CN110119677A - Carbon fiber composite core cable damage testing method based on image classification network - Google Patents
Carbon fiber composite core cable damage testing method based on image classification network Download PDFInfo
- Publication number
- CN110119677A CN110119677A CN201910246680.8A CN201910246680A CN110119677A CN 110119677 A CN110119677 A CN 110119677A CN 201910246680 A CN201910246680 A CN 201910246680A CN 110119677 A CN110119677 A CN 110119677A
- Authority
- CN
- China
- Prior art keywords
- carbon fiber
- fiber composite
- core cable
- composite core
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N23/00—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00
- G01N23/02—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00 by transmitting the radiation through the material
- G01N23/04—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00 by transmitting the radiation through the material and forming images of the material
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/60—Static or dynamic means for assisting the user to position a body part for biometric acquisition
- G06V40/63—Static or dynamic means for assisting the user to position a body part for biometric acquisition by static guides
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Human Computer Interaction (AREA)
- Chemical & Material Sciences (AREA)
- Analytical Chemistry (AREA)
- Biochemistry (AREA)
- General Health & Medical Sciences (AREA)
- Immunology (AREA)
- Pathology (AREA)
- Image Analysis (AREA)
- Analysing Materials By The Use Of Radiation (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
Abstract
The present invention provide it is a kind of based on the carbon fiber composite core cable damage testing method of image classification network using X imaging by the way of to carbon fiber composite core cable carry out Image Acquisition；Pretreatment operation is carried out to the carbon fiber composite core cable image of acquisition, respectively obtains training sample and detection sample；Determine the input and output of starting residual error network and the structure of middle section, building carbon carbon fiber composite core cable failure modes detect network；Network is detected using training sample training carbon fiber composite core cable failure modes, saves the optimal model of training effect；It is detected using detection sample, detects network class output according to carbon fiber composite core cable failure modes, auto mark out the damage location in image.This method can realize carbon fiber composite core cable damage testing automatically, and detection efficiency and precision are high.
Description
Technical field
The carbon fiber composite core cable damage testing method based on image classification network that the present invention relates to a kind of.
Background technique
As modernization industry is constantly progressive, the electricity consumption of all trades and professions is in geometric growth, and the load that cable is born is more next
It is bigger.Due to the complicated multiplicity of cable local environment, either in construction or use process, it is easy to lead to carbon fiber complex core
There are certain damage situations in part, directly influences normal power supply, the life of strong influence resident.Therefore cable failure position
Detection method is particularly important.
In traditional method for detection fault detection, Chinese DianKeYuan ten thousand is built up with Shandong University Zhu Bo et al., the detection skill of use
Art is mainly three kinds: ray detection, frequency analysis and arc sag monitoring.Through with applied the unit of these technologies to exchange, it is believed that: penetrate
Line detection device is heavy, low efficiency；Frequency analysis is disturbed greatly, and False Rate is high；Arc sag monitoring effect is unobvious.
Ultrasound, ultrasound were respectively adopted for carbon fiber complex core plug and similar carbon fiber complex core component both at home and abroad
The methods of guided wave, vortex, ray, sound emission, optics achieve certain effect, but are limited only to plug, for plug outer layer
After wrapping up aluminium stock, generally it is difficult to be detected.Shandong University adds the additive of radiation-sensitive, middle carbon restoration core in core in plug
The related art schemes such as pre-buried optical fiber also can be realized the monitoring or detection of plug inside stick, but for the transmission line of electricity of storage without
Method is applied to ensure carbon fiber composite core wire transmission line safety, overhauls from manufacturing O&M, there is a series of examination
Detection technique is tested, part of mature application, some is still in the development phase.Wherein conventional detection method is mostly
Be by the way that image is sharpened, greyscale transformation, the image pre-processing method after smooth and Fourier transformation, then carry out two-value
Change processing, Edge extraction and image recognition technology extract defect image.But such method is in carbon fiber composite core cable
It is less applicable in damage testing, because damaged area is not to be difficult to detect it is obvious that be not much different with background contrasts.
It all can not the damaged this kind of damaged and back of effective solution carbon fiber composite core cable based on current all kinds of methods of detection
Minimum situation is not spent at scenic spot, therefore the deep learning method having outstanding performance in computer vision at present is introduced to carbon fiber
The detection of composite core cable failure has a very important significance.
Summary of the invention
The carbon fiber composite core cable damage testing method based on image classification network that the object of the present invention is to provide a kind of
It solves existing in the prior art since background and damaged difference are not difficult to the carbon fiber composite core cable damage testing detected very much
The problem of.
The technical solution of the invention is as follows:
A kind of carbon fiber composite core cable damage testing method based on image classification network, includes the following steps,
S1, it is directed to the characteristics of carbon fiber composite core cable, figure is carried out to carbon fiber composite core cable by the way of X imaging
As acquisition；
S2, the feature distribution according to the carbon fiber composite core cable image entirety of acquisition, to the carbon fiber complex core of acquisition
Cable image carries out pretreatment operation, respectively obtains training sample and detection sample；
S3, according to the sample characteristics of the pretreated training sample of step S2, determine that the input of starting residual error v2 network is defeated
Out and the structure level number of the intermediate convolutional layer for realizing feature extraction functions, building carbon fiber composite core cable failure modes detect net
Network；
Carbon fiber composite core cable breakage point in S4, the training sample training step S3 obtained after being handled using step S2
Class detects network, the optimal model of training effect is saved, as damage testing model；
S5, it is examined using detection sample through sample make after of the damage testing model obtained in step S4 to input
It surveys, network class output is detected according to carbon fiber composite core cable failure modes, by normalization exponential function to the sample of cutting
This type judgement through the detection network output of carbon fiber composite core cable failure modes and the corresponding position positioning in original image, automatically
Mark the damage location in image.
Further, in step S2, the feature distribution situation of the carbon fiber composite core cable image entirety of acquisition is cable
Image acquires the picture space vacant relative to other at whole, and at whole to acquire accounting for for picture smaller, and carbon fiber is compound
Core cable has the tendency that part oblique bending, other vacant spaces are black areflexia region entirely.
Further, in step S2, pretreatment operation is carried out to the carbon fiber composite core cable image of acquisition, specifically,
S21, thunder east is used to the carbon fiber composite core cable for tilting expression in the carbon fiber composite core cable image of acquisition
Transformation sciagraphy finds out inclination angle and then carries out slant correction；
S22, the by hand damaged area in mark carbon fiber composite core cable, automatic cutting original image, generating multiple has part
It is overlapped and includes the samples pictures of carbon fiber complex core carbon core segment.
Further, in step S22, automatic cutting original image generates multiple and overlaps and include carbon fiber complex core
The samples pictures of carbon core segment, specifically,
S221, accurate carbon core image is cut into from x-ray imaging carbon fiber composite core cable original image；
S222, it draws rectangle frame and marks damage location；
S223, according to rectangle frame position, the part of both ends setting ratio is removed, according to step-length cutting image, i.e., every step
Length, the picture of interception setting width and height, wherein accurate carbon core obtained by the sample image center made and step S231
Image is consistent, in the sample cut when cutting image and the rectangle frame marked before when partly overlapping, step-length becomes former
Step-length 1/3, step-length is constant when not being overlapped, this step is equivalent to, increasing more intensive in the damaged area cutting where rectangle frame
More breakage sample sizes.
Further, in step S3, determine that carbon fiber composite core cable failure modes detect network, specifically, according to sample
This figure size, the input of adjustment starting residual error v2 network, according to the pattern of defect image, will without it is damaged, fracture, sawed-off, gap,
Crack is classified respectively, and the output as carbon fiber composite core cable failure modes detection network；Due to the picture of input
Size is reduced, and accordingly reduces starting residual error v2 network down-sampling layer, and last corresponding adjustment global pool layer, output figure
Piece classification results.
Further, in step S3, in step S3, determine that carbon fiber composite core cable failure modes detect network, including
The trunk layer that sets gradually, starting residual error layer A, down-sampling layer A, starting residual error layer B, starting residual error layer C, global pool layer, with
Machine layer, classification layer.
Further, it is determined that in carbon fiber composite core cable failure modes detection network,
Trunk layer: as the structure tentatively extracted to characteristics of image, preliminary extraction feature is carried out, and is tentatively reduced special
Levy vector dimension；
It originates residual error layer A: strengthening characteristic extraction part, increase that its is non-linear by increasing network range, make feature extraction
More effectively；
Down-sampling layer A: making down-sampling effect, and reducing vector dimension reduces calculation amount, and by being added under mulitpath
Sample mode addition is non-linear, improves the validity of feature extraction；
Starting residual error layer B: strengthening characteristic extraction part again, increases that its is non-linear by increasing network range, makes feature
It extracts more effective；
Originate residual error layer C: the characteristic extraction part before strengthening again increases that its is non-linear by increasing network range,
Keep feature extraction more effective；
Global pool layer: retain notable feature, reduce characteristic dimension, increase the receptive field of convolution kernel；
Random layer: as over-fitting structure is reduced, generalization is improved；
Classification layer: realizing classification feature, according to class probability is calculated, judges to export last classification, and output situation is made
Finally to export；
Wherein associated order is successively attached from top to bottom, wherein starting residual error layer A is recycled 5 times, starting residual error layer B is followed
Ring 10 times, starting residual error layer C is recycled 5 times.
Further, in step S4, the carbon fiber in the training sample training step S3 obtained after step S2 processing is used
When composite core cable failure classification and Detection network, cross entropy loss function used by training:
Wherein, Loss is loss, SjIt is j-th of value of the output vector S of cross entropy loss function, expression is this sample
Originally belong to the probability of j-th of classification, yjIndicate the classification situation under true tag, k represents k-th of sample, and T representative sample is total
Number,Feature extraction before indicating k-th sample input carbon fiber composite core cable failure modes detection network class part to
Volume index value, similarlyFeature extraction vector index value before indicating j-th sample input classified part, in training process, net
Network successively calculates renewal amount according to back-propagation algorithm and Adam's optimization algorithm to update network according to the numerical value of loss function
Weight and biasing save the optimal model of training effect, as best model.
Further, in step S5, network class output is detected according to carbon fiber composite core cable failure modes, it is automatic to mark
The damage location in image out is remembered, specifically, the carbon fiber composite core cable image detected is carried out sample production
Afterwards, by treated, cutting image addition model is detected, and unabroken cutting image will be classified as in image to be distinguished, right
Should into original image corresponding position, find the corresponding damaged area of original image, mark, which is set to damaged area.
The beneficial effects of the present invention are:
One, carbon fiber composite core cable damage testing method of this kind based on image classification network, by using X-ray at
Picture is put forward for the first time the automatic detection that depth learning technology is used for carbon fiber composite core cable breakage；And it is put forward for the first time image
Sorting technique is used for the fast automatic positioning of damaged area, can realize carbon fiber composite core cable damage testing automatically, and examine
It surveys efficiency and precision is high, it is ineffective using general flaw detection imaging method to solve current carbon fiber composite core cable, and need
The problem of manual operation manual identified after certain image procossing.
Two, carbon fiber composite core cable damage testing method of this kind based on image classification network, in advance answers carbon fiber
It closes core cable image to be handled, carbon fiber composite core cable carbon core in training sample is made to be in picture centre position, while sample
This size becomes smaller, and can increase substantially the accuracy rate when efficiency and detection of sample training, and sample has the cutting of overlapping to enable detection
Possess the function of automatic positioning damage location, is similar to sliding window and positions, by there is the cutting image of overlapping to carry out one by one to these
The image that classification output is not zero is set to damage location by classification, is mapped to label in original image and is.
Three, carbon fiber composite core cable damage testing method of this kind based on image classification network, targetedly improves
Carbon fiber composite core cable failure modes detect network structure, so that training process is more suitable for and is currently maked sample, improve inspection
Survey efficiency and precision.
Four, in the present invention, gone out by the sample extraction that would be classified as damaged, be equivalent to while navigating to damage location, this
Kind method is compared to positioning function in mask-r-cnn series, and operating rate is faster.
Detailed description of the invention
Fig. 1 is the process of carbon fiber composite core cable damage testing method of the embodiment of the present invention based on image classification network
Schematic diagram.
Fig. 2 is that the starting residual error network infrastructure of carbon fiber composite core cable failure modes detection network in embodiment shows
It is intended to.
Fig. 3 is that trunk layer illustrates schematic diagram in embodiment.
Fig. 4 is to originate residual error layer A in embodiment to illustrate schematic diagram.
Fig. 5 is to originate residual error layer B in embodiment to illustrate schematic diagram.
Fig. 6 is to originate residual error layer C in embodiment to illustrate schematic diagram.
Fig. 7 is that down-sampling layer A illustrates schematic diagram in embodiment.
Fig. 8 is that specific example illustrates schematic diagram in embodiment.
Specific embodiment
The preferred embodiment that the invention will now be described in detail with reference to the accompanying drawings.
Embodiment
A kind of carbon fiber composite core cable damage testing method based on image classification network, such as Fig. 1 and Fig. 8, including with
Lower step,
S1, it is directed to the characteristics of carbon fiber composite core cable, figure is carried out to carbon fiber composite core cable by the way of X imaging
As acquisition.
S2, the feature distribution according to the carbon fiber composite core cable image entirety of acquisition, to the carbon fiber complex core of acquisition
Cable image carries out pretreatment operation, respectively obtains training sample and detection sample.In embodiment, the size of samples pictures is preferred
For 100 pixel *, 43 pixel.
In step S2, the feature distribution situation of the carbon fiber composite core cable image entirety of acquisition is cable image at whole
It is smaller that the acquisition picture space vacant relative to other at whole acquires accounting for for picture, and carbon fiber composite core cable has portion
Divide the tendency of oblique bending, other vacant spaces are black areflexia region entirely.
In step S2, pretreatment operation is carried out to the carbon fiber composite core cable image of acquisition, specifically,
S21, thunder east is used to the carbon fiber composite core cable for tilting expression in the carbon fiber composite core cable image of acquisition
Transformation sciagraphy finds out inclination angle and then carries out slant correction；
S23, the by hand damaged area in mark carbon fiber composite core cable, automatic cutting original image, generating multiple has part
It is overlapped and includes the samples pictures of carbon fiber complex core carbon core segment.
Step S22 specifically,
S221, accurate carbon core image is cut into from x-ray imaging carbon fiber composite core cable original image；
S222, it draws rectangle frame and marks damage location；
S223, according to rectangle frame position, remove the part of both ends setting ratio preferably 1/10, according to step-length cutting image,
I.e. every step length, the picture of interception setting width and height, wherein obtained by the sample image center made and step S231
The corresponding position of accurate carbon core image is consistent, the part in the sample cut when cutting image and the rectangle frame marked before
When overlapping, step-length becomes former step-length 1/3, and step-length is constant when not being overlapped, this step is equivalent in the damage zone where rectangle frame
Domain cutting is more intensive, increases damaged sample size.
Carbon fiber composite core cable damage testing method of this kind based on image classification network, in advance by carbon fiber complex core
Cable image is handled, and carbon fiber composite core cable carbon core in training sample is made to be in picture centre position, while sample ruler
It is very little to become smaller, the accuracy rate when efficiency and detection of sample training can be increased substantially, sample has the cutting of overlapping that detection is enabled to possess
It is automatically positioned the function of damage location, sliding window is similar to and positions, by there is the cutting image of overlapping to classify one by one to these,
The image that classification output is not zero is set to damage location, being mapped to label in original image is.
S3, according to the sample characteristics of the pretreated training sample of step S2, determine that the input of starting residual error v2 network is defeated
Out and the structure level number of the intermediate convolutional layer for realizing feature extraction functions, building carbon fiber composite core cable failure modes detect net
Network；Such as Fig. 2, centre realizes that the structure level number of the convolutional layer of feature extraction functions includes the trunk layer set gradually, starting residual error
Layer A, down-sampling layer A, starting residual error layer B, starting residual error layer C, global pool layer, random layer, classification layer.
In step S3, according to sample graph size (100 pixel *, 43 pixel), the input (100 of adjustment starting residual error v2 network
1 channel 43 pixel * of pixel *), it will be 0 without failure modes, fractureing is divided into 1, sawed-off be divided into 2, be empty according to the pattern of defect image
It is 4 output as network that gap, which is divided into 3, crack,；Since the dimension of picture of input is reduced, residual error v2 network will be originated accordingly
Down-sampling layer is reduced, and last corresponding adjustment global pool layer pond degree, exports picture classification result.
Wherein for down-sampling layer the purpose is to reduce characteristic image size to reduce calculation amount, down-sampling layer corresponds to carbon fiber
Down-sampling A module section in composite core damage testing network；The locations of structures of global pool layer is in carbon fiber composite core cable
Failure modes detect network as shown in Fig. 2, being 1 pixel *, 1 Pixel Dimensions vector by feature vector down-sampling, and more preferable extraction is semantic,
So as to classification later.
In step S3, determine that carbon fiber composite core cable failure modes detect network, structure specifically:
Trunk layer: as the structure tentatively extracted to characteristics of image, preliminary extraction feature is carried out, and is tentatively reduced special
Levy vector dimension；Such as Fig. 3.
It originates residual error layer A: strengthening characteristic extraction part, increase that its is non-linear by increasing network range, make feature extraction
More effectively；Such as Fig. 4, the structured loop 5 times；
Down-sampling layer A: making down-sampling effect, and reducing vector dimension reduces calculation amount, and by being added under mulitpath
Sample mode addition is non-linear, improves the validity of feature extraction；Such as Fig. 7, make down-sampling effect；
Starting residual error layer B: strengthening characteristic extraction part again, increases that its is non-linear by increasing network range, makes feature
It extracts more effective；Such as Fig. 5, the structured loop 5 times；
Starting residual error layer C strengthens again before characteristic extraction part, increase that its is non-linear by increase network range, make
Feature extraction is more effective；
Global pool layer: retain notable feature, reduce characteristic dimension, increase the receptive field of convolution kernel；
Random layer: as over-fitting structure is reduced, generalization is effectively improved；
Classification layer: realizing classification feature, according to class probability is calculated, judges to export last classification, and output situation is made
Finally to export；
Wherein associated order is successively attached from top to bottom, wherein starting residual error layer A is recycled 5 times, starting residual error layer B is followed
Ring 10 times, starting residual error layer C is recycled 5 times.
Input and Output respectively indicate input and output in the structure of Fig. 3-Fig. 7；Concat layers of Filter effect are groups
Layer is closed, the input of each path is stitched together；BN layers are Batch_normalization, i.e. batch preparation standards layer；?
Z-score normalized is carried out to the output data of image classification network middle layer when image classification network training, i.e.,
Subtract mean value except variance layer；Relu Activation represents line rectification function, detects for carbon fiber composite core cable failure modes
The activation primitive that network uses, expression formula are relu (x)=max (0, x)；Conv indicates convolution kernel, and wherein stride is indicated
Step-length number, i.e. sampling interval, size indicate convolution kernel size (such as 3*3*32, stride=2 indicate 3 pixel *, 3 picture
Plain * 32 channel, step-length 2), the convolution mode being not particularly illustrated uses all the filling mode for being input front and back with size, defeated
Enter that size is identical with Output Size, the activation primitive for there are linear printed words to use in part of convolution is linearly to activate, and is not had
The convolution activation primitive of specified otherwise is line rectification activation；Maxpool indicates global pool, size Expressing mode and volume
Product nuclear phase is same.
Carbon fiber composite core cable damage testing method of this kind based on image classification network, targetedly improves network
Structure makes training process be more suitable for and currently make sample, improves detection efficiency and precision.
Carbon fiber composite core cable breakage point in S4, the training sample training step S3 obtained after being handled using step S2
Class detects network, the optimal model of training effect is saved, as damage testing model.Wherein, training include training package include before to
It propagates and backpropagation, the optimal model of final preservation training effect, supplying step S5 test uses；
In step S4, in step S4, the carbon fiber in the training sample training step S3 obtained after step S2 processing is used
When composite core cable failure classification and Detection network, cross entropy loss function used by training:
Wherein, Loss is loss, SjIt is j-th of value of the output vector S of cross entropy loss function, expression is this sample
Originally belong to the probability of j-th of classification, yjIndicate the classification situation under true tag, k represents k-th of sample, and T representative sample is total
Number,Feature extraction before indicating k-th sample input carbon fiber composite core cable failure modes detection network class part to
Volume index value, similarlyFeature extraction vector index value before indicating j-th sample input classified part, in training process, net
Network successively calculates renewal amount according to back-propagation algorithm and Adam's optimization algorithm to update carbon fiber according to the numerical value of loss function
The weight of composite core cable failure classification and Detection network and biasing save the optimal model of training effect, as best model.
S5, it is examined using detection sample through sample make after of the damage testing model obtained in step S4 to input
It surveys, network class output is detected according to carbon fiber composite core cable failure modes, by normalization exponential function to the sample of cutting
This auto marks out the damage location in image through the type judgement of network output and the corresponding position positioning in original image.
In step S5, network class output is detected according to carbon fiber composite core cable failure modes, auto marks out image
In damage location, specifically, by the carbon fiber composite core cable image detected carry out sample production after, will handle
Cutting image afterwards is added model and is detected, and unabroken cutting image will be classified as in image and is distinguished, original image is corresponded to
Middle corresponding position, finds the corresponding damaged area of original image, marks, which is set to damaged area.
Embodiment is gone out by would be classified as damaged sample extraction, is equivalent to while navigating to damage location, this method
Compared to positioning function in mask-r-cnn series, and operating rate is faster；
In embodiment, carbon fiber composite core cable failure modes detect each parameter setting of network: the number of iterations is 50 times, batch
Secondary size is 128, and initial learning rate is 0.0001, and loss function uses cross entropy loss function, and optimizer is Adam's optimizer
Image carbon fiber composite core cable failure modes detection network is created with this to be trained and test.
Explanation is made that the specific implementation of this method above, certain this method can also there are other a variety of specific embodiment parties
Formula, those skilled in the art can make various changes and deformation, but this on the premise of without prejudice to spirit of the invention
A little change should be contained in the required protection scope limited of the application patent with deformation.
Claims (8)
1. a kind of carbon fiber composite core cable damage testing method based on image classification network, it is characterised in that: including following
Step,
S1, it is directed to the characteristics of carbon fiber composite core cable, image is carried out to carbon fiber composite core cable by the way of X imaging and is adopted
Collection；
S2, the feature distribution according to the carbon fiber composite core cable image entirety of acquisition, to the carbon fiber composite core cable of acquisition
Image carries out pretreatment operation, respectively obtains training sample and detection sample；
S3, according to step S2 production after training sample sample characteristics, determine starting residual error v2 network input and output and in
Between realize feature extraction functions convolutional layer structure level number, including set gradually trunk layer, starting residual error layer A, down-sampling layer
A, residual error layer B, starting residual error layer C, global pool layer, random layer, classification layer, building carbon fiber composite core cable breakage point are originated
Class detects network；
Carbon fiber composite core cable failure modes in S4, the training sample training step S3 obtained after being handled using step S2 are examined
Survey grid network saves the optimal model of training effect, as damage testing model；
S5, it is detected using detection sample through sample make after of the damage testing model obtained in step S4 to input,
Network class output is detected according to carbon fiber composite core cable failure modes, is passed through by sample of the normalization exponential function to cutting
The type judgement of network output and the corresponding position positioning in original image, auto mark out the damage location in image.
2. the carbon fiber composite core cable damage testing method based on image classification network as described in claim 1, feature
Be: in step S2, the feature distribution situation of the carbon fiber composite core cable image entirety of acquisition is that cable image is adopted at whole
It is smaller that the collection picture space vacant relative to other at whole acquires accounting for for picture, and carbon fiber composite core cable has part
The tendency of oblique bending, other vacant spaces are black areflexia region entirely.
3. the carbon fiber composite core cable damage testing method based on image classification network as described in claim 1, feature
It is: in step S2, pretreatment operation is carried out to the carbon fiber composite core cable image of acquisition, specifically,
S21, Radon transform is used to the carbon fiber composite core cable for tilting expression in the carbon fiber composite core cable image of acquisition
Sciagraphy finds out inclination angle and then carries out slant correction；
S22, the damaged area in mark carbon fiber composite core cable, automatic cutting original image generate multiple and overlap by hand
It and include the samples pictures of carbon fiber complex core carbon core segment.
4. the carbon fiber composite core cable damage testing method based on image classification network as claimed in claim 3, feature
Be: in step S22, automatic cutting original image generates multiple samples for overlapping and include carbon fiber complex core carbon core segment
This picture, specifically,
S221, accurate carbon core image is cut into from x-ray imaging carbon fiber composite core cable original image；
S222, it draws rectangle frame and marks damage location；
S223, according to rectangle frame position, remove the part of both ends setting ratio, it is according to step-length cutting image, i.e., long every step-length
Degree, the picture of interception setting width and height, wherein accurate carbon core image obtained by the sample image center made and step S231
Unanimously, in the sample cut when cutting image and the rectangle frame marked before when partly overlapping, step-length becomes former step-length
1/3, step-length is constant when not being overlapped, this step be equivalent to where rectangle frame damaged area cutting it is more intensive, increase brokenly
Damage sample size.
5. in step S3, carbon fiber composite core cable failure modes detect the intermediate convolution for realizing feature extraction functions in network
The structure level number of layer, including trunk layer, starting residual error layer A, down-sampling layer A, the starting residual error layer B, starting residual error set gradually
Layer C, global pool layer, random layer, classification layer.
6. the carbon fiber composite core cable damage testing method based on image classification network as claimed in claim 5, feature
It is: determines in carbon fiber composite core cable failure modes detection network,
Trunk layer: as the structure tentatively extracted to characteristics of image, carry out it is preliminary extract feature, and tentatively reduce feature to
It takes measurements；
It originates residual error layer A: strengthening characteristic extraction part, increase that its is non-linear by increasing network range, make feature extraction more
Effectively；
Down-sampling layer A: making down-sampling effect, and reducing vector dimension reduces calculation amount, and the down-sampling by the way that mulitpath is added
Mode is added non-linear, improves the validity of feature extraction；
Starting residual error layer B: strengthening characteristic extraction part again, increases that its is non-linear by increasing network range, makes feature extraction
More effectively；
Originate residual error layer C: the characteristic extraction part before strengthening again increases that its is non-linear by increasing network range, makes spy
Sign is extracted more effective；
Global pool layer: retain notable feature, reduce characteristic dimension, increase the receptive field of convolution kernel；
Random layer: as over-fitting structure is reduced, generalization is improved；
Classification layer: realizing classification feature, according to class probability is calculated, judges to export last classification, will export situation as most
After export,
Wherein associated order is successively attached from top to bottom, wherein starting residual error layer A is recycled 5 times, starting residual error layer B circulation 10
Secondary, starting residual error layer C is recycled 5 times.
7. the carbon fiber composite core cable damage testing side as claimed in any one of claims 1 to 6 based on image classification network
Method, it is characterised in that: in step S4, the carbon fiber in training sample training step S3 obtained after being handled using step S2 is compound
When core cable failure classification and Detection network, cross entropy loss function used by training:
Wherein, Loss is loss, SjIt is j-th of value of the output vector S of cross entropy loss function, expression is this sample category
In the probability of j-th of classification, yjIndicating the classification situation under true tag, k represents k-th of sample, and T representative sample is total,
Feature extraction vector index before indicating k-th of sample input carbon fiber composite core cable failure modes detection network class part
Value, similarlyFeature extraction vector index value before indicating j-th sample input classified part, in training process, carbon fiber is multiple
Core cable failure classification and Detection network is closed according to the numerical value of loss function, it is layer-by-layer according to back-propagation algorithm and Adam's optimization algorithm
Renewal amount is calculated to update the weight of carbon fiber composite core cable failure modes detection network and biasing, it is optimal to save training effect
Model, as best model.
8. the carbon fiber composite core cable damage testing side as claimed in any one of claims 1 to 6 based on image classification network
Method, it is characterised in that: in step S5, detect network output according to carbon fiber composite core cable failure modes, auto mark out figure
Damage location as in, specifically, will locate after the carbon fiber composite core cable image detected is carried out sample production
Cutting image after reason is added model and is detected, and unabroken cutting image will be classified as in image and is distinguished, original is corresponded to
Corresponding position in figure, finds the corresponding damaged area of original image, marks, which is set to damaged area.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910246680.8A CN110119677B (en) | 2019-03-28 | 2019-03-28 | Image classification network-based carbon fiber composite core cable damage detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910246680.8A CN110119677B (en) | 2019-03-28 | 2019-03-28 | Image classification network-based carbon fiber composite core cable damage detection method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110119677A true CN110119677A (en) | 2019-08-13 |
| CN110119677B CN110119677B (en) | 2022-03-11 |
Family
ID=67520586
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910246680.8A Active CN110119677B (en) | 2019-03-28 | 2019-03-28 | Image classification network-based carbon fiber composite core cable damage detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110119677B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111046943A (en) * | 2019-12-09 | 2020-04-21 | 国网智能科技股份有限公司 | Method and system for automatically identifying state of isolation switch of transformer substation |
| CN111583197A (en) * | 2020-04-23 | 2020-08-25 | 浙江大学 | Power box picture rust damage identification method combining SSD and Resnet50 network |
| CN113487534A (en) * | 2021-05-20 | 2021-10-08 | 浙江大华技术股份有限公司 | Method and device for detecting package damage and computer readable storage medium |
| CN114140637A (en) * | 2021-10-21 | 2022-03-04 | 阿里巴巴达摩院(杭州)科技有限公司 | Image classification method, storage medium and electronic device |
| CN114202541A (en) * | 2022-02-17 | 2022-03-18 | 湖北中海电线电缆有限公司 | Cable defect detection method based on artificial intelligence |
| CN116318406A (en) * | 2023-05-19 | 2023-06-23 | 深圳市飞思卓科技有限公司 | Signal compensation method and system for optical fiber communication system |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106875381A (en) * | 2017-01-17 | 2017-06-20 | 同济大学 | A kind of phone housing defect inspection method based on deep learning |
| CN107941828A (en) * | 2018-01-03 | 2018-04-20 | 国网安徽省电力有限公司电力科学研究院 | A kind of power cable and connector nondestructive detection system and method based on X-ray Real Time Imaging Technology |
| CN109461149A (en) * | 2018-10-31 | 2019-03-12 | 泰州市创新电子有限公司 | The intelligent checking system and method for lacquered surface defect |
| CN109934821A (en) * | 2019-03-22 | 2019-06-25 | 杭州睿工科技有限公司 | A kind of part defect detection method and system |
-
2019
- 2019-03-28 CN CN201910246680.8A patent/CN110119677B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106875381A (en) * | 2017-01-17 | 2017-06-20 | 同济大学 | A kind of phone housing defect inspection method based on deep learning |
| CN107941828A (en) * | 2018-01-03 | 2018-04-20 | 国网安徽省电力有限公司电力科学研究院 | A kind of power cable and connector nondestructive detection system and method based on X-ray Real Time Imaging Technology |
| CN109461149A (en) * | 2018-10-31 | 2019-03-12 | 泰州市创新电子有限公司 | The intelligent checking system and method for lacquered surface defect |
| CN109934821A (en) * | 2019-03-22 | 2019-06-25 | 杭州睿工科技有限公司 | A kind of part defect detection method and system |
Non-Patent Citations (1)
| Title |
|---|
| CHRISTIAN SZEGEDY ET AL.: "Inception-v4,Inception-ResNet and the Impact of Residual Connections on Learning", 《ARXIV》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111046943A (en) * | 2019-12-09 | 2020-04-21 | 国网智能科技股份有限公司 | Method and system for automatically identifying state of isolation switch of transformer substation |
| CN111583197A (en) * | 2020-04-23 | 2020-08-25 | 浙江大学 | Power box picture rust damage identification method combining SSD and Resnet50 network |
| CN111583197B (en) * | 2020-04-23 | 2022-05-13 | 浙江大学 | Power box picture rust damage identification method combining SSD and Resnet50 network |
| CN113487534A (en) * | 2021-05-20 | 2021-10-08 | 浙江大华技术股份有限公司 | Method and device for detecting package damage and computer readable storage medium |
| CN113487534B (en) * | 2021-05-20 | 2022-08-09 | 浙江大华技术股份有限公司 | Method and device for detecting package damage and computer readable storage medium |
| CN114140637A (en) * | 2021-10-21 | 2022-03-04 | 阿里巴巴达摩院(杭州)科技有限公司 | Image classification method, storage medium and electronic device |
| CN114140637B (en) * | 2021-10-21 | 2023-09-12 | 阿里巴巴达摩院(杭州)科技有限公司 | Image classification method, storage medium and electronic device |
| CN114202541A (en) * | 2022-02-17 | 2022-03-18 | 湖北中海电线电缆有限公司 | Cable defect detection method based on artificial intelligence |
| CN114202541B (en) * | 2022-02-17 | 2022-04-29 | 湖北中海电线电缆有限公司 | Cable defect detection method based on artificial intelligence |
| CN116318406A (en) * | 2023-05-19 | 2023-06-23 | 深圳市飞思卓科技有限公司 | Signal compensation method and system for optical fiber communication system |
| CN116318406B (en) * | 2023-05-19 | 2023-08-08 | 深圳市飞思卓科技有限公司 | Signal compensation method and system for optical fiber communication system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110119677B (en) | 2022-03-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110119677A (en) | Carbon fiber composite core cable damage testing method based on image classification network | |
| CN110956126B (en) | Small target detection method combined with super-resolution reconstruction | |
| CN110349126A (en) | A kind of Surface Defects in Steel Plate detection method based on convolutional neural networks tape label | |
| CN107564002A (en) | Plastic tube detection method of surface flaw, system and computer-readable recording medium | |
| CN102590247A (en) | Steel wire rope defect detection method based on X-ray image processing technology | |
| CN109872303A (en) | Surface defect visible detection method, device and electronic equipment | |
| CN106530271A (en) | Infrared image significance detection method | |
| CN111783616B (en) | Nondestructive testing method based on data-driven self-learning | |
| Li et al. | TireNet: A high recall rate method for practical application of tire defect type classification | |
| CN110390341B (en) | Convolutional neural network-based image recognition method for radioactive source of green traffic vehicle carried goods | |
| CN114549446A (en) | Cylinder sleeve defect mark detection method based on deep learning | |
| CN109697459A (en) | One kind is towards optical coherence tomography image patch Morphology observation method | |
| CN115439693A (en) | Training method of target recognition network model, electronic device and program product | |
| Kähler et al. | Anomaly detection for industrial surface inspection: Application in maintenance of aircraft components | |
| US20230341354A1 (en) | Intelligent inversion method for pipeline defects based on heterogeneous field signals | |
| CN113628170A (en) | Laser line extraction method and system based on deep learning | |
| Li et al. | Wafer crack detection based on yolov4 target detection method | |
| Zhang et al. | FDNet: Lightweight train image fault detection network in edge computing environments | |
| CN116721291A (en) | Metal surface defect detection method based on improved YOLOv7 model | |
| CN116630263A (en) | Weld X-ray image defect detection and identification method based on deep neural network | |
| Zhang et al. | An improved DCGAN for fabric defect detection | |
| CN112508862B (en) | Method for enhancing magneto-optical image of crack by improving GAN | |
| CN116596878B (en) | Strip steel surface defect detection method, system, electronic equipment and medium | |
| CN117132541A (en) | Weak supervision fabric surface flaw identification method based on VGG-9 reduced network | |
| Xiong et al. | Research on a method of X-ray graph analysis for power equipment |
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 | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20201215 Address after: No.2, Sipailou, Xinjiekou Street, Xuanwu District, Nanjing, Jiangsu Province Applicant after: SOUTHEAST University Applicant after: STATE GRID JIANGSU ELECTRIC POWER Research Institute Address before: 210000 No. 2, Four Architecture Building, Xinjiekou Street, Xuanwu District, Nanjing City, Jiangsu Province Applicant before: SOUTHEAST University |
|
| TA01 | Transfer of patent application right | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |