CN110243934A - A kind of ultrasonic weld seam detection method based on wavelet convolution neural network - Google Patents
A kind of ultrasonic weld seam detection method based on wavelet convolution neural network Download PDFInfo
- Publication number
- CN110243934A CN110243934A CN201910461584.5A CN201910461584A CN110243934A CN 110243934 A CN110243934 A CN 110243934A CN 201910461584 A CN201910461584 A CN 201910461584A CN 110243934 A CN110243934 A CN 110243934A
- Authority
- CN
- China
- Prior art keywords
- layer
- data
- convolutional layer
- neural network
- weld
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- 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
-
- 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
- G01N29/44—Processing the detected response signal, e.g. electronic circuits specially adapted therefor
- G01N29/4481—Neural networks
-
- 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
- G01N2291/2675—Seam, butt welding
Landscapes
- Physics & Mathematics (AREA)
- Analytical Chemistry (AREA)
- Biochemistry (AREA)
- Pathology (AREA)
- Immunology (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Chemical & Material Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Signal Processing (AREA)
- Acoustics & Sound (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of ultrasonic weld seam detection methods based on wavelet convolution neural network, comprising the following steps: step 1: obtaining weld data；Step 2: weld data pretreatment；Step 3: training wavelet convolution neural network；Step 4: weld data to be detected is obtained；Step 5: weld data pretreatment to be detected；Step 6: testing result is calculated；Step 7；Judge whether weldment to be detected is qualified.Cannot accurately it classify when the present invention is for the weld seam detection of different defect types in the prior art, influence of the different scale to testing result of signal is not accounted for when detection, the technical problems such as the accuracy rate of detection is low improve, the high accuracy for examination that there is the present invention weld seam for different defect types can accurately classify when detecting, detect, detect for different signal scales.
Description
Technical field
The present invention relates to weld seam detection technical field more particularly to a kind of supersonic weldings based on wavelet convolution neural network
Stitch detection method.
Background technique
Welded unit refers to the metal structure formed by welding, in numerous necks such as building, automobile manufacture, ship, bridge
There is the figure of welded unit in domain, welded unit quality it is good with it is bad not only largely affect its use it is safe
Property, or even very big hidden danger caused to the safety of the person, and insufficient (incomplete) penetration or it is lack of penetration be the most danger for causing welding structure to fail
Dangerous factor.In order to prevent the accident as caused by problems of welded quality, economic loss caused by welding point failure and really is reduced
The normal operation for protecting welded unit just seems extremely urgent and important with the presence or absence of the detection of defect to weld seam nugget, has pole
Big practical meaning in engineering.
In the welding process, defect is frequently found in inside weld seam.It is exactly to cut that most direct detection method is carried out to it
Interior tissue is observed later, but this belongs to destructive test, it is difficult to meet the needs of automatic online detection.?
In lossless detection method, ultrasound detection has advantage at low cost and can detecte internal structure, therefore uses ultrasonic butt welding
It connects quality and carries out detection as a kind of natural good method.But ultrasonic detecting technology still has technological deficiency at present, it is existing
Supersonic detection method cannot accurately classify when being detected for the weld seams of different defect types, when detection, does not examine
Consider influence of the signal of different scale to testing result, the accuracy rate of detection is low.
Against the above technical problems, the ultrasonic weld seam detection based on wavelet convolution neural network that the invention discloses a kind of
There is the weld seam for different defect types can accurately classify, when detecting for different signals by method, the present invention
The high accuracy for examination that scale is detected, detected.
Summary of the invention
It is super based on wavelet convolution neural network that it is an object of the invention to overcome the deficiencies of the prior art and provide a kind of
Sound wave weld inspection method, cannot be accurate when solving the weld seam detection of different defect types in the prior art
Classify, the technologies such as when detection does not account for influence of the different scale to testing result of signal, and the accuracy rate of detection is low
Problem improves, and there is the present invention weld seam for different defect types can accurately classify when detecting, for not
The high accuracy for examination that same signal scale is detected, detected.
The invention is realized by the following technical scheme: the invention discloses a kind of ultrasounds based on wavelet convolution neural network
Wave weld inspection method, comprising the following steps:
Step 1: weld data is obtained, defect weldment and zero defect weldment are scanned using ultrasonic sensor, obtained
Weld data is obtained, weld data includes defective data and normal data；
Step 2: weld data pretreatment pre-processes weld data according to Preprocessing Algorithm, obtains pretreatment weldering
Data are stitched, pre-processing weld data includes pretreatment defective data and pretreatment normal data, first to weld data progress bar
Te Wozi filtering；Then mean filter is carried out to the weld data after butterworth filter using mean filter；Finally
Weld data after mean filter is normalized；
Step 3: training wavelet convolution neural network, using weld data and pretreatment weld data according to training algorithm
Training wavelet convolution neural network；
Step 4: obtaining weld data to be detected, is scanned using ultrasonic sensor to weldment to be detected, obtain to
Detect weld data；
Step 5: weld data pretreatment to be detected pre-processes weld data to be detected according to Preprocessing Algorithm,
It obtains pre-processing weld data to be detected, butterworth filter is carried out to weld data to be detected first；Then it is filtered using mean value
Wave device carries out mean filter to the weld data to be detected after butterworth filter；Finally to after the mean filter to
Detection weld data is normalized；
Step 6: calculating testing result, will pre-process weld data to be detected and is input to trained wavelet convolution nerve
It is calculated in network, obtains testing result；
Step 7；Judge whether weldment to be detected is qualified, will test result and 0.5 comparison, if testing result is greater than 0.5,
Then the weld seam of weldment to be detected is defective, if testing result less than 0.5, the weld seam zero defect of weldment to be detected.
Further, in order to improve the accuracy of data acquisition, in step 1, the frequency of the center probe of ultrasonic sensor
Rate is 5MHz.
Further, in order to improve the accuracy of weld defect type classification, in step 1, defect weldment includes in weld seam
Has the lack of penetration weldment of leachy weldment, weld seam with the weldment being mingled with, in weld seam, in the weldment and weld seam that weld seam does not merge
Weldment with crackle.
Further, in order to improve the accuracy of testing result, in step 2 and step 5, Preprocessing Algorithm is specifically wrapped
It includes:
Butterworth filter is carried out to data first；The characteristics of butterworth filter is the frequency response curve in passband
It is flat to greatest extent, do not rise and fall, and being then gradually reduced in suppressed frequency band is zero.In the Bode diagram of the logarithm diagonal frequencies of amplitude
On, since a certain boundary angular frequency, amplitude is gradually reduced with the increase of angular frequency, tends to minus infinity.
Then mean filter is carried out to the data after butterworth filter using mean filter；Mean filter is allusion quotation
The linear filtering algorithm of type, it refers to that the template includes surrounding closes on the image to object pixel to a template
Pixel (8 pixels, constitute a Filtering Template, that is, remove object pixel itself around centered on object pixel), then use
The average value of entire pixels in template replaces original pixel value；
Finally the data after mean filter are normalized, data are become 0~1 by normalized
Between decimal.
Further, in order to improve the effect of data prediction, the window of mean filter is dimensioned to 0.0002s.
Further, in order to improve the training effect of wavelet convolution neural network, in step 3, training algorithm includes random
Gradient descent algorithm, Adam algorithm, RMSProp algorithm, Adagrad algorithm, Adadelta algorithm or Adamax algorithm.
Further, in order to improve the training effect of wavelet convolution neural network, the accuracy that data calculate, step are improved
In three, wavelet convolution neural network successively includes the first convolutional layer, the first pond layer, the second convolutional layer, the second pond layer, third
Convolutional layer, third pond layer, Volume Four lamination, the 4th pond layer, the 5th convolutional layer, the 5th pond layer, the first full articulamentum and
Second full articulamentum.Every layer of convolutional layer is made of several convolution units, and the parameter of each convolution unit is to pass through backpropagation
What algorithm optimized.The purpose of convolution algorithm is to extract the different characteristic of input, and first layer convolutional layer can only extract
The rudimentary feature such as levels such as edge, lines and angle, the convolutional layer of more layers can from low-level features iterative extraction it is more complicated
Feature, therefore wavelet convolution neural network is designed to five layers of convolutional layer by the present invention, can extract the more complicated spy of weld seam
Sign improves the accuracy of weld classification detection.In wavelet convolution neural network, a pond is often added between convolutional layer
Layer.Pond layer can effectively reduce the size of parameter matrix, to reduce the number of parameters finally connected entirely in layer.It uses
Pond layer, which can accelerate calculating speed, also prevents over-fitting, and in field of image recognition, image is too big sometimes, we
Need to reduce the quantity of training parameter, it is required periodically to introduce pond layer between subsequent convolutional layer.Pond is only
One the purpose is to reduce the space sizes of image.Pond is completed alone in each depth dimension, therefore the depth of image is kept
It is constant.In wavelet convolution neural network structure, after multiple convolutional layers and pond layer, be connected to 1 or 1 or more entirely connect
Layer is connect, each node of full articulamentum is connected with upper one layer of all nodes, for the characteristic synthetic that front is extracted
Get up.
Further, in order to improve the training effect of wavelet convolution neural network, the first convolutional layer, the second convolutional layer,
The number of three convolutional layers, Volume Four lamination and the filter in the 5th convolutional layer is followed successively by 11, seven, five, three, three
It is a, the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination and the 5th convolutional layer convolution kernel size be followed successively by
(39,1), (27,11), (23,7), (17,5), (13,3), convolution kernel is also referred to as Filter, is with one group of fixed power
The neuron of weight, the two-dimensional matrix of usually n*m, n and m are also the receptive field of neuron.What is deposited in n*m matrix is to receptive field
The coefficient of middle data processing.The filtering of one convolution kernel can be used to extract specific feature (such as can extract contour of object,
Shade etc.).The process for extracting new feature from initial data by convolutional layer becomes featuremap again (feature is reflected
Penetrate), multiple characteristic patterns of former weld image are obtained by the convolution of convolution kernel, improve the accuracy of image recognition, the first convolution
The step-length of layer, the second convolutional layer, third convolutional layer, Volume Four lamination and the 5th convolutional layer is 1, the first convolutional layer, volume Two
Lamination, third convolutional layer, Volume Four lamination and the 5th convolutional layer excitation function be LeakyReLUctant function, the first pond
Change layer, the second pond layer, third pond layer, the 4th pond layer and the 5th pond layer and is all made of maximization pond method, the first pond
The size for changing the pond window of layer, the second pond layer, third pond layer, the 4th pond layer and the 5th pond layer is disposed as 2, first
The number of the neuron of full articulamentum output is ten, and the excitation function of the first full articulamentum is wavelet function, the second full connection
The number of the neuron of layer output is one, and the excitation function of the second full articulamentum is sigmoid function.It is every in neural network
A node receives input value, and input value is passed to next layer, and attribute value can be directly passed to next by input node
Layer (hidden layer or output layer) has letter between hidden layer and output the outputting and inputting of node layer in wavelet convolution neural network
Number relationship, this function are known as excitation function, and without using if excitation function, every layer of neural network all only does linear change
It changes, also still linear transformation after multilayer input superposition.Because the ability to express of linear model is inadequate, excitation function can introduce non-
Linear factor improves the recognition capability of image.
Further, in order to improve the accuracy of the first full articulamentum output result, wavelet function are as follows:
y1For output, it be α is 0.7, β 0.5 that x, which is input,.α and β is that variable element, α and β are obtained by test of many times.
The accuracy that the second full articulamentum exports result is improved by the calculating of wavelet function.
Further, in order to improve the accuracy of the first full articulamentum output result, wavelet function are as follows:
y2For output, x is input, α 0.9, β 0.4.α and β is that variable element, α and β are obtained by test of many times.It is logical
The calculating for crossing wavelet function improves the accuracy of the second full articulamentum output result.
The invention has the following advantages that the present invention is based on the ultrasonic waves of wavelet convolution neural network to be directed to different defect types
Weld seam and different signal scales detected, obtain the weld seam of different defect types and the data of normal weld first,
Then it is calculated by pretreatment, initial data and calculated result training wavelet convolution neural network, then obtains weld seam to be detected
Data are sent in wavelet convolution neural network after pretreatment and are calculated as a result, finally comparing judgement to result
Weld seam is obtained with the presence or absence of defect, the present invention can accurately classify when detecting for the weld seam of different defect types, can
It is high with the accuracy rate for being detected, being detected for different signal scales.
Detailed description of the invention
Fig. 1 is flow chart of the present invention；
Fig. 2 is wavelet convolution neural network structure schematic diagram.
Specific embodiment
It elaborates below to the embodiment of the present invention, the present embodiment carries out under the premise of the technical scheme of the present invention
Implement, the detailed implementation method and specific operation process are given, but protection scope of the present invention is not limited to following implementation
Example.
Embodiment 1
Embodiment 1 discloses a kind of ultrasonic weld seam detection method based on wavelet convolution neural network, as shown in Figure 1,
The following steps are included:
The following steps are included:
Step 101: weld data is obtained, defect weldment and zero defect weldment are scanned using ultrasonic sensor,
Weld data is obtained, weld data includes defective data and normal data；
Step 102: weld data pretreatment pre-processes weld data according to Preprocessing Algorithm, is pre-processed
Weld data, pretreatment weld data include pretreatment defective data and pretreatment normal data；
Step 103: training wavelet convolution neural network, using weld data and pretreatment weld data according to training algorithm
Training wavelet convolution neural network；
Step 104: obtaining weld data to be detected, weldment to be detected is scanned using ultrasonic sensor, obtain
Weld data to be detected；
Step 105: weld data pretreatment to be detected locates weld data to be detected according to Preprocessing Algorithm in advance
Reason, obtains pre-processing weld data to be detected；
Step 106: calculating testing result, weld data to be detected will be pre-processed and be input to trained wavelet convolution nerve
It is calculated in network, obtains testing result；
Step 107；Judge whether weldment to be detected is qualified, will test result and 0.5 comparison, if testing result is greater than 0.5,
Then the weld seam of weldment to be detected is defective, if testing result less than 0.5, the weld seam zero defect of weldment to be detected.
In step 101, the frequency of the center probe of ultrasonic sensor is 5MHz, and defect weldment includes having folder in weld seam
Have the lack of penetration weldment of leachy weldment, weld seam in miscellaneous weldment, weld seam, have in the weldment and weld seam that weld seam does not merge and split
The weldment of line.
In step 102 and step 105, Preprocessing Algorithm is specifically included:
Butterworth filter is carried out to data first；
Then mean filter is carried out to the data after butterworth filter using mean filter；
Finally the data after mean filter are normalized.
The window of mean filter is dimensioned to 0.0002s.
In step 103, training algorithm includes stochastic gradient descent algorithm, Adam algorithm, RMSProp algorithm, Adagrad calculation
Method, Adadelta algorithm or Adamax algorithm.
As shown in Fig. 2, wavelet convolution neural network successively includes the first convolutional layer, the first pond layer, the second convolutional layer,
Two pond layers, third convolutional layer, third pond layer, Volume Four lamination, the 4th pond layer, the 5th convolutional layer, the 5th pond layer,
One full articulamentum and the second full articulamentum, every layer of convolutional layer are made of several convolution units, and the parameter of each convolution unit is
It is optimized by back-propagation algorithm.The purpose of convolution algorithm is to extract the different characteristic of input, first layer convolutional layer
Can only extract some rudimentary features such as levels such as edge, lines and angle, the convolutional layer of more layers can from low-level features iteration
More complicated feature is extracted, therefore wavelet convolution neural network is designed to five layers of convolutional layer by the present invention, can extract weld seam
More complicated feature, improve weld classification detection accuracy.In wavelet convolution neural network, between convolutional layer often
In addition a pond layer.Pond layer can effectively reduce the size of parameter matrix, finally connect in layer entirely to reduce
Number of parameters.Can accelerate calculating speed using pond layer also prevents over-fitting, in field of image recognition, sometimes
Image is too big, it would be desirable to reduce the quantity of training parameter, it is required periodically to introduce pond between subsequent convolutional layer
Change layer.The sole purpose in pond is to reduce the space size of image.Pond is completed alone in each depth dimension, therefore is schemed
The depth of picture remains unchanged.In wavelet convolution neural network structure, after multiple convolutional layers and pond layer, it is connected to 1 or 1
A above full articulamentum, each node of full articulamentum is connected with upper one layer of all nodes, for front is extracted
To characteristic synthetic get up.
Of first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination and the filter in the 5th convolutional layer
Number is followed successively by 11, seven, five, three, three, the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four product
The convolution kernel size of layer and the 5th convolutional layer is followed successively by (39,1), (27,11), (23,7), (17,5), (13,3), convolution kernel
Referred to as Filter is with the neuron of one group of fixed weight, the two-dimensional matrix of usually n*m, and n and m are also nerve
The receptive field of member.What is deposited in n*m matrix is the coefficient to data processing in receptive field.The filtering of one convolution kernel can be used to mention
Take specific feature (such as contour of object, shade can be extracted etc.).It is extracted from initial data by convolutional layer new
The process of feature become feature map (Feature Mapping) again, the multiple of former weld image are obtained by the convolution of convolution kernel
Characteristic pattern improves the accuracy of image recognition, the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination and the 5th
The step-length of convolutional layer is 1, and the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination and the 5th convolutional layer swash
Encouraging function is LeakyReLUctant function, the first pond layer, the second pond layer, third pond layer, the 4th pond layer and
Five pond layers are all made of the pond method that maximizes, the first pond layer, the second pond layer, third pond layer, the 4th pond layer and the
The size of the pond window of five pond layers is disposed as 2, and the number of the neuron of the first full articulamentum output is ten, and first connects entirely
The excitation function for connecing layer is wavelet function, and the number of the neuron of the second full articulamentum output is one, the second full articulamentum
Excitation function is sigmoid function.Each node in neural network receives input value, and input value is passed to next layer,
Attribute value can be directly passed to next layer (hidden layer or output layer) by input node, hidden in wavelet convolution neural network
There is functional relation between layer and output the outputting and inputting of node layer, this function is known as excitation function, without using excitation letter
If number, every layer of neural network all only does linear transformation, also still linear transformation after multilayer input superposition.Because of linear mould
The ability to express of type is inadequate, and excitation function can introduce non-linear factor, improves the recognition capability of image.
Wavelet function are as follows:
y1For output, it be α is 0.7, β 0.5 that x, which is input,.α and β is that variable element, α and β are obtained by test of many times.
By the way that the weld data of various defect types and normal weld data to be brought into wavelet function calculate respectively, and will
Calculated result is compared, and repeatedly carries out bringing test into, the final optimal value for obtaining α and β, α 0.7, β 0.5.
Embodiment 2
Embodiment 2 discloses a kind of ultrasonic weld seam detection method based on wavelet convolution neural network, including following step
It is rapid:
Step 101: weld data is obtained, defect weldment and zero defect weldment are scanned using ultrasonic sensor,
Weld data is obtained, weld data includes defective data and normal data；
Step 102: weld data pretreatment pre-processes weld data according to Preprocessing Algorithm, is pre-processed
Weld data, pretreatment weld data include pretreatment defective data and pretreatment normal data；
Step 103: training wavelet convolution neural network, using weld data and pretreatment weld data according to training algorithm
Training wavelet convolution neural network；
Step 104: obtaining weld data to be detected, weldment to be detected is scanned using ultrasonic sensor, obtain
Weld data to be detected；
Step 105: weld data pretreatment to be detected locates weld data to be detected according to Preprocessing Algorithm in advance
Reason, obtains pre-processing weld data to be detected；
Step 106: calculating testing result, weld data to be detected will be pre-processed and be input to trained wavelet convolution nerve
It is calculated in network, obtains testing result；
Step 107；Judge whether weldment to be detected is qualified, will test result and 0.5 comparison, if testing result is greater than 0.5,
Then the weld seam of weldment to be detected is defective, if testing result less than 0.5, the weld seam zero defect of weldment to be detected.
In step 101, the frequency of the center probe of ultrasonic sensor is 5MHz, and defect weldment includes having folder in weld seam
Have the lack of penetration weldment of leachy weldment, weld seam in miscellaneous weldment, weld seam, have in the weldment and weld seam that weld seam does not merge and split
The weldment of line.
In step 102 and step 105, Preprocessing Algorithm is specifically included:
Butterworth filter is carried out to data first；
Then mean filter is carried out to the data after butterworth filter using mean filter；
Finally the data after mean filter are normalized.
The window of mean filter is dimensioned to 0.0002s.
In step 103, training algorithm includes stochastic gradient descent algorithm, Adam algorithm, RMSProp algorithm, Adagrad calculation
Method, Adadelta algorithm or Adamax algorithm.
Wavelet convolution neural network successively include the first convolutional layer, the first pond layer, the second convolutional layer, the second pond layer,
Third convolutional layer, third pond layer, Volume Four lamination, the 4th pond layer, the 5th convolutional layer, the 5th pond layer, the first full connection
Layer and the second full articulamentum, every layer of convolutional layer are made of several convolution units, and the parameter of each convolution unit is by reversed
What propagation algorithm optimized.The purpose of convolution algorithm is to extract the different characteristic of input, and first layer convolutional layer can only extract
Some rudimentary features such as levels such as edge, lines and angle, the convolutional layer of more layers can from low-level features iterative extraction it is more multiple
Miscellaneous feature, therefore wavelet convolution neural network is designed to five layers of convolutional layer by the present invention, can extract the more complicated of weld seam
Feature, improve weld classification detection accuracy.In wavelet convolution neural network, one is often added between convolutional layer
Pond layer.Pond layer can effectively reduce the size of parameter matrix, to reduce the number of parameters finally connected entirely in layer.
Can accelerate calculating speed using pond layer also prevents over-fitting, and in field of image recognition, image is too big sometimes,
We need to reduce the quantity of training parameter, it is required periodically to introduce pond layer between subsequent convolutional layer.Chi Hua
Sole purpose be reduce image space size.Pond is completed alone in each depth dimension, therefore the depth of image
It remains unchanged.In wavelet convolution neural network structure, after multiple convolutional layers and pond layer, it is connected to 1 or 1 or more
Full articulamentum, each node of full articulamentum is connected with upper one layer of all nodes, for the feature that front is extracted
It integrates.
Of first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination and the filter in the 5th convolutional layer
Number is followed successively by 11, seven, five, three, three, the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four product
The convolution kernel size of layer and the 5th convolutional layer is followed successively by (39,1), (27,11), (23,7), (17,5), (13,3), convolution kernel
Referred to as Filter is with the neuron of one group of fixed weight, the two-dimensional matrix of usually n*m, and n and m are also nerve
The receptive field of member.What is deposited in n*m matrix is the coefficient to data processing in receptive field.The filtering of one convolution kernel can be used to mention
Take specific feature (such as contour of object, shade can be extracted etc.).It is extracted from initial data by convolutional layer new
The process of feature become feature map (Feature Mapping) again, the multiple of former weld image are obtained by the convolution of convolution kernel
Characteristic pattern improves the accuracy of image recognition, the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination and the 5th
The step-length of convolutional layer is 1, and the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination and the 5th convolutional layer swash
Encouraging function is LeakyReLUctant function, the first pond layer, the second pond layer, third pond layer, the 4th pond layer and
Five pond layers are all made of the pond method that maximizes, the first pond layer, the second pond layer, third pond layer, the 4th pond layer and the
The size of the pond window of five pond layers is disposed as 2, and the number of the neuron of the first full articulamentum output is ten, and first connects entirely
The excitation function for connecing layer is wavelet function, and the number of the neuron of the second full articulamentum output is one, the second full articulamentum
Excitation function is sigmoid function.Each node in neural network receives input value, and input value is passed to next layer,
Attribute value can be directly passed to next layer (hidden layer or output layer) by input node, hidden in wavelet convolution neural network
There is functional relation between layer and output the outputting and inputting of node layer, this function is known as excitation function, without using excitation letter
If number, every layer of neural network all only does linear transformation, also still linear transformation after multilayer input superposition.Because of linear mould
The ability to express of type is inadequate, and excitation function can introduce non-linear factor, improves the recognition capability of image.
In the present embodiment, wavelet function are as follows:
y2For output, x is input, α 0.9, β 0.4.α and β is that variable element, α and β are obtained by test of many times.It is logical
It crosses and the weld data of various defect types and normal weld data is brought into wavelet function calculates respectively, and will meter
It calculates result to be compared, repeatedly carries out bringing test into, the final optimal value for obtaining α and β, α 0.9, β 0.4.
Claims (10)
1. a kind of ultrasonic weld seam detection method based on wavelet convolution neural network, which comprises the following steps:
Step 1: weld data is obtained, defect weldment and zero defect weldment are scanned using ultrasonic sensor, welded
Data are stitched, the weld data includes defective data and normal data；
Step 2: weld data pretreatment pre-processes the weld data according to Preprocessing Algorithm, obtains pretreatment weldering
Data are stitched, the pretreatment weld data includes pretreatment defective data and pretreatment normal data；
Step 3: training wavelet convolution neural network, using the weld data and the pretreatment weld data according to training
Algorithm trains wavelet convolution neural network；
Step 4: obtaining weld data to be detected, is scanned using the ultrasonic sensor to weldment to be detected, obtain to
Detect weld data；
Step 5: weld data pretreatment to be detected carries out the weld data to be detected according to the Preprocessing Algorithm pre-
Processing, obtains pre-processing weld data to be detected；
Step 6: calculating testing result, and the pretreatment weld data to be detected is input to the trained wavelet convolution
It is calculated in neural network, obtains testing result；
Step 7；Judge whether the weldment to be detected is qualified, by the testing result and 0.5 comparison, if the testing result
Greater than 0.5, then the weld seam of the weldment to be detected is defective, if the testing result less than 0.5, the weldment to be detected
Weld seam zero defect.
2. as described in claim 1 based on the ultrasonic weld seam detection method of wavelet convolution neural network, which is characterized in that step
In rapid one, the frequency of the center probe of the ultrasonic sensor is 5MHz.
3. as described in claim 1 based on the ultrasonic weld seam detection method of wavelet convolution neural network, which is characterized in that step
In rapid one, the defect weldment includes having that the weldment being mingled with, to have leachy weldment, weld seam in weld seam lack of penetration in weld seam
With the weldment of crackle in the weldment and weld seam that weldment, weld seam do not merge.
4. as described in claim 1 based on the ultrasonic weld seam detection method of wavelet convolution neural network, which is characterized in that step
Rapid two and step 5 in, the Preprocessing Algorithm specifically includes:
Butterworth filter is carried out to data first；
Then mean filter is carried out to the data after butterworth filter using mean filter；
Finally the data after mean filter are normalized.
5. as claimed in claim 4 based on the ultrasonic weld seam detection method of wavelet convolution neural network, which is characterized in that institute
The window for stating mean filter is dimensioned to 0.0002s.
6. as described in claim 1 based on the ultrasonic weld seam detection method of wavelet convolution neural network, which is characterized in that step
In rapid three, the training algorithm include stochastic gradient descent algorithm, Adam algorithm, RMSProp algorithm, Adagrad algorithm,
Adadelta algorithm or Adamax algorithm.
7. as described in claim 1 based on the ultrasonic weld seam detection method of wavelet convolution neural network, which is characterized in that step
In rapid three, the wavelet convolution neural network successively includes the first convolutional layer, the first pond layer, the second convolutional layer, the second pond
Layer, third convolutional layer, third pond layer, Volume Four lamination, the 4th pond layer, the 5th convolutional layer, the 5th pond layer, first connect entirely
Connect layer and the second full articulamentum.
8. as claimed in claim 7 based on the ultrasonic weld seam detection method of wavelet convolution neural network, which is characterized in that institute
State the number of the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination and the filter in the 5th convolutional layer
It is followed successively by 11, seven, five, three, three, first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four
The convolution kernel size of lamination and the 5th convolutional layer is followed successively by (39,1), (27,11), (23,7), (17,5), (13,3), institute
The step-length for stating the first convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination and the 5th convolutional layer is 1, described
First convolutional layer, the second convolutional layer, third convolutional layer, Volume Four lamination and the 5th convolutional layer excitation function be
LeakyReLUctant function, first pond layer, the second pond layer, third pond layer, the 4th pond layer and the described 5th
Pond layer be all made of maximize pond method, first pond layer, the second pond layer, third pond layer, the 4th pond layer and
The size of the pond window of the 5th pond layer is disposed as 2, and the number of the neuron of the first full articulamentum output is ten
A, the excitation function of the first full articulamentum is wavelet function, and the number of the neuron of the second full articulamentum output is
One, the excitation function of the second full articulamentum is sigmoid function.
9. as claimed in claim 8 based on the ultrasonic weld seam detection method of wavelet convolution neural network, which is characterized in that institute
State wavelet function are as follows:
The y1For output, the x is input, and described be α is 0.7, and the β is 0.5.
10. as claimed in claim 8 based on the ultrasonic weld seam detection method of wavelet convolution neural network, which is characterized in that
The wavelet function are as follows:
The y2For output, the x is input, and the α is 0.9, and the β is 0.4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910461584.5A CN110243934A (en) | 2019-05-30 | 2019-05-30 | A kind of ultrasonic weld seam detection method based on wavelet convolution neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910461584.5A CN110243934A (en) | 2019-05-30 | 2019-05-30 | A kind of ultrasonic weld seam detection method based on wavelet convolution neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110243934A true CN110243934A (en) | 2019-09-17 |
Family
ID=67885309
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910461584.5A Pending CN110243934A (en) | 2019-05-30 | 2019-05-30 | A kind of ultrasonic weld seam detection method based on wavelet convolution neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110243934A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111754463A (en) * | 2020-06-02 | 2020-10-09 | 石家庄铁道大学 | Ballastless track CA mortar layer defect detection method based on convolutional neural network |
| 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 |
| CN112150410A (en) * | 2020-08-24 | 2020-12-29 | 浙江工商大学 | Automatic detection method and system for weld defects |
| CN112264731A (en) * | 2020-10-20 | 2021-01-26 | 李小兵 | Control method and device for improving welding quality |
| 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 |
| CN113607817A (en) * | 2021-08-03 | 2021-11-05 | 重庆兆光科技股份有限公司 | Pipeline girth weld detection method and system, electronic equipment and medium |
| CN114088817A (en) * | 2021-10-28 | 2022-02-25 | 扬州大学 | Deep learning flat ceramic membrane ultrasonic defect detection method based on deep features |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105841961A (en) * | 2016-03-29 | 2016-08-10 | 中国石油大学(华东) | Bearing fault diagnosis method based on Morlet wavelet transformation and convolutional neural network |
| CN107451997A (en) * | 2017-07-31 | 2017-12-08 | 南昌航空大学 | A kind of automatic identifying method of the welding line ultrasonic TOFD D scanning defect types based on deep learning |
| CN107766828A (en) * | 2017-10-29 | 2018-03-06 | 北京工业大学 | UAV Landing Geomorphological Classification method based on wavelet convolution neutral net |
| CN108074231A (en) * | 2017-12-18 | 2018-05-25 | 浙江工业大学 | A kind of magnetic sheet detection method of surface flaw based on convolutional neural networks |
| CN108562589A (en) * | 2018-03-30 | 2018-09-21 | 慧泉智能科技（苏州）有限公司 | A method of magnetic circuit material surface defect is detected |
| CN108596922A (en) * | 2018-05-10 | 2018-09-28 | 广东工业大学 | A kind of handware Method of Defect Segmentation and system based on convolutional neural networks |
| CN108918527A (en) * | 2018-05-15 | 2018-11-30 | 佛山市南海区广工大数控装备协同创新研究院 | A kind of printed matter defect inspection method based on deep learning |
| CN109389599A (en) * | 2018-10-25 | 2019-02-26 | 北京阿丘机器人科技有限公司 | A kind of defect inspection method and device based on deep learning |
-
2019
- 2019-05-30 CN CN201910461584.5A patent/CN110243934A/en active Pending
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105841961A (en) * | 2016-03-29 | 2016-08-10 | 中国石油大学(华东) | Bearing fault diagnosis method based on Morlet wavelet transformation and convolutional neural network |
| CN107451997A (en) * | 2017-07-31 | 2017-12-08 | 南昌航空大学 | A kind of automatic identifying method of the welding line ultrasonic TOFD D scanning defect types based on deep learning |
| CN107766828A (en) * | 2017-10-29 | 2018-03-06 | 北京工业大学 | UAV Landing Geomorphological Classification method based on wavelet convolution neutral net |
| CN108074231A (en) * | 2017-12-18 | 2018-05-25 | 浙江工业大学 | A kind of magnetic sheet detection method of surface flaw based on convolutional neural networks |
| CN108562589A (en) * | 2018-03-30 | 2018-09-21 | 慧泉智能科技（苏州）有限公司 | A method of magnetic circuit material surface defect is detected |
| CN108596922A (en) * | 2018-05-10 | 2018-09-28 | 广东工业大学 | A kind of handware Method of Defect Segmentation and system based on convolutional neural networks |
| CN108918527A (en) * | 2018-05-15 | 2018-11-30 | 佛山市南海区广工大数控装备协同创新研究院 | A kind of printed matter defect inspection method based on deep learning |
| CN109389599A (en) * | 2018-10-25 | 2019-02-26 | 北京阿丘机器人科技有限公司 | A kind of defect inspection method and device based on deep learning |
Non-Patent Citations (2)
| Title |
|---|
| 姜文强 等: "一种融合小波变换与卷积神经网络的高相似度图像识别与分类算法", 《计算机工程与科学》 * |
| 郑志远: "焊缝典型缺陷的超声TOFD-D扫成像技术研究", 《中国优秀硕士学位论文全文数据库工程科技Ⅰ辑》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111754463A (en) * | 2020-06-02 | 2020-10-09 | 石家庄铁道大学 | Ballastless track CA mortar layer defect detection method based on convolutional neural network |
| CN111754463B (en) * | 2020-06-02 | 2024-05-14 | 石家庄铁道大学 | Method for detecting CA mortar layer defects of ballastless track based on convolutional neural network |
| 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 |
| CN112150410A (en) * | 2020-08-24 | 2020-12-29 | 浙江工商大学 | Automatic detection method and system for weld defects |
| CN112150410B (en) * | 2020-08-24 | 2023-12-12 | 浙江工商大学 | Automatic detection method and system for weld defects |
| CN112264731A (en) * | 2020-10-20 | 2021-01-26 | 李小兵 | Control method and device for improving welding quality |
| 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 |
| CN113607817A (en) * | 2021-08-03 | 2021-11-05 | 重庆兆光科技股份有限公司 | Pipeline girth weld detection method and system, electronic equipment and medium |
| CN113607817B (en) * | 2021-08-03 | 2023-07-18 | 重庆兆光科技股份有限公司 | Pipeline girth weld detection method, system, electronic equipment and medium |
| CN114088817A (en) * | 2021-10-28 | 2022-02-25 | 扬州大学 | Deep learning flat ceramic membrane ultrasonic defect detection method based on deep features |
| CN114088817B (en) * | 2021-10-28 | 2023-10-24 | 扬州大学 | Deep learning flat ceramic membrane ultrasonic defect detection method based on deep features |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110243934A (en) | A kind of ultrasonic weld seam detection method based on wavelet convolution neural network | |
| Yang et al. | An automatic welding defect location algorithm based on deep learning | |
| CN112232400B (en) | Stainless steel weld ultrasonic defect detection method based on depth feature fusion | |
| CN105118044B (en) | A kind of wheel shape cast article defect automatic testing method | |
| Sambath et al. | Automatic detection of defects in ultrasonic testing using artificial neural network | |
| CN112184693B (en) | Intelligent detection method for welding line defects of ray industrial negative film | |
| CN108426713A (en) | Rolling bearing Weak fault diagnostic method based on wavelet transformation and deep learning | |
| Legendre et al. | Neural classification of Lamb wave ultrasonic weld testing signals using wavelet coefficients | |
| CN107369155A (en) | A kind of cloth surface defect detection method and its system based on machine vision | |
| CN104914164B (en) | A kind of ultrasonic C-scanning imaging method based on DAC curves | |
| Merazi Meksen et al. | Automatic crack detection and characterization during ultrasonic inspection | |
| CN106663198A (en) | Image-based liveness detection for ultrasonic fingerprints | |
| Posilović et al. | Deep learning-based anomaly detection from ultrasonic images | |
| CN107563364A (en) | The discriminating conduct of the fingerprint true and false and fingerprint identification method based on sweat gland | |
| CN112668382A (en) | Support vector machine-based ultrasonic signal identification method for defects in steel ingot | |
| Yuan et al. | Automatic recognition and positioning of wheel defects in ultrasonic B-Scan image using artificial neural network and image processing | |
| Puspitasari et al. | Partial discharge waveform identification using image with convolutional neural network | |
| CN114544920A (en) | Vibration detection device and method for capping beam construction | |
| Filipović et al. | Automated ultrasonic testing of materials based on C-scan flaw classification | |
| Rale et al. | Comparison of different ANN techniques for automatic defect detection in X-Ray images | |
| Zhou et al. | Segmentation of Main Weld Seam Area Based on MGLNS-Retinex Image Enhancement Algorithm | |
| CN117900981B (en) | Production process optimization method and system applied to TC4 tape processing | |
| CN114563473A (en) | Feature extraction and identification method for phased array ultrasonic non-fusion defect | |
| Zhang et al. | Defect detection and recognition technology in wheel images based on deep learning | |
| WO2022209169A1 (en) | Information processing device, determination method, and determination program |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190917 |