CN107451997A - A kind of automatic identifying method of the welding line ultrasonic TOFD D scanning defect types based on deep learning - Google Patents
A kind of automatic identifying method of the welding line ultrasonic TOFD D scanning defect types based on deep learning Download PDFInfo
- Publication number
- CN107451997A CN107451997A CN201710642922.6A CN201710642922A CN107451997A CN 107451997 A CN107451997 A CN 107451997A CN 201710642922 A CN201710642922 A CN 201710642922A CN 107451997 A CN107451997 A CN 107451997A
- Authority
- CN
- China
- Prior art keywords
- networks
- network
- defect
- rpn
- tofd
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
- G01N29/04—Analysing solids
- G01N29/06—Visualisation of the interior, e.g. acoustic microscopy
- G01N29/0654—Imaging
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
- G01N29/22—Details, e.g. general constructional or apparatus details
- G01N29/26—Arrangements for orientation or scanning by relative movement of the head and the sensor
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N2291/00—Indexing codes associated with group G01N29/00
- G01N2291/26—Scanned objects
- G01N2291/267—Welds
-
- 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
- G06T2207/10132—Ultrasound 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
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Chemical & Material Sciences (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Analytical Chemistry (AREA)
- Biochemistry (AREA)
- Immunology (AREA)
- Pathology (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Acoustics & Sound (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of automatic identifying method of the welding line ultrasonic TOFD D scanning defect types based on depth learning technology.First, the D scan image datas of weld seam typical defect are gathered by ultrasonic TOFD D scanning techniques；Secondly, built in MATLAB programmed environments and propose that network RPN and FRCN network forms FasterR CNN deep learning network frames by characteristic pattern extraction convolutional neural networks VGG16, region；Finally, Faster R CNN deep learnings networks are trained stage by stage.Network test results are shown：Faster R CNN networks possess the ability of defect type in efficient identification weld seam D scan images.The method disclosed in the present, take full advantage of advantage of the FasterR CNN identification networks to image recognition, applied in the detection of welding line ultrasonic TOFD D scan images, avoid the influence of testing staff's subjective factor, effectively raise the recognition capability and efficiency to defect type in TOFD D scan images, with the advantages of Detection accuracy is high, robustness is good, strong interference immunity, can be applied in intelligent non-destructive testing technology.
Description
Technical field
The present invention relates to it is a kind of based on deep learning welding line ultrasonic TOFD-D scanning defect type automatic identifying method,
Belong to Ultrasonic Nondestructive and image automatic identification field.
Background technology
Ultrasonic diffraction time difference method（Time of flight Diffraction, TOFD）It is most widely used weld seam imaging
One of detection method.The defects of based on TOFD-D scan images the examined member's experience of type identification, professional knowledge, detection environment
Influence, detection efficiency be present, testing result dispute is big, lacks the problems such as reliability.Therefore, using automatic distinguishing method for image pair
It is significant that D scan images carry out Classifcation of flaws.
The feature combinations conventional machines learning method that generally use is manually extracted realizes the automatic identification of image, manually
Feature selecting and optimization hardly possible are usually faced during feature extraction, theory analysis is difficult, needs the problems such as experience and skill.Faster
R-CNN identification networks have automatically extract feature, generalization ability it is strong and quasi real time property the characteristics of, recognition result is in image
Translation, proportional zoom, inclination or the deformation of other forms of target have the consistency of height, available for easy examined condition
Influence, the automatic identification of baroque ultrasonic TOFD-D scanning defect images.
The present invention devises a kind of automatic identification side of the welding line ultrasonic TOFD-D scanning defect types based on deep learning
Method.Extended by sample, network configuration and training optimization improve the recognition efficiency of network, realize to weld seam ultrasonic TOFD-D
The automatic identification of typical defect type in scan image.
Abbreviation and Key Term definition
TOFD：Time Of Flight Diffraction ultrasonic wave diffraction time difference methods.
PCS：ProbesCenterSpacing probe spacing.
CNN：Convolutional Neural Networks convolutional neural networks.
RCNN：RegionBased Convolutional Neural Networks regions convolutional neural networks.
FRCN：Fast Region based Convolutional Neural Networks fast area nerve nets
Network.
Faster R-CNN：FasterRegion-based Convolutional Neural Network high-speed regions
Neutral net.
RPN：Propose network in RegionProposalNetworks regions.
NMS：Non-maximum Suppression non-maxima suppression algorithms.
The content of the invention
Goal of the invention：The present invention is directed to ultrasonic diffraction time difference method（TOFD）The manual identified of defect type in D scan images
A kind of the problem of reliability is low, dispute is big and efficiency is low, there is provided welding line ultrasonic TOFD-D scanning defects based on deep learning
The automatic identifying method of type, to improve the recognition capability and efficiency of defect type in welding line ultrasonic TOFD-D scan images.
The training process S1 of deep learning network is：Collection weld seam TOFD-D scan images numerical example simultaneously carries out sample expansion
Fill；Faster R-CNN deep learning networks are configured in MATLAB programmed environments；By ImageNet data sets to identifying net
Network carries out pre-training；RPN networks and FRCN networks are respectively trained using D scan images sample；Shared VGG16 networks are to RPN
Network and FRCN networks carry out overall training until network convergence, obtains Faster R-CNN identification network final masks；
Defect recognition process S2 is：The confidence threshold value of defect type is set, higher than the judgement defect of the confidence threshold value；Will
The Faster R-CNN networks that the input training of TOFD-D scan images is completed are tested, and obtain test result.
Training process S1's includes following sub-step：
Step S11, collection weld seam typical defect is tested by ultrasonic TOFD-D Scanning Detctions（Lack of penetration, incomplete fusion, slag inclusion, split
Line, stomata）TOFD-D scan image datas, and by test method expand image pattern, image pattern expand pass through change
Frequency probe, wedge angle, PCS, depth bounds, scanning direction obtain the special with certain scaling and deformation of same weld defect
The D scan images of sign.Training set, checking collection, test set will be divided into by the image expanded.
Step S12, in MATLAB programmed environments, configuration extracts convolutional neural networks VGG16 by characteristic pattern, region is proposed
Network RPN and FRCN network form FasterR-CNN deep learning network frames.
Step S13, using ImageNet data sets to VGG16 networks in the FasterR-CNN network frames that are obtained in S12
Pre-training is carried out, initializes the weights of each layer convolutional neural networks layer in VGG16 networks, obtaining one has extraction characteristic pattern energy
The VGG16 feature extraction networks of power；
Step S14, the TOFD-D scannings image pattern of gained in S11 is sent into VGG16 feature extraction networks, to RPN networks
Be respectively trained and regularized learning algorithm speed with FRCN networks, obtain a RPN network with preliminary aim predictive ability and
The FRCN networks of preliminary classification ability.
Specifically, when RPN networks and FRCN networks are respectively trained, first RPN networks should be trained, then it is right
FRCN networks are trained；During training, the learning rate of the weights and universe network of tackling each network is adjusted.Further
, target should be proposed that frame be configured to the ratio of width to height 1 during RPN network trainings:1 or 1:2 pixel frame, width range set 128 and
256 two kinds of Pixel Dimensions, the i.e. Pixel Dimensions of RPN target areas Suggestion box are：128×128、256×256、128×256、
256 × 512 totally four kinds.During each layer network weight initialization, should directly by the parameter value in VGG16 networks be copied to RPN networks,
The convolution layer parameter that FRCN networks share with VGG16 networks, remaining convolution layer parameter are equal to 0.01 Gauss by standard deviation
Distribution is configured；Learning rate method of adjustment is to tune up learning rate when whole network is poor to checking collection fitting degree,
And then turn learning rate when being fitted preferable down, until network when the error on checking collection reaches certain threshold value i.e. deconditioning.
Step S15, VGG16 feature extractions network is accessed in S14 to the RPN networks and FRCN nets for being respectively trained to obtain simultaneously
Network carries out joint training and use and step S14 identical method regularized learning algorithm speed, utilizes FRCN netinit RPN networks
And shared VGG16 depth convolutional layers, joint tuning is carried out to RPN networks, show that FasterR-CNN identifies network final mask.
So far, a kind of welding line ultrasonic TOFD-D scanning defect types based on deep learning provided by the present invention is automatic
The model training stage of recognition methods completes.
Identification test process S2's includes following sub-step：
Step S21, confidence threshold value is set, the confidence level of output is more than the threshold value and represents to can recognize that the defect；Otherwise can not know
Other defect.Confidence threshold value is arranged to 0.6 in the present invention, i.e., is set to confirm defect when confidence level is higher than 0.6, sets during less than 0.6
It whether there is defect to be uncertain.
Step S22, the FasterR-CNN of completion will be trained in the test set input S15 in the image pattern of gained in S11
Identification network final mask is tested, and is calculated the type of defect and confidence level in image pattern respectively and is obtained test result, lacks
Type identification is fallen into by setting target to propose that frame makes a distinction for different colours, wherein：Slag inclusion-yellow, stomata-green, not
Through welding-white, incomplete fusion-grey, crackle-red.
So far, a kind of welding line ultrasonic TOFD-D scanning defect types based on deep learning provided by the present invention is automatic
The model measurement stage of recognition methods completes.
The present invention constructs the deep learning god applied to welding line ultrasonic TOFD-D scan image defect type automatic identifications
Through network FasterR-CNN networks, extended by sample, the optimization of network training and relevant parameter, using corresponding training method
FasterR-CNN neutral nets are obtained, the network possesses the weld seam typical defect type in identification ultrasonic TOFD-D scan images,
Such as：Crackle, slag inclusion, lack of penetration, incomplete fusion, the ability of stomata, have that strong robustness, discrimination be high, fireballing feature.The party
Method can be widely used in the automatic Classification and Identification of the defects of automatic detection of weld seam type, commented for improving defect synthesis
Valency ability plays an important roll.
Brief description of the drawings
Fig. 1 is that a kind of welding line ultrasonic TOFD-D based on deep learning disclosed in this invention scans the automatic of defect type
The overall flow figure of recognition methods.
Fig. 2 is FasterR-CNN image recognitions network frame schematic diagram described in step S12 of the present invention.
Fig. 3 is VGG16 network structure block diagrams described in step S12 of the present invention.
Fig. 4 is RPN network structure block diagrams described in step S12 of the present invention.
Fig. 5 is FRCN network structure block diagrams described in step S12 of the present invention.
Fig. 6 is prediction block described in step S12 of the present invention with returning frame explanation.
Fig. 7 is type identification test result the defects of TOFD-D scan images described in step S22 of the present invention.
Embodiment
With reference to specific embodiment, the present invention is furture elucidated, it should be understood that these embodiments are merely to illustrate the present invention
Rather than limitation the scope of the present invention, after reading this disclosure, various equivalences of the those skilled in the art to the present invention
The modification of form lags behind the application appended claims limited range.
A kind of as shown in figure 1, welding line ultrasonic TOFD-D scanning defect types based on deep learning disclosed in this invention
Automatic identifying method include the training stage and identification test phase.The groundwork of training stage includes：Weld seam TOFD-D is swept
Retouch image pattern extension collection, FasterR-CNN frameworks are built, network pre-training, RPN networks and FRCN networks are respectively trained,
Network is integrally trained.After the completion of network training, the network is tested, test process is equivalent to the identification in practical application
Journey.Training stage and test phase are described in detail below.
Step S1 is network training process.The training stage that the present invention provides is intended to obtain one can be to welding line ultrasonic
The character network model of classification is identified in defect in TOFD-D scan images.Step S1 network training includes five steps point
It is not：S11, the extension collection of weld seam TOFD-D scan images sample, S12, FasterR-CNN framework are built, and S13, network are instructed in advance
Practice, S14, RPN network and FRCN networks are respectively trained, S15 networks are integrally trained.Details are as follows：
Step S11, the ultrasonic TOFD-D scan images of collection should include weld seam typical defect, such as：Lack of penetration, incomplete fusion, folder
Slag, crackle, gas hole defect totally five class.The extended method of image pattern includes：1）Ultrasonic TOFD detection parameters are adjusted, such as：Probe
Frequency, wedge angle, probe spacing PCS obtain several in same test block and have the defects of different characteristics of image；2）Change
Become the sampling time width of detection signal（Picture traverse）Can obtain has the defects of certain zooming effect in weld penetration scope
Characteristic image；3）In opposite direction, feature identical Defect Scanning image can be obtained by changing scanning direction.The defects of after expanded, schemes
Decent totally 537 defect images；Wherein, training set has 207 images, and checking collection has 129 images, and test set has 201
Image：Training set is used for training network weights；Checking collection is used to verify whether Current Situation of Neural Network model is fitted completely, and with this
Regularized learning algorithm speed（Weighed value adjusting stepping）And deconditioning；Test set is used to test final network performance.Image expansion side
Method not only contributes to expand training set, avoids over-fitting, can also strengthening system noise resisting ability and robustness.
Step S12, Faster R-CNN deep learning network frames are configured in MATLAB programmed environments.Fig. 2 is
The integral frame structure schematic diagram of FasterR-CNN image recognition networks.FasterR-CNN image recognition networks are carried by characteristic pattern
Convolutional neural networks VGG16, region is taken to propose that network RPN and FRCN network are formed, its specific operation workflow is as follows：D is scanned
Image pattern inputs the characteristic pattern for carrying out convolution algorithm to VGG16 networks and exporting every width D scan images, VGG16 nets by input
Network exports primitive image features figure to RPN networks and carries out region proposal, and region proposed issue exports after NMS algorithms reduce redundancy
2000 targets propose frame to FRCN networks, and the target that FRCN networks export according to RPN proposes that frame information is entered to every characteristic pattern
Row down-sampling, output target propose the confidence level of defect type and the positions and dimensions of target proposal frame in frame.VGG16 networks
Structure by the use of VGG16 in the present invention as shown in figure 3, be used as image characteristics extraction instrument.Image pattern is sent into VGG16 by input
Convolutional layer in network carries out convolution algorithm, and view data feature is gradually abstracted with the intensification of convolutional layer after convolution algorithm
Change, by the characteristic pattern of last convolutional layer output image（FeatureMap）, now subject image had on characteristic pattern
Good distinction.VGG16 network structures are described below：VGG16 networks are configured with 16 layers of convolutional neural networks, the network point altogether
For 8 parts, it is respectively：Data input unit；First group of convolutional layer Conv1, this, which is assembled, puts two convolutional layers, and every layer sets shifting
The convolution kernel that dynamic step-length is 1, pixel size is 3x3；Second group of convolutional layer Conv2, this, which is assembled, puts two convolutional layers, every layer of setting
The convolution kernel that moving step length is 1, pixel size is 3x3；3rd group of convolutional layer Conv3, this, which is assembled, puts three convolutional layers, and every layer sets
Put the convolution kernel that moving step length is 1, pixel size is 3x3；4th group of convolutional layer Conv4, this, which is assembled, puts three convolutional layers, every layer
Setting moving step length is 1, the convolution kernel that pixel size is 3x3；5th group of convolutional layer Conv5, this, which is assembled, puts three convolutional layers, often
It is 1 that layer, which sets moving step length, the convolution kernel that pixel size is 3x3；Full articulamentum FC, this, which is assembled, puts three convolutional layers, and first and second
Layer is the full articulamentum for including 4096 neurons, and third layer is the full articulamentum for including 1000 neurons；Classification output layer
Out.The characteristic pattern of last convolutional layer Conv5_3 outputs of VGG16 networks, which is inputted to RPN networks, carries out target prediction and seat
Mark prediction, generates the output that target proposes frame after the completion of prediction.RPN network structures are as shown in figure 4, RPN network settings addition volume
Lamination, target prediction network, coordinate prediction four structures of Recurrent networks and NMS de-redundancy computation layer.The feature exported by VGG16
Figure enters the scanning of line slip convolution to characteristic pattern by adding the 3x3 windows of convolutional layer, obtains 512 dimensional feature Input matrixes to RPN
In target prediction and coordinate the prediction Recurrent networks of network.Target prediction network can typically exportkKind（k=9, including：It is 3 kinds wide
High ratio 1:1、1:2、2:1；The target that 3 kinds of pixel dimensions are 128,256,512 proposes 2 confidence levels that frame is background or defect（0
~1）, 2* altogetherkIndividual prediction output.Coordinate prediction network output is to thiskIndividual target propose frame position and correction size [dx,dy,w,h], whereindx、dyPropose offset of the frame center relative to defect target for target,w、hWhat is represented is that target carries
Discuss the pantograph ratio amendment of frame size（wScaled for width,hHighly to scale）, 4* altogetherkIndividual output.By target prediction network and
Coordinate prediction Recurrent networks can obtain substantial amounts of target and propose frame, but which part target propose the confidence level of frame it is very low and
There is a situation where it is overlapping, therefore RPN networks carry out target propose frame output when also need to by NMS algorithms carry out de-redundancy meter
Calculate, NMS algorithms propose that frame is exported to FRCN networks by 2000 targets before searching for confidence score highest.FRCN networks are set
ROIPooling ponds layer and two full articulamentums are put, as shown in Figure 5.The target obtained through RPN network processes proposes frame input
To the ROIPooling ponds layer of FRCN networks, layer effect in pond is mainly compressed to the characteristic pattern of input, is on the one hand made
Characteristic pattern diminishes, and simplifies network calculations complexity；On the one hand Feature Compression is carried out, extracts principal character.Pond layer is according to region
Propose that information carries out down-sampling to the characteristic pattern of TOFD scan images, the classification output target score after full articulamentum（Certain class lacks
Sunken possibility）With the external surrounding frame amendment of destination object.So far, to the Faster R-CNN depth based on MATLAB programmed environments
The configuration of learning network framework is completed.
Step S13, the purpose of pre-training is mainly that the weights for making each layer convolutional neural networks layer in VGG16 networks have one
Individual initial value, avoid causing model not due to the very few generation over-fitting of sample when being trained using TOFD-D scan images
Convergence.The existing ImageNet data sets for including 14,000,000 width images are directly sent into the Faster obtained by S12 during pre-training
VGG16 convolutional neural networks in R-CNN network frames are trained, and now VGG16 networks will scheme in the ImageNet of input
As data carry out characteristic pattern extraction, characteristic pattern is contrasted with original image after the calculating of network weight algorithm, according to both it
Between each layer neutral net of error transfer factor weights, until characteristic image ties afterwards within the specific limits with primitive image features error
Shu Xunlian.After the completion of training, VGG16 networks have the ability of preliminary extraction characteristics of image.
Step S14, model training is divided into two processes：First process is individually training RPN networks and FRCN networks；The
Two processes are joint training RPN networks and FRCN networks.The purpose individually trained is by VGG16 netinits RPN
Convolution layer parameter in network and FRCN networks, it can effectively prevent whole network from not convergent situation occur when calculating, together
When trained two networks is strengthened respective adaptability, the time of joint training can be reduced.It is first individually during training
The training set in TOFD-D scan images sample in S11 and checking collection are first sent into VGG16 feature extraction networks, while by RPN
Network connection to VGG16 feature extraction networks are trained.Before training RPN networks, it should also adjust target in network and propose frame
Pixel request and framework the ratio of width to height.Picture pixels are higher, and extractible minutia is more, but the amount of calculation faced also will be aobvious
Write increase.Therefore, understand to set target to propose that frame pixel in the direction of the width is up to that 600 can meet that identification will through experiment
Ask.When target proposes framework the ratio of width to height adjustment of frame, Fig. 6 gives the correcting mode that target in the present invention proposes frame, wherein
[dx,dy] represent target prediction frame center with respect to actual defects position offset coordinates, [w, h] what is represented is the target
Propose that frame surrounds the scaling needed for actual defects.Target propose frame by former target frame by center movement [dx,dy] after,
Holding center is constant, and the length and width of target frame scales respectivelyw、hTimes, finally give revised target frame.According to defect
TOFD-D scan image characteristic statisticses display defect aspect ratios are generally 1:1 and 1:2 two kinds, therefore setting target to propose
The ratio of width to height is also set as 1 during frame:1 and 1:2 two kinds.It is up to 600 in view of pixel on width, the present invention is provided with
128 and 256 two kind of Pixel Dimensions, thus final goal propose that the size of frame is：128×128、256×256、128×256、
256 × 512 totally four kinds, i.e., nowk=4.It should be noted that：The 1 of target prediction network output:1 and 2:1 two kinds of frameworks, are being passed through
After coordinate prediction Recurrent networks adjustment position, pantograph ratio, final framework the ratio of width to height is not limited to both ratios, can be any
Ratio.When training RPN networks, the shared convolution layer parameter with VGG16 networks（Network weight）It can be passed through with direct copying
ImageNet trains the parameter in obtained VGG16 networks；Remaining convolution layer parameter Gauss point of the standard deviation equal to 0.01
Cloth carries out Initialize installation, a new RPN network is obtained after the completion of training, now the network has preliminary target frame（Mesh
Mark region）Predictive ability.Then, classification based training then by FRCN network connections to VGG16 feature extractions network is carried out, equally will altogether
Parameter value in some convolution layer parameter direct copying VGG16 networks, remaining convolutional layer are initialized by Gaussian Profile,
Training is completed to obtain a new FRCN network, and now the network has preliminary classification capacity.It is worth noting that, carry out
Also need to be adjusted network overall learning rate during training.Learning rate is a most important ginseng for training pattern
Number, when learning rate is too big, easily there is over control in overall network, i.e., is constantly dissipated at extreme point both ends, or acutely shake
Swing, now with the increase of iterations, Softmax Loss（Detect class probability）With Smooth L1 Loss（Detect frame
Return）Layer is without the trend reduced；It is too small, it can lead to not be quickly found out the preferable descent direction of Loss layers, with iteration time
Number increase Loss layers are basically unchanged.It should be deferred to during regularized learning algorithm speed：Adjusted when whole network is poor to checking collection fitting degree
Big learning rate, and learning rate is then turned down when being fitted preferable, until network is when the error on checking collection reaches certain threshold value
That is deconditioning.For example, initial learning rate lr is arranged to 0.1, using checking collection come determine when to reduce learning rate and
When deconditioning, when checking collect meet with bottleneck, lr is now adjusted to 0.05 or 0.02, then proceedes to train.Final lr meetings
Become a very small value, now can deconditioning.
Step S15, during joint training, obtained RPN networks and FRCN networks will individually be trained while accesses VGG16 features
Network is extracted to carry out.The parameter sharing of RPN networks and FRCN networks and unrealized convolutional network layer during due to individually training, therefore
FRCN netinit RPN networks and shared VGG16 depth convolutional layers should be utilized in joint training, RPN networks are joined
Close tuning so that two networks form a joint network.The identical convolutional layer parameter of each network is completely shared in joint network,
Possesses the Classification and Identification ability of defect in TOFD-D scan images.Also need to adjust e-learning speed during joint training
Whole, its method of adjustment is with reference to step S14.So far, a final Faster R-CNN identification network model is obtained.
Step S2 is Network Recognition test process, and Model Identification test is by by ultrasonic TOFD-D scan image samples
Test set sample is inputted in the FasterR-CNN identification network final masks completed to training, so as to reach test model identification
The purpose of ability.Trained FasterR-CNN identifications network model can export the confidence level of defect type, so-called confidence
Degree is the possibility of certain class defect（0~1）, defect is judged by defect confidence level.Faster R-CNN algorithms are by ultrasound
The position that defect is there may be in TOFD-D scannings imaging is marked with rectangle frame, and identifies the type of the defect.Described in step S2
Model Identification test phase include following sub-step：
S21, the Faster R-CNN knowledges for finding, being obtained when confidence threshold value is arranged to 0.6 in S15 are contrasted by test of many times
Other network final mask has the defects of optimal discrimination.Therefore, it is 0.6 that the present invention, which sets the confidence level of defect, and score is higher than
It is set to confirm defect when 0.6, score is set to zero defect when being less than 0.6.
S22, TOFD-D is swept to the Faster R-CNN identification nets obtained in scan image sample set in test set input S15
Network final mask, the stability of model is verified with this.Suggestion box the defects of variety classes is arranged to different face by programming
Color：Slag inclusion-yellow, stomata-green, lack of penetration-white, incomplete fusion-grey, crackle-red；The upper left corner of Suggestion box
It also show defect type and confidence level（0~1）, Fig. 7 is recognition result figure.
The recognition effect statistical form of the test set of table 1
| Defect type | It is lack of penetration | Incomplete fusion | Slag inclusion | Crackle | Stomata |
| Participate in the sample size of test | 16 | 10 | 47 | 80 | 48 |
| It is non-defective to be known by mistake | 2 | 0 | 8 | 11 | 4 |
| Defect type is judged by accident | 0 | 0 | 0 | 0 | 1 (by crackle is known into by mistake) |
| Recognition accuracy | 0.87 | 1.00 | 0.85 | 0.86 | 0.90 |
So far, the automatic identification of a kind of welding line ultrasonic TOFD-D scanning defect types based on deep learning provided by the present invention
The training stage of method and test phase are completed.
Method disclosed by the invention is tested in ultrasonic TOFD-D scanning images, and test result is as shown in table 2.Table
In as can be seen that participate in testing the defects of in image, be up to more than 0.90 to the recognition accuracy of incomplete fusion and stomata,
To crackle, slag inclusion and lack of penetration recognition accuracy also up to more than 0.85.It can be seen that ultrasound provided by the invention
The accuracy rate of TOFD-D scan image recognition methods is in higher level.
In summary, the invention discloses oneself of a kind of welding line ultrasonic TOFD-D scanning defect types based on deep learning
Dynamic recognition methods.Training and the method for testing of related FasterR-CNN networks are mainly elaborated, including：1）Sample collection is with expanding
Exhibition method, 2）FasterR-CNN network frames are built, 3）Network pre-training method, 3）FasterR-CNN networks have been carried out continuously
Training so that model obtains optimal learning rate and each layer network weights, last, by test model recognition effect, obtains
Ultrasonic TOFD-D scan image automatic identification the networks of the very high weld defect of one accuracy rate.The method disclosed in the present,
Advantages of the FasterR-CNN to image recognition is taken full advantage of, is applied in the detection of ultrasonic TOFD-D scan images, had
The advantages of designing preferable simple, Detection accuracy height, robustness and strong antijamming capability, the detection for improving TOFD-D scannings is imitated
Rate.
Claims (10)
1. the automatic identifying method of the welding line ultrasonic TOFD-D scanning defect types based on deep learning, methods described include model
Training process and identification process are as follows：
The training process S1 of deep learning network is：Collection weld seam TOFD-D scan images numerical example simultaneously carries out sample expansion；
Faster R-CNN deep learning networks are configured in MATLAB programmed environments；Identification network is carried out by ImageNet data sets
Pre-training；RPN networks and FRCN networks are respectively trained using D scan images sample；Shared VGG16 networks to RPN networks and
FRCN networks carry out overall training until network convergence, obtains Faster R-CNN identification network final masks；
Defect recognition process S2 is：The confidence threshold value of defect is set, and confidence level is defined as such defect the defects of higher than threshold value；
The Faster R-CNN networks that TOFD-D scan images are inputted to training completion are tested, and obtain test result.
2. the automatic identification side of the welding line ultrasonic TOFD-D scanning defect types based on deep learning according to claim 1
Method, it is characterised in that as follows the step of training process S1：
S11, the TOFD-D scan image datas by ultrasonic TOFD-D scanning techniques collection weld seam typical defect, and pass through experiment
Method expands image pattern, and image pattern is divided into training set, checking collection, test set；
S12, build in MATLAB programmed environments by characteristic pattern extraction convolutional neural networks VGG16, region propose network RPN with
And FRCN networks form FasterR-CNN deep learning network frames；
S13, using ImageNet data sets VGG16 networks in the FasterR-CNN network frames that are obtained in S12 are instructed in advance
Practice, initialize the weights of each layer convolutional neural networks layer in VGG16 networks, obtaining one has extraction characteristics of image figure ability
VGG16 feature extraction networks；
S14, the TOFD-D scannings image pattern of gained in S11 inputted into VGG16 feature extraction networks, to RPN networks and
FRCN networks be respectively trained and regularized learning algorithm speed, obtain a RPN network with preliminary aim predictive ability and just
Walk the FRCN networks of classification capacity；
S15, the output of VGG16 feature extraction networks is accessed in S14 to the RPN networks and FRCN networks for being respectively trained to obtain simultaneously
Progress joint training and use and step S14 identical method regularized learning algorithm speed, using FRCN netinit RPN networks simultaneously
Shared VGG16 depth convolutional layers, joint tuning is carried out to RPN network weights, show that FasterR-CNN identifies the final mould of network
Type.
3. the automatic identification side of the welding line ultrasonic TOFD-D scanning defect types based on deep learning according to claim 1
Method, it is characterised in that as follows the step of test process S2：
S21, setting confidence threshold value, the confidence level of output are more than the threshold value and represent to can recognize that the defect；Otherwise None- identified lacks
Fall into；
S22, in the test set input S15 in the image pattern of gained in S11 the FasterR-CNN of completion will be trained to identify network
Final mask is tested, and obtains the automatic identification effect of defect type.
4. weld seam typical defect type according to claim 2, it is characterised in that weld seam typical defect type is in S11：Not
Through welding, incomplete fusion, slag inclusion, crackle, stomata.
5. image pattern extending method according to claim 2, it is characterised in that image pattern expands and passes through change in S11
Frequency probe, wedge angle, PCS, depth bounds, scanning direction obtain the special with certain scaling and deformation of same weld defect
The D scan images of sign.
6. method is respectively trained according to claim 2, it is characterised in that be respectively trained in S14 first to be carried out to RPN networks
Training, then FRCN networks are trained；During training, the weights of each network and the learning rate of universe network should be adjusted
It is whole.
7. RPN network training methods according to claim 6, it is characterised in that target should be proposed frame during RPN network trainings
It is configured to the ratio of width to height 1:1 or 1:2 pixel frame, width range set 128 and 256 two kind of Pixel Dimensions, i.e. RPN target areas build
View frame Pixel Dimensions be：128 × 128,256 × 256,128 × 256,256 × 512 totally four kinds.
8. according to network weight described in right 6 and learning rate method of adjustment, it is characterised in that each layer network weight initialization
When, the parameter value in VGG16 networks directly should be copied to RPN networks, FRCN networks and the convolutional layer that VGG16 networks share and joined
Number, remaining convolution layer parameter are configured by Gaussian Profile of the standard deviation equal to 0.01；Learning rate method of adjustment is, when
Learning rate is then turned down when tuning up learning rate when whole network is to verifying that collection fitting degree is poor, and being fitted preferable, until net
Network i.e. deconditioning when the error on checking collection reaches certain threshold value.
9. confidence threshold value method to set up according to claim 3, it is characterised in that confidence threshold value is arranged in S21
0.6, i.e.,：Confidence level is set to confirm defect when being higher than 0.6, is set to uncertain during less than 0.6 and whether there is defect.
10. defect type method of discrimination according to claim 3, it is characterised in that defect type, which differentiates, in S22 passes through setting
Target proposes that frame makes a distinction for different colours, wherein：Slag inclusion-yellow, stomata-green, lack of penetration-white, incomplete fusion-
Grey, crackle-red.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710642922.6A CN107451997A (en) | 2017-07-31 | 2017-07-31 | A kind of automatic identifying method of the welding line ultrasonic TOFD D scanning defect types based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710642922.6A CN107451997A (en) | 2017-07-31 | 2017-07-31 | A kind of automatic identifying method of the welding line ultrasonic TOFD D scanning defect types based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107451997A true CN107451997A (en) | 2017-12-08 |
Family
ID=60489995
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710642922.6A Pending CN107451997A (en) | 2017-07-31 | 2017-07-31 | A kind of automatic identifying method of the welding line ultrasonic TOFD D scanning defect types based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107451997A (en) |
Cited By (52)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108154508A (en) * | 2018-01-09 | 2018-06-12 | 北京百度网讯科技有限公司 | Method, apparatus, storage medium and the terminal device of product defects detection positioning |
| CN108182402A (en) * | 2017-12-27 | 2018-06-19 | 同济大学 | A kind of steel box-girder method for inspecting for using acquisition robot and deep learning algorithm |
| CN108428247A (en) * | 2018-02-27 | 2018-08-21 | 广州视源电子科技股份有限公司 | Method and system for detecting direction of soldering tin point |
| CN108509919A (en) * | 2018-04-03 | 2018-09-07 | 哈尔滨哈船智控科技有限责任公司 | A kind of detection and recognition methods based on deep learning to waterline in video or picture |
| CN108932713A (en) * | 2018-07-20 | 2018-12-04 | 成都指码科技有限公司 | A kind of weld porosity defect automatic testing method based on deep learning |
| CN109035204A (en) * | 2018-06-25 | 2018-12-18 | 华南理工大学 | A kind of weld seam target real-time detection method |
| CN109064459A (en) * | 2018-07-27 | 2018-12-21 | 江苏理工学院 | A kind of Fabric Defect detection method based on deep learning |
| CN109087294A (en) * | 2018-07-31 | 2018-12-25 | 深圳辰视智能科技有限公司 | A kind of product defects detection method, system and computer readable storage medium |
| CN109115805A (en) * | 2018-10-25 | 2019-01-01 | 广东华中科技大学工业技术研究院 | Transparent component defect detecting device and method based on ultrasound and the double imagings of optics |
| CN109285139A (en) * | 2018-07-23 | 2019-01-29 | 同济大学 | A kind of x-ray imaging weld inspection method based on deep learning |
| CN109615611A (en) * | 2018-11-19 | 2019-04-12 | 国家电网有限公司 | A kind of insulator self-destruction defect inspection method based on inspection image |
| CN109767427A (en) * | 2018-12-25 | 2019-05-17 | 北京交通大学 | The detection method of train rail fastener defect |
| CN109900706A (en) * | 2019-03-20 | 2019-06-18 | 易思维（杭州）科技有限公司 | A kind of weld seam and weld defect detection method based on deep learning |
| CN109946303A (en) * | 2017-12-21 | 2019-06-28 | 东友精细化工有限公司 | Check device and method |
| CN109977948A (en) * | 2019-03-20 | 2019-07-05 | 哈尔滨工业大学 | A kind of stirring friction welding seam defect identification method based on convolutional neural networks |
| CN110110798A (en) * | 2019-05-13 | 2019-08-09 | 广东工业大学 | A kind of weld joint recognition method based on Mask-RCNN network |
| CN110120035A (en) * | 2019-04-17 | 2019-08-13 | 杭州数据点金科技有限公司 | A kind of tire X-ray defect detection method differentiating defect grade |
| CN110135513A (en) * | 2019-05-22 | 2019-08-16 | 广东工业大学 | A kind of weld joint recognition method of the welding robot based on deep learning |
| CN110175985A (en) * | 2019-04-22 | 2019-08-27 | 国网江苏省电力有限公司电力科学研究院 | Carbon fiber composite core wire damage detecting method, device and computer storage medium |
| CN110243934A (en) * | 2019-05-30 | 2019-09-17 | 中国计量大学上虞高等研究院有限公司 | A kind of ultrasonic weld seam detection method based on wavelet convolution neural network |
| CN110276403A (en) * | 2019-06-25 | 2019-09-24 | 北京百度网讯科技有限公司 | Method for establishing model and device |
| CN110297041A (en) * | 2019-07-01 | 2019-10-01 | 天津工业大学 | A kind of 3D woven composite defect inspection method based on FCN and GRU |
| CN110322430A (en) * | 2019-05-17 | 2019-10-11 | 杭州数据点金科技有限公司 | A kind of two stage tire X-ray defect detection method |
| CN110414299A (en) * | 2018-04-28 | 2019-11-05 | 中山大学 | A kind of monkey face Genetic relationship method based on computer vision |
| CN110458828A (en) * | 2019-08-12 | 2019-11-15 | 广东工业大学 | A kind of laser welding defect identification method and device based on multi-modal fusion network |
| CN110470729A (en) * | 2019-07-31 | 2019-11-19 | 浙江树人学院（浙江树人大学） | A kind of sleeve in oil field defect lossless detection method based on current vortex |
| CN110490166A (en) * | 2019-08-26 | 2019-11-22 | 石家庄铁道大学 | High iron catenary pipe cap automatic testing method based on depth learning technology |
| CN110728657A (en) * | 2019-09-10 | 2020-01-24 | 江苏理工学院 | Annular bearing outer surface defect detection method based on deep learning |
| CN110910373A (en) * | 2019-11-25 | 2020-03-24 | 西南交通大学 | Identification method of orthotropic steel bridge deck fatigue crack detection image |
| CN110988140A (en) * | 2019-11-25 | 2020-04-10 | 西南交通大学 | Intelligent identification method for fatigue cracks of orthotropic steel bridge deck slab |
| CN111257422A (en) * | 2020-02-28 | 2020-06-09 | 北京新联铁集团股份有限公司 | Wheel axle defect identification model construction method and defect identification method based on machine vision |
| CN111311542A (en) * | 2020-01-15 | 2020-06-19 | 歌尔股份有限公司 | Product quality detection method and device |
| CN111462193A (en) * | 2019-01-22 | 2020-07-28 | 斯特拉德视觉公司 | Object detection method and device for monitoring multiple cameras or surrounding views |
| CN111461120A (en) * | 2020-04-01 | 2020-07-28 | 济南浪潮高新科技投资发展有限公司 | Method for detecting surface defects of convolutional neural network object based on region |
| CN111612730A (en) * | 2019-02-23 | 2020-09-01 | 妙微(杭州)科技有限公司 | Welding seam detection and positioning method |
| US10769774B2 (en) | 2018-01-09 | 2020-09-08 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method and device for detecting a defect in a steel plate, as well as apparatus and server therefor |
| CN111882557A (en) * | 2020-09-28 | 2020-11-03 | 成都睿沿科技有限公司 | Welding defect detection method and device, electronic equipment and storage medium |
| CN111912910A (en) * | 2020-08-12 | 2020-11-10 | 上海核工程研究设计院有限公司 | Intelligent identification method for polyethylene pipeline hot-melt weld joint hybrid ultrasonic scanning defects |
| CN112198232A (en) * | 2020-09-14 | 2021-01-08 | 昆明理工大学 | Drainage pipeline working condition detection and identification method |
| CN112215843A (en) * | 2019-12-31 | 2021-01-12 | 无锡祥生医疗科技股份有限公司 | Ultrasonic intelligent imaging navigation method and device, ultrasonic equipment and storage medium |
| CN112232400A (en) * | 2020-10-12 | 2021-01-15 | 太原科技大学 | Stainless steel weld ultrasonic defect detection method based on depth feature fusion |
| CN112329687A (en) * | 2020-11-16 | 2021-02-05 | 山东天岳先进科技股份有限公司 | Automatic recognition method and system for dislocation of silicon carbide substrate |
| CN112651341A (en) * | 2020-12-28 | 2021-04-13 | 长江大学 | Processing method of welded pipe weld joint real-time detection video |
| CN112816557A (en) * | 2019-11-18 | 2021-05-18 | 中国商用飞机有限责任公司 | Defect detection method, device, equipment and storage medium |
| CN112949666A (en) * | 2019-11-26 | 2021-06-11 | 北矿机电科技有限责任公司 | Image identification method for fault cathode plate of zinc stripping machine |
| CN113034446A (en) * | 2021-03-08 | 2021-06-25 | 国网山东省电力公司平邑县供电公司 | Automatic transformer substation equipment defect identification method and system |
| CN113207286A (en) * | 2018-12-19 | 2021-08-03 | 松下知识产权经营株式会社 | Appearance inspection device, method for improving accuracy of determination of presence or absence of shape defect in welded part and type of shape defect using the appearance inspection device, welding system, and method for welding workpiece using the welding system |
| CN113256566A (en) * | 2021-04-29 | 2021-08-13 | 广州杰赛科技股份有限公司 | Pipeline weld defect identification method |
| CN113808065A (en) * | 2020-05-28 | 2021-12-17 | 青岛海尔工业智能研究院有限公司 | Weld joint detection method and device, industrial robot and storage medium |
| CN114359193A (en) * | 2021-12-23 | 2022-04-15 | 华中科技大学 | Defect classification method and system based on ultrasonic phased array imaging |
| CN114820499A (en) * | 2022-04-20 | 2022-07-29 | 上海电气集团股份有限公司 | Pipeline welding seam detection method |
| CN117740960A (en) * | 2023-12-04 | 2024-03-22 | 绍兴市特种设备检测院 | TOFD spectrum feature detection method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105891215A (en) * | 2016-03-31 | 2016-08-24 | 浙江工业大学 | Welding visual detection method and device based on convolutional neural network |
| US20170061625A1 (en) * | 2015-08-26 | 2017-03-02 | Digitalglobe, Inc. | Synthesizing training data for broad area geospatial object detection |
| CN106504233A (en) * | 2016-10-18 | 2017-03-15 | 国网山东省电力公司电力科学研究院 | Image electric power widget recognition methodss and system are patrolled and examined based on the unmanned plane of Faster R CNN |
-
2017
- 2017-07-31 CN CN201710642922.6A patent/CN107451997A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170061625A1 (en) * | 2015-08-26 | 2017-03-02 | Digitalglobe, Inc. | Synthesizing training data for broad area geospatial object detection |
| CN105891215A (en) * | 2016-03-31 | 2016-08-24 | 浙江工业大学 | Welding visual detection method and device based on convolutional neural network |
| CN106504233A (en) * | 2016-10-18 | 2017-03-15 | 国网山东省电力公司电力科学研究院 | Image electric power widget recognition methodss and system are patrolled and examined based on the unmanned plane of Faster R CNN |
Non-Patent Citations (1)
| Title |
|---|
| 钟晓明，余贵珍，马亚龙，任毅龙: "《基于快速区域卷积神经网络的交通标志识别算法研究》", 《2 0 1 6 中国汽车工程学会年会论文集》 * |
Cited By (66)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109946303A (en) * | 2017-12-21 | 2019-06-28 | 东友精细化工有限公司 | Check device and method |
| CN108182402A (en) * | 2017-12-27 | 2018-06-19 | 同济大学 | A kind of steel box-girder method for inspecting for using acquisition robot and deep learning algorithm |
| US10769774B2 (en) | 2018-01-09 | 2020-09-08 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method and device for detecting a defect in a steel plate, as well as apparatus and server therefor |
| CN108154508A (en) * | 2018-01-09 | 2018-06-12 | 北京百度网讯科技有限公司 | Method, apparatus, storage medium and the terminal device of product defects detection positioning |
| CN108428247A (en) * | 2018-02-27 | 2018-08-21 | 广州视源电子科技股份有限公司 | Method and system for detecting direction of soldering tin point |
| CN108509919A (en) * | 2018-04-03 | 2018-09-07 | 哈尔滨哈船智控科技有限责任公司 | A kind of detection and recognition methods based on deep learning to waterline in video or picture |
| CN108509919B (en) * | 2018-04-03 | 2022-04-29 | 哈尔滨哈船智控科技有限责任公司 | Method for detecting and identifying waterline in video or picture based on deep learning |
| CN110414299A (en) * | 2018-04-28 | 2019-11-05 | 中山大学 | A kind of monkey face Genetic relationship method based on computer vision |
| CN110414299B (en) * | 2018-04-28 | 2024-02-06 | 中山大学 | Monkey face affinity analysis method based on computer vision |
| CN109035204A (en) * | 2018-06-25 | 2018-12-18 | 华南理工大学 | A kind of weld seam target real-time detection method |
| CN109035204B (en) * | 2018-06-25 | 2021-06-08 | 华南理工大学 | Real-time detection method for weld joint target |
| CN108932713A (en) * | 2018-07-20 | 2018-12-04 | 成都指码科技有限公司 | A kind of weld porosity defect automatic testing method based on deep learning |
| CN109285139A (en) * | 2018-07-23 | 2019-01-29 | 同济大学 | A kind of x-ray imaging weld inspection method based on deep learning |
| CN109064459A (en) * | 2018-07-27 | 2018-12-21 | 江苏理工学院 | A kind of Fabric Defect detection method based on deep learning |
| CN109087294A (en) * | 2018-07-31 | 2018-12-25 | 深圳辰视智能科技有限公司 | A kind of product defects detection method, system and computer readable storage medium |
| CN109115805A (en) * | 2018-10-25 | 2019-01-01 | 广东华中科技大学工业技术研究院 | Transparent component defect detecting device and method based on ultrasound and the double imagings of optics |
| CN109615611A (en) * | 2018-11-19 | 2019-04-12 | 国家电网有限公司 | A kind of insulator self-destruction defect inspection method based on inspection image |
| CN109615611B (en) * | 2018-11-19 | 2023-06-27 | 国家电网有限公司 | Inspection image-based insulator self-explosion defect detection method |
| CN113207286A (en) * | 2018-12-19 | 2021-08-03 | 松下知识产权经营株式会社 | Appearance inspection device, method for improving accuracy of determination of presence or absence of shape defect in welded part and type of shape defect using the appearance inspection device, welding system, and method for welding workpiece using the welding system |
| CN109767427A (en) * | 2018-12-25 | 2019-05-17 | 北京交通大学 | The detection method of train rail fastener defect |
| CN111462193B (en) * | 2019-01-22 | 2023-09-12 | 斯特拉德视觉公司 | Object detection method and device for multi-camera or surround view monitoring |
| CN111462193A (en) * | 2019-01-22 | 2020-07-28 | 斯特拉德视觉公司 | Object detection method and device for monitoring multiple cameras or surrounding views |
| CN111612730A (en) * | 2019-02-23 | 2020-09-01 | 妙微(杭州)科技有限公司 | Welding seam detection and positioning method |
| CN111612730B (en) * | 2019-02-23 | 2022-04-22 | 妙微(杭州)科技有限公司 | Welding seam detection and positioning method |
| CN109900706A (en) * | 2019-03-20 | 2019-06-18 | 易思维（杭州）科技有限公司 | A kind of weld seam and weld defect detection method based on deep learning |
| CN109977948A (en) * | 2019-03-20 | 2019-07-05 | 哈尔滨工业大学 | A kind of stirring friction welding seam defect identification method based on convolutional neural networks |
| CN110120035A (en) * | 2019-04-17 | 2019-08-13 | 杭州数据点金科技有限公司 | A kind of tire X-ray defect detection method differentiating defect grade |
| CN110175985A (en) * | 2019-04-22 | 2019-08-27 | 国网江苏省电力有限公司电力科学研究院 | Carbon fiber composite core wire damage detecting method, device and computer storage medium |
| CN110110798A (en) * | 2019-05-13 | 2019-08-09 | 广东工业大学 | A kind of weld joint recognition method based on Mask-RCNN network |
| CN110110798B (en) * | 2019-05-13 | 2023-02-10 | 广东工业大学 | Mask-RCNN (Richardson network) -based weld joint identification method, device, equipment and readable storage medium |
| CN110322430A (en) * | 2019-05-17 | 2019-10-11 | 杭州数据点金科技有限公司 | A kind of two stage tire X-ray defect detection method |
| CN110135513A (en) * | 2019-05-22 | 2019-08-16 | 广东工业大学 | A kind of weld joint recognition method of the welding robot based on deep learning |
| CN110243934A (en) * | 2019-05-30 | 2019-09-17 | 中国计量大学上虞高等研究院有限公司 | A kind of ultrasonic weld seam detection method based on wavelet convolution neural network |
| CN110276403A (en) * | 2019-06-25 | 2019-09-24 | 北京百度网讯科技有限公司 | Method for establishing model and device |
| CN110297041A (en) * | 2019-07-01 | 2019-10-01 | 天津工业大学 | A kind of 3D woven composite defect inspection method based on FCN and GRU |
| CN110470729B (en) * | 2019-07-31 | 2022-12-02 | 浙江树人学院（浙江树人大学） | Eddy current-based nondestructive testing method for oil field casing pipe defects |
| CN110470729A (en) * | 2019-07-31 | 2019-11-19 | 浙江树人学院（浙江树人大学） | A kind of sleeve in oil field defect lossless detection method based on current vortex |
| CN110458828A (en) * | 2019-08-12 | 2019-11-15 | 广东工业大学 | A kind of laser welding defect identification method and device based on multi-modal fusion network |
| CN110490166A (en) * | 2019-08-26 | 2019-11-22 | 石家庄铁道大学 | High iron catenary pipe cap automatic testing method based on depth learning technology |
| CN110728657A (en) * | 2019-09-10 | 2020-01-24 | 江苏理工学院 | Annular bearing outer surface defect detection method based on deep learning |
| CN112816557A (en) * | 2019-11-18 | 2021-05-18 | 中国商用飞机有限责任公司 | Defect detection method, device, equipment and storage medium |
| CN110988140A (en) * | 2019-11-25 | 2020-04-10 | 西南交通大学 | Intelligent identification method for fatigue cracks of orthotropic steel bridge deck slab |
| CN110910373A (en) * | 2019-11-25 | 2020-03-24 | 西南交通大学 | Identification method of orthotropic steel bridge deck fatigue crack detection image |
| CN110988140B (en) * | 2019-11-25 | 2021-10-01 | 西南交通大学 | Intelligent identification method for fatigue cracks of orthotropic steel bridge deck slab |
| CN112949666B (en) * | 2019-11-26 | 2024-04-26 | 北矿机电科技有限责任公司 | Image recognition method for fault cathode plate of zinc stripping machine |
| CN112949666A (en) * | 2019-11-26 | 2021-06-11 | 北矿机电科技有限责任公司 | Image identification method for fault cathode plate of zinc stripping machine |
| CN112215843A (en) * | 2019-12-31 | 2021-01-12 | 无锡祥生医疗科技股份有限公司 | Ultrasonic intelligent imaging navigation method and device, ultrasonic equipment and storage medium |
| CN112215843B (en) * | 2019-12-31 | 2021-06-11 | 无锡祥生医疗科技股份有限公司 | Ultrasonic intelligent imaging navigation method and device, ultrasonic equipment and storage medium |
| CN111311542A (en) * | 2020-01-15 | 2020-06-19 | 歌尔股份有限公司 | Product quality detection method and device |
| CN111311542B (en) * | 2020-01-15 | 2023-09-19 | 歌尔股份有限公司 | Product quality detection method and device |
| CN111257422A (en) * | 2020-02-28 | 2020-06-09 | 北京新联铁集团股份有限公司 | Wheel axle defect identification model construction method and defect identification method based on machine vision |
| CN111257422B (en) * | 2020-02-28 | 2023-09-08 | 北京新联铁集团股份有限公司 | Wheel axle defect identification model construction method and defect identification method based on machine vision |
| CN111461120A (en) * | 2020-04-01 | 2020-07-28 | 济南浪潮高新科技投资发展有限公司 | Method for detecting surface defects of convolutional neural network object based on region |
| CN113808065A (en) * | 2020-05-28 | 2021-12-17 | 青岛海尔工业智能研究院有限公司 | Weld joint detection method and device, industrial robot and storage medium |
| CN111912910A (en) * | 2020-08-12 | 2020-11-10 | 上海核工程研究设计院有限公司 | Intelligent identification method for polyethylene pipeline hot-melt weld joint hybrid ultrasonic scanning defects |
| CN112198232A (en) * | 2020-09-14 | 2021-01-08 | 昆明理工大学 | Drainage pipeline working condition detection and identification method |
| CN111882557A (en) * | 2020-09-28 | 2020-11-03 | 成都睿沿科技有限公司 | Welding defect detection method and device, electronic equipment and storage medium |
| CN112232400A (en) * | 2020-10-12 | 2021-01-15 | 太原科技大学 | Stainless steel weld ultrasonic defect detection method based on depth feature fusion |
| CN112232400B (en) * | 2020-10-12 | 2023-06-20 | 太原科技大学 | Stainless steel weld ultrasonic defect detection method based on depth feature fusion |
| CN112329687A (en) * | 2020-11-16 | 2021-02-05 | 山东天岳先进科技股份有限公司 | Automatic recognition method and system for dislocation of silicon carbide substrate |
| CN112651341A (en) * | 2020-12-28 | 2021-04-13 | 长江大学 | Processing method of welded pipe weld joint real-time detection video |
| CN113034446A (en) * | 2021-03-08 | 2021-06-25 | 国网山东省电力公司平邑县供电公司 | Automatic transformer substation equipment defect identification method and system |
| CN113256566A (en) * | 2021-04-29 | 2021-08-13 | 广州杰赛科技股份有限公司 | Pipeline weld defect identification method |
| CN114359193A (en) * | 2021-12-23 | 2022-04-15 | 华中科技大学 | Defect classification method and system based on ultrasonic phased array imaging |
| CN114820499A (en) * | 2022-04-20 | 2022-07-29 | 上海电气集团股份有限公司 | Pipeline welding seam detection method |
| CN117740960A (en) * | 2023-12-04 | 2024-03-22 | 绍兴市特种设备检测院 | TOFD spectrum feature detection method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107451997A (en) | A kind of automatic identifying method of the welding line ultrasonic TOFD D scanning defect types based on deep learning | |
| CN109615611A (en) | A kind of insulator self-destruction defect inspection method based on inspection image | |
| CN109285139A (en) | A kind of x-ray imaging weld inspection method based on deep learning | |
| CN110047073A (en) | A kind of X-ray weld image fault grading method and system | |
| CN109816012A (en) | A kind of multiscale target detection method of integrating context information | |
| CN109446992A (en) | Remote sensing image building extracting method and system, storage medium, electronic equipment based on deep learning | |
| CN105069746A (en) | Video real-time human face substitution method and system based on partial affine and color transfer technology | |
| CN110705457A (en) | Remote sensing image building change detection method | |
| CN106127204A (en) | A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks | |
| CN114548278A (en) | In-service tunnel lining structure defect identification method and system based on deep learning | |
| CN110287932A (en) | Route denial information extraction based on the segmentation of deep learning image, semantic | |
| CN106530284A (en) | Welding spot type detection and device based on image recognition | |
| CN111275171B (en) | Small target detection method based on parameter sharing multi-scale super-division reconstruction | |
| CN108573222A (en) | The pedestrian image occlusion detection method for generating network is fought based on cycle | |
| CN109635748A (en) | The extracting method of roadway characteristic in high resolution image | |
| CN114296075A (en) | Ground penetrating radar image artificial intelligence identification method and device | |
| CN112241950A (en) | Detection method of tower crane crack image | |
| CN115345881B (en) | Pavement disease detection method based on computer vision | |
| CN113096138A (en) | Weak supervision semantic image segmentation method for selective pixel affinity learning | |
| CN114170527B (en) | Remote sensing target detection method using rotating frame representation | |
| CN107993228A (en) | A kind of vulnerable plaque automatic testing method and device based on cardiovascular OCT images | |
| CN114926511A (en) | High-resolution remote sensing image change detection method based on self-supervision learning | |
| CN110532914A (en) | Building analyte detection method based on fine-feature study | |
| CN109508661A (en) | A kind of person's of raising one's hand detection method based on object detection and Attitude estimation | |
| CN114897865A (en) | Industrial defect detection method based on small number of defect samples |
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: 20171208 |
|
| RJ01 | Rejection of invention patent application after publication |