | 2020-09-29 | 2021-08-24 | 湖南长天自控工程有限公司 | Detection method for wheel falling of circular cooler |
| CN112232222A (en) * | 2020-10-19 | 2021-01-15 | 哈尔滨市科佳通用机电股份有限公司 | Bullet train axle box end cover bolt loss fault detection method based on image processing |
| CN112232222B (en) * | 2020-10-19 | 2021-05-07 | 哈尔滨市科佳通用机电股份有限公司 | Bullet train axle box end cover bolt loss fault detection method based on image processing |
| CN112819779A (en) * | 2021-01-29 | 2021-05-18 | 北京明略软件***有限公司 | Method, device and equipment for identifying bolt looseness |
| CN113236653A (en) * | 2021-04-21 | 2021-08-10 | 西南交通大学 | Bolt structure with loosening detection function and bolt loosening detection method |
| CN113469966B (en) * | 2021-06-25 | 2023-04-18 | 西南交通大学 | Train bolt looseness detection method based on anti-loosening line identification |
| CN113469966A (en) * | 2021-06-25 | 2021-10-01 | 西南交通大学 | Train bolt looseness detection method based on anti-loosening line identification |
| CN113643252A (en) * | 2021-08-10 | 2021-11-12 | 深圳市道通科技股份有限公司 | Method for determining screw torque of vehicle and torque wrench |
| CN113643252B (en) * | 2021-08-10 | 2024-05-31 | 深圳市道通科技股份有限公司 | Screw torque force determining method and torque wrench for vehicle |
| CN113962961A (en) * | 2021-10-22 | 2022-01-21 | 常州市新创智能科技有限公司 | Screw loosening detection method based on high-speed machine |
| CN114723714A (en) * | 2022-04-12 | 2022-07-08 | 西南交通大学 | Image recognition detection method for train bottom bolt |
| CN114723714B (en) * | 2022-04-12 | 2023-04-07 | 西南交通大学 | Image recognition detection method for train bottom bolt |
| CN117593515A (en) * | 2024-01-17 | 2024-02-23 | 中数智科(杭州)科技有限公司 | Bolt loosening detection system and method for railway vehicle and storage medium |
| CN117593515B (en) * | 2024-01-17 | 2024-03-29 | 中数智科(杭州)科技有限公司 | Bolt loosening detection system and method for railway vehicle and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110246132A (en) | Rail vehicle bolt looseness detection method and system | |
| CN104809732B (en) | A kind of power equipment appearance method for detecting abnormality compared based on image | |
| CN109767427A (en) | The detection method of train rail fastener defect | |
| CN109324604A (en) | A kind of intelligent train resultant fault analysis method based on source signal | |
| CN108537154A (en) | Transmission line of electricity Bird's Nest recognition methods based on HOG features and machine learning | |
| CN110399816A (en) | A kind of high-speed train bottom foreign matter detecting method based on Faster R-CNN | |
| CN104268588B (en) | Railway wagon brake shoe pricker loses the automatic testing method of failure | |
| CN106778833A (en) | Small object loses the automatic identifying method of failure under a kind of complex background | |
| CN107610128A (en) | The method for inspecting and device of a kind of oil level indicator | |
| CN109165541A (en) | Coding method for vehicle component in intelligent recognition rail traffic vehicles image | |
| CN105260744A (en) | Automatic on-line diagnosis method for freight train coupler tail cotter position faults and system | |
| CN110334750A (en) | Iron tower of power transmission line bolt corrosion degree image classification recognition methods | |
| CN113516629A (en) | Intelligent detection system for TFDS passing operation | |
| CN107392247A (en) | Atural object safe distance real-time detection method below a kind of power line | |
| CN111311567A (en) | Method for identifying fastener and steel rail diseases of track line image | |
| CN114202540B (en) | Intelligent detection method for split pin defect of high-speed rail contact network | |
| CN111881970A (en) | Intelligent outer broken image identification method based on deep learning | |
| CN114581764B (en) | Underground structure crack disease discriminating method based on deep learning algorithm | |
| CN109389146A (en) | Equipment state feedback method and system based on neural network image sorting algorithm | |
| CN112464846B (en) | Automatic recognition method for abnormal faults of freight train carriage of station | |
| CN106444665A (en) | Fault classification diagnosis method based on non-Gaussian similarity matching | |
| CN111105389A (en) | Detection method for pavement crack by fusing Gabor filter and convolutional neural network | |
| CN115527170A (en) | Method and system for identifying closing fault of door stopper handle of automatic freight car derailing brake device | |
| CN103279823A (en) | Vehicle axle overhaul working procedure detecting method and system | |
| KR20210037199A (en) | Apparatus for dividing, tagging an image and for detecting defect of facilities using the same |
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: 20190917 |
|
| RJ01 | Rejection of invention patent application after publication |