CN109559301A - A kind of EMU end to end enters the twin network method of institute's defects detection - Google Patents
A kind of EMU end to end enters the twin network method of institute's defects detection Download PDFInfo
- Publication number
- CN109559301A CN109559301A CN201811382037.XA CN201811382037A CN109559301A CN 109559301 A CN109559301 A CN 109559301A CN 201811382037 A CN201811382037 A CN 201811382037A CN 109559301 A CN109559301 A CN 109559301A
- Authority
- CN
- China
- Prior art keywords
- target
- defect
- error
- twin network
- grid
- 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.)
- Withdrawn
Links
Classifications
-
- 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
- G06T7/0008—Industrial image inspection checking presence/absence
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- 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
- 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
- G06T2207/30168—Image quality inspection
Abstract
A kind of EMU end to end enters the twin network method of institute's defects detection, the visible detection method is a twin network, including two identical convolutional neural networks and subsequent cyclic convolution layer similar to YOLO V3, cyclic convolution can be obtained by the dot product of Fast Fourier Transform (FFT), characteristic pattern is exported, the positioning of target and the detection of defect are obtained.The present invention provides the twin network method that a kind of EMU end to end enters institute's defects detection, covers the positioning and defects detection of target part, so as to greatly improve Detection accuracy, promote overhaul efficiency.
Description
Technical field
The present invention relates to field of artificial intelligence, enter institute's defects detection more particularly to a kind of EMU end to end
Twin network method.
Background technique
By to the end of the year 2017, China Railway High-speed operating mileage reaches 2.5 ten thousand kilometers, ranks first in the world.It expects
The year two thousand thirty, Motor Car Institute ownership are up to 2000 institutes, and EMU reserved is up to 5000 standard groups or so.EMU operation
Time span and spatial extent determine the necessity and complexity of safety detection.With the increasing of Motor Car Institute and EMU ownership
It is long, with can not contradiction between the service ability of rapid growth it is increasing, thus it is extremely urgent to carry out intelligent measurement.Motor-car
Group enter institute's intelligent measurement refer to motor-car stop after, 360 degree are carried out to Motor train unit body by camera and laser equipment
Comprehensive detection, the maintenance corner of covering level-one maintenance 90% detect defect with artificial intelligence technology analysis data in turn
Process, existing solution are based on the change detecting method that target part+traditional is positioned manually, and overhaul efficiency is low；Using
The deep learning training pattern problem on the low side there are EMU defect sample, therefore applicant proposes a kind of EMU end to end
Enter the twin network method of institute's defects detection to solve problem above.
Summary of the invention
In order to solve problem above, the present invention provides the twin network side that a kind of EMU end to end enters institute's defects detection
Method, cover target part positioning and defects detection, so as to greatly improve Detection accuracy, promoted overhaul efficiency, for up to
This purpose, the present invention provides the twin network method that a kind of EMU end to end enters institute's defects detection, described to move end to end
The twin network that vehicle group enters institute's defects detection includes 2 convolutional neural networks and subsequent cyclic convolution layer, the cyclic convolution
Output characteristic pattern is obtained by the dot product of Fast Fourier Transform (FFT), output includes the positioning of target part and the detection of defect；
The model of the twin network includes two identical convolutional neural networks and subsequent cyclic convolution layer, mould
Type input is two images, and a width is the history image (standard picture) of acquisition, and a width is the figure in same channels collection in worksite
Picture, the output of the cyclic convolution layer are characterized figure, and maximum value is the position of target, and minimum value position is the position of defect
It sets；
The building of the training set of the convolutional neural networks is made of vehicle bottom, vehicle side, roof data three parts, and for not
With the data label target in channel and type and the position of defect, EMU image data base is constructed；
The loss function of the twin network model is made of error of coordinate, IOU error and error in classification；
Shown in formula specific as follows:
Coordinates of targets error
Target IOU error
Target classification error
Defect coordinate error
Defect IOU error
In formula, input picture is divided into S × S grid, if the center of certain target or defect falls into some grid
Lattice, then this grid is just responsible for detecting this object；
Each grid predicts B Bounding Box and its confidence level and C class probability；
Bounding Box information (x, y, w, h) be the center opposing grid position of object or defect offset and
Width and height, are normalized；
Confidence level reflects whether to be defined as comprising target and the accuracy comprising position under target conditions
Wherein Pr (Object) ∈ { 0,1 }；
λcoordFor the weight of position error, λnoobjAnd λnodefectFor the weight of target and defect error in classification；
Judge whether j-th of Bounding Box is responsible for this target in i-th of grid, the Ground with target
The maximum Bounding Box of the IOU of truth Box is responsible for the coordinate prediction of target；
The central point for judging whether there is target is fallen in grid i, containing the center of target in grid, is just responsible for prediction
The class probability of the target；
Judge whether j-th of Bounding Box is responsible for this defect in i-th of grid, the Ground with defect
The maximum Bounding Box of the IOU of truth Box is responsible for the coordinate prediction of defect；
The training of SGD algorithm obtains suitable network parameter, is predicted to obtain the positioning of target and defect, classification,
The output of present networks is characterized figure, directly returns the length and width of BoundingBox.
Number of allocated passenger trains of the present invention according to need to position motor car wheel to, discharge outlet, trailer wheels to, sleeper beam, connecting cable, trailer
To six class targets, the vehicle side data needs to position skirtboard bolt, air-conditioner air outlet, cover board lock, sewage draining exit switch, turns to wheel
Axle bolt, empty spring, vertical damper, anti-snake damper, brake lining, ten class target of passing phase insulator device, the roof data need
Position top antenna, roof high pressure wire jumper, outer hood, air-conditioner wind hood net, cable connector insulator, roof high-tension apparatus, by
Eight class target of pantograph and MUB and pod, the at present building of training set are for vehicle bottom, vehicle side, roof data three parts structure
At.
Defect of the present invention is divided into defect, permeability, missing, foreign matter suspension and five class of other defects, at present defect master
It will be using the above classification.
A kind of EMU end to end of the invention enters the twin network method of institute's defects detection, has the following advantages that；
1) present invention situation on the low side for EMU defect sample, gives a kind of effective defect inspection method, from
And Detection accuracy can be greatlyd improve, promote overhaul efficiency；
2) present invention obtains network parameter by training method end to end, and traditional target part positioning, image are matched
Quasi- and variation detection, defect recognition are unified into a network.
Detailed description of the invention
Fig. 1 is the twin schematic network structure of defects detection of the invention.
Specific embodiment
Present invention is further described in detail with specific embodiment with reference to the accompanying drawing:
The present invention provides the twin network method that a kind of EMU end to end enters institute's defects detection, and it is fixed to cover target part
Position and defects detection, so as to greatly improve Detection accuracy, promote overhaul efficiency.
In conjunction with Fig. 1, the invention proposes the network method that a kind of EMU enters institute's vision-based detection, the vision-based detection sides
Method is that a twin network is followed including two identical convolutional neural networks and subsequent cyclic convolution layer similar to YOLO V3
Ring convolution can obtain characteristic pattern by the dot product of Fast Fourier Transform (FFT), obtain the positioning of target and the detection of defect.
The building of the training set is to constitute for vehicle bottom, vehicle side, roof data three parts, for vehicle bottom, needs to position
Motor car wheel to, discharge outlet, trailer wheels to, sleeper beam, connecting cable, trailer wheels are to six class targets；For vehicle side, need to position skirtboard
Bolt, air-conditioner air outlet, cover board lock, sewage draining exit switch, turn to axle bolt, empty spring, vertical damper, anti-snake damper,
Brake lining, ten class target of passing phase insulator device；For roof, need to position top antenna, roof high pressure wire jumper, outer hood, air-conditioner wind
Hood net, cable connector insulator, roof high-tension apparatus, pantograph and MUB and eight class target of pod.Corresponding defect is total
It is divided into five class defects, permeability, missing, foreign matter suspension and other defects.Data label target and defect for different channels
Type and position construct EMU image data base.
The twin network model includes two identical convolutional neural networks and subsequent cyclic convolution layer, defeated
Enter for two images, a width is the history image (standard picture) of acquisition, and a width is the image in same channels collection in worksite；It follows
The output of ring convolutional layer is characterized figure (feature map), and maximum value is the position of target, and minimum value position is defect
Position.
The loss function of the twin network model is missed by error of coordinate, IOU (Intersection over Union)
Difference and error in classification composition.Shown in formula specific as follows:
Coordinates of targets error
Target IOU error
Target classification error
Defect coordinate error
Defect IOU error
In formula, input picture is divided into S × S grid, if the centre bit of certain target or defect Ground truth
It sets and falls into
Some grid, then this grid is just responsible for detecting this object.Each grid predicts B Bounding Box
And its confidence level and C class probability.Bounding Box information (x, y, w, h) is the center of object or defect
The offset of opposing grid position and width and height, are normalized.Confidence level is reflected whether comprising target and comprising target
In the case of position accuracy, be defined asWherein Pr (Object) ∈ { 0,1 }.λcoordIt is fixed
The weight of position error, λnoobjAnd λnodefectFor the weight of target and defect error in classification；Judge in i-th of grid j-th
Whether Bounding Box is responsible for this target, the maximum Bounding Box of IOU with the Ground truth Box of target
It is responsible for the coordinate prediction of target；The central point for judging whether there is target is fallen in grid i, containing in target in grid
The heart is just responsible for predicting the class probability of the target；Judge whether j-th of Bounding Box is responsible in i-th of grid
This defect is predicted with the maximum Bounding Box of IOU of the Ground truth Box of the defect coordinate for being responsible for defect.
The training of SGD algorithm obtains suitable network parameter, is predicted to obtain the positioning of target and defect, classification.The output of present networks
It is characterized figure, directly returns the length and width of BoundingBox.
The above described is only a preferred embodiment of the present invention, being not the limit for making any other form to the present invention
System, and made any modification or equivalent variations according to the technical essence of the invention, still fall within present invention model claimed
It encloses.
Claims (5)
1. a kind of EMU end to end enters the twin network method of institute's defects detection, the EMU end to end enters institute's defect
The twin network method of detection uses twin network, it is characterised in that: the twin network includes 2 convolutional neural networks with after
Continuous cyclic convolution layer, the cyclic convolution obtain output characteristic pattern by the dot product of Fast Fourier Transform (FFT), and output includes mesh
Mark the positioning of part and the detection of defect；
The model of the twin network includes that two identical convolutional neural networks and subsequent cyclic convolution layer, model are defeated
Enter for two images, a width is the history image (standard picture) of acquisition, and a width is the image in same channels collection in worksite, institute
The output for stating cyclic convolution layer is characterized figure, and maximum value is the position of target part, and minimum value position is the position of defect；
The building of the training set of the convolutional neural networks is made of vehicle bottom, vehicle side, roof data three parts, and for different logical
The data label target part in road and type and the position of defect construct EMU image data base；
The loss function of the twin network model is made of error of coordinate, IOU error and error in classification；
Shown in formula specific as follows:
Coordinates of targets error
Target IOU error
Target classification error
Defect coordinate error
Defect IOU errorFormula
In, input picture is divided into S × S grid, if the center of certain target or defect falls into some grid, this
Grid is just responsible for detecting this object；
Each grid predicts B Bounding Box and its confidence level and C class probability；
Bounding Box information (x, y, w, h) is offset and the width of the center opposing grid position of object or defect
And height, it is normalized；
Confidence level reflects whether to be defined as comprising target and the accuracy comprising position under target conditions
Wherein Pr (Object) ∈ { 0,1 }；
λcoordFor the weight of position error, λnoobjAnd λnodefectFor the weight of target and defect error in classification；
Judge whether j-th of Bounding Box is responsible for this target in i-th of grid, the Ground truth with target
The maximum Bounding Box of the IOU of Box is responsible for the coordinate prediction of target；
The central point for judging whether there is target is fallen in grid i, containing the center of target in grid, is just responsible for predicting the mesh
Target class probability；
Judge whether j-th of Bounding Box is responsible for this defect in i-th of grid, the Ground with defect
The maximum Bounding Box of the IOU of truth Box is responsible for the coordinate prediction of defect；
The training of SGD algorithm obtains suitable network parameter, is predicted to obtain the positioning of target part and defect, classification, Home Network
The output of network is characterized figure, directly returns the length and width of BoundingBox.
2. a kind of EMU end to end according to claim 1 enters the twin network method of institute's defects detection, feature
Be: the number of allocated passenger trains according to need to position motor car wheel to, discharge outlet, trailer wheels to, sleeper beam, connecting cable, trailer wheels are to six classes
Target part.
3. a kind of EMU end to end according to claim 1 enters the twin network method of institute's defects detection, feature
Be: the vehicle side data need to position skirtboard bolt, air-conditioner air outlet, cover board lock, sewage draining exit switch, turn to axle bolt,
Empty spring, vertical damper, anti-snake damper, brake lining, ten class target part of passing phase insulator device.
4. a kind of EMU end to end according to claim 1 enters the twin network method of institute's defects detection, feature
Be: it is exhausted that the roof data need to position top antenna, roof high pressure wire jumper, outer hood, air-conditioner wind hood net, cable connector
Edge, roof high-tension apparatus, pantograph and MUB and eight class target part of pod.
5. a kind of EMU end to end according to claim 1 enters the twin network method of institute's defects detection, feature
Be: the defect is divided into defect, permeability, missing, foreign matter suspension and five class of other defects.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811382037.XA CN109559301A (en) | 2018-11-20 | 2018-11-20 | A kind of EMU end to end enters the twin network method of institute's defects detection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811382037.XA CN109559301A (en) | 2018-11-20 | 2018-11-20 | A kind of EMU end to end enters the twin network method of institute's defects detection |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109559301A true CN109559301A (en) | 2019-04-02 |
Family
ID=65866593
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811382037.XA Withdrawn CN109559301A (en) | 2018-11-20 | 2018-11-20 | A kind of EMU end to end enters the twin network method of institute's defects detection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109559301A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110009622A (en) * | 2019-04-04 | 2019-07-12 | 武汉精立电子技术有限公司 | A kind of display panel open defect detection network and its defect inspection method |
| CN110532886A (en) * | 2019-07-31 | 2019-12-03 | 国网江苏省电力有限公司 | A kind of algorithm of target detection based on twin neural network |
| CN110533654A (en) * | 2019-08-30 | 2019-12-03 | 北京明略软件系统有限公司 | The method for detecting abnormality and device of components |
| CN111179251A (en) * | 2019-12-30 | 2020-05-19 | 上海交通大学 | Defect detection system and method based on twin neural network and by utilizing template comparison |
| 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 |
| CN111325708A (en) * | 2019-11-22 | 2020-06-23 | 济南信通达电气科技有限公司 | Power transmission line detection method and server |
| CN111652228A (en) * | 2020-05-21 | 2020-09-11 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon sleeper beam hole foreign matter detection method |
| CN111855667A (en) * | 2020-07-17 | 2020-10-30 | 成都盛锴科技有限公司 | Novel intelligent train inspection system and detection method suitable for metro vehicle |
| CN112465818A (en) * | 2020-12-18 | 2021-03-09 | 哈尔滨市科佳通用机电股份有限公司 | Method for detecting foreign matter fault of apron board |
| CN112699730A (en) * | 2020-12-01 | 2021-04-23 | 贵州电网有限责任公司 | Machine room character re-identification method based on YOLO and convolution-cycle network |
| CN113269236A (en) * | 2021-05-10 | 2021-08-17 | 青岛理工大学 | Assembly body change detection method, device and medium based on multi-model integration |
-
2018
- 2018-11-20 CN CN201811382037.XA patent/CN109559301A/en not_active Withdrawn
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110009622A (en) * | 2019-04-04 | 2019-07-12 | 武汉精立电子技术有限公司 | A kind of display panel open defect detection network and its defect inspection method |
| CN110532886A (en) * | 2019-07-31 | 2019-12-03 | 国网江苏省电力有限公司 | A kind of algorithm of target detection based on twin neural network |
| CN110533654A (en) * | 2019-08-30 | 2019-12-03 | 北京明略软件系统有限公司 | The method for detecting abnormality and device of components |
| CN111325708A (en) * | 2019-11-22 | 2020-06-23 | 济南信通达电气科技有限公司 | Power transmission line detection method and server |
| CN111325708B (en) * | 2019-11-22 | 2023-06-30 | 济南信通达电气科技有限公司 | Transmission line detection method and server |
| CN111179251B (en) * | 2019-12-30 | 2021-04-02 | 上海交通大学 | Defect detection system and method based on twin neural network and by utilizing template comparison |
| CN111179251A (en) * | 2019-12-30 | 2020-05-19 | 上海交通大学 | Defect detection system and method based on twin neural network and by utilizing template comparison |
| 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 |
| CN111652228A (en) * | 2020-05-21 | 2020-09-11 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon sleeper beam hole foreign matter detection method |
| CN111652228B (en) * | 2020-05-21 | 2020-12-04 | 哈尔滨市科佳通用机电股份有限公司 | Railway wagon sleeper beam hole foreign matter detection method |
| CN111855667A (en) * | 2020-07-17 | 2020-10-30 | 成都盛锴科技有限公司 | Novel intelligent train inspection system and detection method suitable for metro vehicle |
| CN112699730A (en) * | 2020-12-01 | 2021-04-23 | 贵州电网有限责任公司 | Machine room character re-identification method based on YOLO and convolution-cycle network |
| CN112465818A (en) * | 2020-12-18 | 2021-03-09 | 哈尔滨市科佳通用机电股份有限公司 | Method for detecting foreign matter fault of apron board |
| CN113269236A (en) * | 2021-05-10 | 2021-08-17 | 青岛理工大学 | Assembly body change detection method, device and medium based on multi-model integration |
| CN113269236B (en) * | 2021-05-10 | 2022-04-01 | 青岛理工大学 | Assembly body change detection method, device and medium based on multi-model integration |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109559301A (en) | A kind of EMU end to end enters the twin network method of institute's defects detection | |
| Nie et al. | Pavement Crack Detection based on yolo v3 | |
| CN110532889B (en) | Track foreign matter detection method based on rotor unmanned aerial vehicle and YOLOv3 | |
| CN104392212B (en) | The road information detection and front vehicles recognition methods of a kind of view-based access control model | |
| CN106652551A (en) | Parking stall detection method and device | |
| CN117836667A (en) | Static and non-static object point cloud identification method based on road side sensing unit | |
| CN112414309B (en) | High-speed rail contact line height-guiding and pull-out value inspection method based on airborne laser radar | |
| CN107380163A (en) | Automobile intelligent alarm forecasting system and its method based on magnetic navigation | |
| CN113569915B (en) | Multi-strategy rail transit obstacle recognition method based on laser radar | |
| CN102346844B (en) | Device and method for identifying fault of losing screw bolts for truck center plates | |
| CN106741890A (en) | A kind of high-speed railway safety detecting system based on the dual-purpose unmanned plane of empty rail | |
| CN110427993A (en) | Bullet train navigational gap localization method based on meteorologic parameter | |
| CN115605777A (en) | Dynamic target point cloud rapid identification and point cloud segmentation method based on road side sensing unit | |
| CN106096504A (en) | A kind of model recognizing method based on unmanned aerial vehicle onboard platform | |
| CN111723706A (en) | Box type freight car door opening monitoring device and system based on raspberry group | |
| CN114898296A (en) | Bus lane occupation detection method based on millimeter wave radar and vision fusion | |
| CN110083099A (en) | One kind meeting automobile function safety standard automatic Pilot architecture system and working method | |
| CN113788051A (en) | Train on-station running state monitoring and analyzing system | |
| CN115909092A (en) | Light-weight power transmission channel hidden danger distance measuring method and hidden danger early warning device | |
| CN111507975A (en) | Method for detecting abnormity of outdoor insulator of traction substation | |
| CN115965915A (en) | Wagon connecting pull rod fracture fault identification method and system based on deep learning | |
| CN114879160A (en) | Rail foreign matter invasion real-time monitoring method and system based on three-dimensional point cloud data | |
| CN107066501A (en) | A kind of telemetering motor vehicle tail equipment points distributing method based on road similitude | |
| CN106803348A (en) | A kind of overpass vehicle travel time Forecasting Approach for Short-term based on bayonet socket | |
| CN116486359A (en) | All-weather-oriented intelligent vehicle environment sensing network self-adaptive selection method |
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 | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20190402 |
|
| WW01 | Invention patent application withdrawn after publication |