CN114088817B - Deep learning flat ceramic membrane ultrasonic defect detection method based on deep features - Google Patents
Deep learning flat ceramic membrane ultrasonic defect detection method based on deep features Download PDFInfo
- Publication number
- CN114088817B CN114088817B CN202111259179.9A CN202111259179A CN114088817B CN 114088817 B CN114088817 B CN 114088817B CN 202111259179 A CN202111259179 A CN 202111259179A CN 114088817 B CN114088817 B CN 114088817B
- Authority
- CN
- China
- Prior art keywords
- ultrasonic
- ceramic membrane
- signals
- defect
- defects
- 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.)
- Active
Links
- 230000007547 defect Effects 0.000 title claims abstract description 85
- 239000000919 ceramic Substances 0.000 title claims abstract description 75
- 239000012528 membrane Substances 0.000 title claims abstract description 62
- 238000001514 detection method Methods 0.000 title claims abstract description 52
- 238000013135 deep learning Methods 0.000 title claims abstract description 8
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 37
- 238000000605 extraction Methods 0.000 claims abstract description 15
- 230000009467 reduction Effects 0.000 claims abstract description 7
- 238000012549 training Methods 0.000 claims abstract description 7
- 239000000523 sample Substances 0.000 claims description 39
- 238000000034 method Methods 0.000 claims description 13
- 230000006870 function Effects 0.000 claims description 12
- 238000001914 filtration Methods 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 6
- 238000011176 pooling Methods 0.000 claims description 5
- 230000011218 segmentation Effects 0.000 claims description 5
- 230000000087 stabilizing effect Effects 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 230000000694 effects Effects 0.000 claims description 4
- 238000005259 measurement Methods 0.000 claims description 4
- 230000007246 mechanism Effects 0.000 claims description 4
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 230000007423 decrease Effects 0.000 claims description 3
- 239000000463 material Substances 0.000 claims description 3
- 238000013178 mathematical model Methods 0.000 claims description 3
- 239000000203 mixture Substances 0.000 claims description 3
- 230000003287 optical effect Effects 0.000 claims description 3
- 239000002245 particle Substances 0.000 claims description 3
- 230000010287 polarization Effects 0.000 claims description 3
- 238000007781 pre-processing Methods 0.000 claims description 3
- 238000011426 transformation method Methods 0.000 claims description 3
- 230000008859 change Effects 0.000 claims description 2
- 238000010606 normalization Methods 0.000 claims description 2
- 230000009466 transformation Effects 0.000 abstract description 7
- 238000010586 diagram Methods 0.000 description 7
- 238000007689 inspection Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 230000004927 fusion Effects 0.000 description 2
- 230000035945 sensitivity Effects 0.000 description 2
- 229910001220 stainless steel Inorganic materials 0.000 description 2
- 239000010935 stainless steel Substances 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 2
- 244000025254 Cannabis sativa Species 0.000 description 1
- PNEYBMLMFCGWSK-UHFFFAOYSA-N aluminium oxide Inorganic materials [O-2].[O-2].[O-2].[Al+3].[Al+3] PNEYBMLMFCGWSK-UHFFFAOYSA-N 0.000 description 1
- 238000005266 casting Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000010612 desalination reaction Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000011010 flushing procedure Methods 0.000 description 1
- 238000003837 high-temperature calcination Methods 0.000 description 1
- 229910010272 inorganic material Inorganic materials 0.000 description 1
- 239000011147 inorganic material Substances 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000035699 permeability Effects 0.000 description 1
- 238000000746 purification Methods 0.000 description 1
- 239000013535 sea water Substances 0.000 description 1
- 239000010865 sewage Substances 0.000 description 1
- 238000005245 sintering Methods 0.000 description 1
- 239000002893 slag Substances 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 238000002604 ultrasonography Methods 0.000 description 1
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
- G01N29/44—Processing the detected response signal, e.g. electronic circuits specially adapted therefor
- G01N29/4409—Processing the detected response signal, e.g. electronic circuits specially adapted therefor by comparison
- G01N29/4418—Processing the detected response signal, e.g. electronic circuits specially adapted therefor by comparison with a model, e.g. best-fit, regression analysis
-
- 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
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Analytical Chemistry (AREA)
- Biochemistry (AREA)
- General Engineering & Computer Science (AREA)
- Chemical & Material Sciences (AREA)
- Immunology (AREA)
- Pathology (AREA)
- Signal Processing (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Investigating Or Analyzing Materials By The Use Of Ultrasonic Waves (AREA)
Abstract
The invention belongs to the field of ultrasonic defect detection, in particular to a deep learning flat ceramic membrane ultrasonic defect detection method based on deep features, which comprises the following steps: step 1: setting up a device, collecting ultrasonic signals and setting up a detection model; step 2: extracting deep features of ultrasonic defect signals; step 3: training a cavity convolutional neural network; step 4: and defining the types of the defects and classifying the defects. According to the invention, the noise reduction treatment is carried out through continuous wavelet transformation, the signals are converted into the time spectrogram, the deep extraction of the ultrasonic signals is realized, the time spectrogram gray level map is used as the input of the cavity convolutional neural network, the time spectrogram gray level map is analyzed and the defects are judged through the cavity convolutional neural network, and the internal defects of the flat ceramic membrane can be accurately, rapidly and objectively detected.
Description
Technical Field
The invention belongs to the technical field of ultrasonic defect detection, and particularly relates to a deep learning flat ceramic membrane ultrasonic defect detection method based on deep features.
Background
The flat ceramic membrane is prepared from inorganic materials such as alumina and the like through a high-temperature sintering process, and has the advantages of high water permeability, strong anti-pollution capability, easy flushing of surface attachments, small working pressure and the like, thus being widely applied to the fields of sea water desalination, sewage purification and the like. The defects of air holes, slag falling and the like of the flat ceramic film in the high-temperature calcination stage need to be screened and removed in the production process. The traditional detection method of the internal defects of the flat ceramic membrane is lamp inspection, the internal defects of the flat ceramic membrane are detected by a detection mode of water inspection, the detection result is inaccurate, the error is large, and the long-term detection of detection personnel is not facilitated.
In 2020, qin Xunpeng et al in an invention patent of a reconstruction and extraction method (publication No. CN111795931 a) for detecting diffraction echo signals aiming at laser ultrasonic defects, a noise adding signal is decomposed by an empirical mode decomposition method to generate a plurality of Intrinsic Mode Function (IMF) components, the IMF components are ordered according to the frequency, the discrete degree of each IMF component is calculated, and corresponding IMF components are selected for reconstruction according to the discrete degree distribution and the IMF component relation, so as to obtain a reconstructed signal. The identification of the signal finally depends on the identification of the detection personnel, and the identification detection has an artificial detection error. In 2020, wei Yunfei et al, an invention patent of a method for detecting internal defects of a cast slab (publication No. CN 112147223A) uses an ultrasonic flaw detector to scan the interior of the cast slab; the sensitivity gain of the ultrasonic flaw detector is adjusted by using the sensitivity gain value at the center thickness of the casting blank, so that the influence of excessive grass waves on the accuracy of a detection result in the detection process is inhibited. The method simply processes the collected signals through functions in the ultrasonic flaw detector, and cannot deeply extract useful information of detection signals. In 2021, zhang Rui et al invented a stainless steel weld ultrasonic defect detection method (publication No. CN 112232400A) based on depth feature fusion by fusing shallow features extracted manually and deep features extracted by CNN, and classifying by using a support vector machine, so as to realize classification and identification of defects of stainless steel. The time resolution and the frequency resolution can not be obtained by using short-time Fourier transformation, and the signal extraction is lack during deep feature extraction, so that the recognition accuracy is reduced.
In summary, although the conventional method can detect the internal defects of the flat ceramic film, the detection accuracy is not high, the detection speed is low, and the characteristics cannot be deeply proposed.
Disclosure of Invention
In order to overcome the defects of the prior art and the method, the invention provides a deep learning flat ceramic membrane ultrasonic defect detection method based on deep features, which can effectively improve the effect of flat ceramic membrane defect detection.
The technical solution for realizing the purpose of the invention is as follows:
a deep learning flat ceramic membrane ultrasonic defect detection method based on deep features comprises the following steps:
step 1: setting up a device, collecting ultrasonic signals and setting up a detection model;
step 2: extracting deep features of ultrasonic defect signals;
step 3: training a cavity convolutional neural network;
step 4: and defining the types of the defects and classifying the defects.
Further, the step 1 specifically includes:
the device comprises an ultrasonic detection moving device and an ultrasonic probe clamping device, the detected ceramic membrane is placed on the ultrasonic detection moving device, the ultrasonic probe is held by the ultrasonic probe clamping device, the ceramic membrane is adjusted to the optimal position of the ultrasonic probe, so that the ultrasonic probe extends over the whole ceramic membrane surface, a plurality of flat ceramic membranes with scratch, fracture and hole defects are replaced, the flat ceramic membranes with scratch, fracture and hole are detected, and the defect detection signals of the flat ceramic membranes at each time are collected,
establishing a detection model, wherein the scattering effect generated when ultrasonic waves strike the defect depends on the relation between the wavelength lambda and the defect diameter d, and Rayleigh scattering occurs when the following conditions are met:
wherein < means less than one tenth of the wavelength lambda, alpha is the ultrasonic attenuation, beta is the average grain volume, s is the optical propagation rate of the material, f is the frequency of the incident light, and → means the inverse ratio.
Establishing a mathematical model h (t) of the single-pulse ultrasonic echo, as shown in a formula (2):
h(t)＝s(t)+δe(t) (2)
where t=0, 1, …, s (t) represents a defect detection signal, δe (t) represents a noise signal, which is distributed as σ 2 (0, 1) a gaussian random variable composition,takes the value of zero as the reflecting surface, tau is the thickness of the detected ceramic film, beta is the peak amplitude of echo wave, f c As the center frequency of the echo, psi is the bandwidth coefficient;
according to the Thompson-Gray measurement model, the transducer occupies V F Echo signal amplitude modeling received by a defect of a region is as follows:
wherein phi is the amplitude of the echo signal, P is the electric power of the propagation cable, i is the wave velocity of the medium, ω is the angular frequency of the ultrasonic wave, δ p To detect density differences between the medium and the defect, U i Representing particle polarization, superscript o represents defect-free field variation, delta c Is the similarity difference of elastic constants, U ij Expressed as the scatterer density.
Neglecting scatterer density fluctuations, then:
as can be seen from the equation (5), if a defect occurs in the propagation path of the received echo, P > 0, and the amplitude Φ of the echo signal decreases as P increases.
Further, ultrasonic testing mobile device includes backup pad, ceramic membrane place the platform, four support columns, stabilizing plate, Y direction advancing device, X direction advancing device, four support nuts and moving platform, the backup pad passes through threaded connection with four support columns, ceramic membrane place the platform and pass through aperture cooperation together through four support nuts and four support columns and support ceramic membrane place the platform through threaded cooperation together, support nuts and corresponding support column threaded connection, support nuts are located between ceramic membrane place the platform and the backup pad, can drive ceramic membrane place the platform through the left and right rotation of support nuts and reciprocate, Y direction advancing device can drive moving platform and advance and retreat in Y direction, and then realizes advancing and retreating in X direction of probe, and X direction advancing device can drive moving platform and retreat in X direction, and then realizes advancing and retreating in X direction of probe, and the top of four support columns passes stabilizing plate and makes through the support of clamp block ultrasonic testing mobile device is more stable, ultrasonic probe clamping device includes probe centre gripping backup pad, guide rail, rotatory clamp block, locating hole and specific ultrasonic signal collection include probe:
when the defect detection is carried out on the flat ceramic membrane, an ultrasonic probe is placed in a probe positioning hole, the clamping block is used for clamping the ultrasonic probe through the rotating handle, then the probe clamping support plate is installed on the X-direction advancing device, the detected ceramic membrane is placed on the ceramic membrane placing table, the ceramic membrane is adjusted to the optimal position of the ultrasonic probe through rotating the supporting nut, the ultrasonic probe extends over the surface of the whole ceramic membrane through the X-direction advancing device and the Y-direction advancing device, a plurality of flat ceramic membranes with scratches, cracks and holes are replaced, the flat ceramic membranes with the scratches, the cracks and the holes are detected, and detection signals of the defects of the flat ceramic membrane at each time are collected.
Further, the step 2 specifically comprises:
the shallow layer features are that the collected signals are directly subjected to simple noise reduction treatment, then the signals are used as the input of a cavity convolutional neural network, and the existing signals are extracted and classified by means of the cavity convolutional neural network. The accuracy of network identification can be greatly reduced by only extracting the characteristics from the signals by virtue of the cavity convolutional neural network; compared with the extraction of shallow features, the extraction of deep features needs to be subjected to multiple image transformation, and a series of image processing such as graying, ROI segmentation, canny operator extraction and the like is performed, so that more deep useful information in the acquired signals is highlighted. These features are not available through the network itself. Deep features combine a series of image transformation and extracted features of the cavity convolutional neural network to increase the accuracy of model identification. Under the same cavity convolutional neural network model, the image extracted by deep features is easier to identify than the image extracted by shallow features.
The acquired ultrasonic signals are one-dimensional time sequence signals, deep features are extracted from the signals acquired in the step 1, the signals are subjected to noise reduction by adopting a continuous wavelet transformation method, useful defect signals of the signals are removed, the denoised time sequence signals are converted into two-dimensional scale diagrams, all the two-dimensional scale diagrams are collected, and the features of the signals in a time domain and a frequency domain can be seen from the two-dimensional scale diagrams. And carrying out image graying and ROI segmentation on the collected two-dimensional scale map, filtering the image, carrying out edge extraction through a Canny operator, extracting the edges of the features of the two-dimensional scale map through the Canny operator, highlighting the edges of the defects, and improving the accuracy of the subsequent model identification.
The preprocessing is to carry out gray processing on the acquired picture, reduce the memory of the picture and accelerate the training and recognition speed of the cavity convolutional neural network. The segmented ROI determines two-dimensional scale map defect feature locations. Filtering the image, and reducing the filtering of the image. The data after each A scan is converted into a two-dimensional scale map, so that not only can useful defect signals be provided in depth, but also the data can be used as input of a cavity convolutional neural network, and the cavity convolutional neural network can be better identified to reach an optimal state.
Further, the step 3 specifically comprises:
the cavity convolution neural network comprises a convolution layer, a pooling layer, an activation function, a batch normalization layer and an attention mechanism, wherein the cavity convolution is to expand the receptive field of the conventional 3*3 convolution core to the receptive field of the 5*5 convolution core through the change of the cavity, and the other expanded cavities are filled with zeros. Under the condition of the same calculated amount, compared with the traditional convolution kernel, the method can obtain a larger receptive field by using the traditional cavity convolution kernel, and obtain more useful information.
Further, the step 4 specifically comprises:
dividing the types of defects of the ceramic membrane into scratches, breaks and holes, establishing a general defect classification formula model, setting a classification task as N, enabling embedding of the classification task to be expressed as x, and evaluating probability of the y-th classification by using a Softmax function, wherein the probability is expressed as shown in a formula (6):
if x belongs to class ζ, the corresponding Softmax loss function is expressed as:
wherein (y) 1 ,…,y k ,…,y N ) Is a classified tag, y k And the value is 0 or 1, and the probability is output through the final Softmax, so that the detection result can be obtained.
Compared with the prior art, the invention has the remarkable advantages that:
the invention provides a deep learning flat ceramic membrane ultrasonic defect detection method based on deep features, which carries out noise reduction treatment through continuous wavelet transformation and converts signals into a time spectrogram, so that deep extraction of ultrasonic signals is realized, the time spectrogram gray level is used as input of a cavity convolutional neural network, the time spectrogram gray level is analyzed and defects are judged through the cavity convolutional neural network, and the internal defects of the flat ceramic membrane can be accurately, rapidly and objectively detected.
Drawings
FIG. 1 is a schematic illustration of a positive three-axis measurement of an ultrasound inspection mobile device.
Fig. 2 is a schematic front view of an ultrasonic inspection mobile device.
Fig. 3 is a top view of an ultrasonic probe holder.
Fig. 4 is a flow chart of ultrasonic detection.
Fig. 5 is a schematic diagram of the structure of a convolutional neural network.
Fig. 6 is a variation of the convolution kernel of a hole convolutional neural network.
FIG. 7 is a flowchart of a method for detecting ultrasonic defects of a flat ceramic membrane based on a cavity convolutional neural network fused with deep features.
Fig. 8 is a flow chart of shallow and deep extraction features.
The device comprises a 1-supporting plate, a 2-ceramic membrane placing table, a 3-supporting column, a 4-stabilizing plate, a 5-Y direction advancing device, a 6-X direction advancing device, a 7-supporting nut, an 8-probe clamping supporting plate, a 9-clamping block, a 10-guide rail, an 11-rotating handle, a 12-probe positioning hole, a 13-clamping block and a 14-moving platform.
Detailed Description
The following describes the implementation of the present invention further with reference to fig. 1, 2, 3 and 7 for detecting defects of ceramic membranes of a flat plate of a cavity convolutional neural network based on fused deep features.
As shown in FIG. 1, the defect detection method for the ceramic membrane of the cavity convolutional neural network flat plate based on the fusion deep layer characteristics comprises the following steps:
step 1: setting up a device, collecting ultrasonic signals and setting up a detection model.
With reference to fig. 1-2, the support plate 1 is connected with the four support columns 3 through threads, the ceramic membrane placing table 2 is matched with the four support columns 3 through apertures, the support nuts 7 are matched with the four support columns 3 through threads to support the ceramic membrane placing table 2, and the support nuts 7 are rotated left and right to drive the ceramic membrane placing table 2 to move up and down. The Y-direction advancing device 5 drives the screw rod to rotate forwards and backwards in the Y direction by the forward and backward rotation of the motor, and the forward and backward rotation of the screw rod drives the moving platform 14 to advance and retreat in the Y direction, so that the forward and backward movement of the probe in the Y direction is realized; similarly, the X-direction advancing means 6 is arranged in the same manner as the Y-direction advancing means 5. The two stabilizing plates 4 are supported by the clamping blocks 13 to make the overall device more stable.
In the case of defect detection of a flat ceramic film, referring to fig. 3, an ultrasonic probe is placed in a probe positioning hole 12, a grip block 9 is held by rotating a handle 11, a probe grip support plate 9 is mounted on an X-direction advancing device 6, the detected ceramic film is placed on a ceramic film placing table 2, the ceramic film is adjusted to an optimal position by rotating a support nut 7, the ultrasonic probe is spread over the entire ceramic film surface by using the X-direction advancing device 6 and the Y-direction advancing device 5, a plurality of flat ceramic films having scratches, breaks, and holes are replaced, the flat ceramic film having scratches, breaks, and holes is detected, and a flat ceramic film defect detection signal for each time is collected.
The detection model is built, and the scattering effect generated when the ultrasonic waves strike the defect depends on the relation between the wavelength lambda and the defect diameter d. Rayleigh scattering occurs when the following conditions are met:
wherein < means less than one tenth of the wavelength lambda, alpha is the ultrasonic attenuation, beta is the average grain volume, s is the optical propagation rate of the material, f is the frequency of the incident light, and → means the inverse ratio.
Establishing a mathematical model h (t) of the single-pulse ultrasonic echo, as shown in a formula (2):
h(t)＝s(t)+δe(t) (2)
where t=0, 1, …, s (t) represents a defect detection signal, δe (t) represents a noise signal, which is distributed as σ 2 (0, 1) a gaussian random variable composition,the value of the reflecting surface is usually zero, tau is the thickness of the flat ceramic film sample, beta is the peak amplitude of echo, and f c Is the center frequency of the echo, and ψ is the bandwidth factor.
According to the Thompson-Gray measurement model, the transducer occupies V F Echo signal received by defect of areaThe magnitude of the number can be modeled as follows:
wherein phi is the amplitude of the echo signal, P is the electric power of the propagation cable, i is the wave velocity of the medium, ω is the angular frequency of the ultrasonic wave, δ p To detect density differences between the medium and the defect, U i Representing particle polarization, superscript o represents defect-free field variation, delta c Is the similarity difference of elastic constants, U ij Expressed as the scatterer density.
If scatterer density fluctuations are negligible:
as can be seen from the equation (5), if a defect occurs in the propagation path of the received echo, P > 0, and the amplitude Φ of the echo signal decreases as P increases.
Step 2: deep features of ultrasonic defect signals are extracted.
As shown in fig. 8, the shallow features are obtained by directly performing simple noise reduction processing on the acquired signals, using the signals as input of a cavity convolutional neural network, and extracting features and classifying the existing signals by means of the cavity convolutional neural network. The accuracy of network identification can be greatly reduced by only extracting the characteristics from the signals by virtue of the cavity convolutional neural network; compared with the extraction of shallow features, the extraction of deep features needs to be subjected to multiple image transformation, and a series of image processing such as graying, ROI segmentation, canny operator extraction and the like is performed, so that more deep useful information in the acquired signals is highlighted. These features are not available through the network itself. Deep features combine a series of image transformation and extracted features of the cavity convolutional neural network to increase the accuracy of model identification. Under the same cavity convolutional neural network model, the image extracted by deep features is easier to identify than the image extracted by shallow features.
The acquired ultrasonic signals are one-dimensional time sequence signals, deep features are extracted from the signals acquired in the step 1, the signals are subjected to noise reduction by adopting a continuous wavelet transformation method, useful defect signals of the signals are removed, the denoised time sequence signals are converted into two-dimensional scale diagrams, all the two-dimensional scale diagrams are collected, and the features of the signals in a time domain and a frequency domain can be seen from the two-dimensional scale diagrams. And carrying out image graying and ROI segmentation on the collected two-dimensional scale map, filtering the image, carrying out edge extraction through a Canny operator, extracting the edges of the features of the two-dimensional scale map through the Canny operator, highlighting the edges of the defects, and improving the accuracy of the subsequent model identification.
The preprocessing is to carry out gray processing on the acquired picture, reduce the memory of the picture and accelerate the training and recognition speed of the cavity convolutional neural network. The segmented ROI determines two-dimensional scale map defect feature locations. Filtering the image, and reducing the filtering of the image. The data after each A scan is converted into a two-dimensional scale map, so that not only can useful defect signals be provided in depth, but also the data can be used as input of a cavity convolutional neural network, and the cavity convolutional neural network can be better identified to reach an optimal state.
Step 3: training of the hole convolutional neural network.
As shown in fig. 5, the hole convolutional neural network includes a convolutional layer, a pooling layer, an activation function, a batch norm, and an attention mechanism. The hole convolutional neural network comprises a convolutional layer, a pooling layer, an activation function, a Batchnorm, an attention mechanism and the like. As shown in fig. 6, the hole convolution is to expand the receptive field of the conventional 3*3 convolution kernel to the receptive field of the 5*5 convolution kernel by changing the hole, and the other holes are filled with zeros. Under the condition of the same calculated amount, compared with the traditional convolution kernel, the method can obtain a larger receptive field by using the traditional cavity convolution kernel, and obtain more useful information.
And taking the gray level images of scratches, cracks and holes as the input of the cavity convolutional neural network, extracting the characteristics of the gray level images through the convolutional layer, and reducing the dimension and the size of the images through the pooling layer. The model of the cavity convolutional neural network can be trained after at least 50 iterations.
Step 4: and defining the types of the defects and classifying the defects.
The invention divides the defect types of ceramic membranes into scratches, cracks and holes. Establishing a general defect classification formula model, setting a classification task as N, wherein the embedding of the classification task is expressed as x, and the probability of the y-th classification can be evaluated by using a Softmax function, as shown in a formula (6):
if x belongs to class ζ, the corresponding Softmax penalty function can be expressed as:
wherein (y) 1 ,…,y k ,…,y N ) Is a classified tag, y k And the value is 0 or 1, and the probability is output through the final Softmax, so that the detection result can be obtained.
The foregoing has outlined and described the basic principles, features, and advantages of the present invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, and that the above embodiments and descriptions are merely illustrative of the principles of the present invention, and various changes and modifications may be made without departing from the spirit and scope of the invention, which is defined in the appended claims. The scope of the invention is defined by the appended claims and equivalents thereof.
Claims (3)
1. The method for detecting the ultrasonic defects of the flat ceramic membrane based on deep learning of deep features is characterized by comprising the following steps:
step 1: setting up a device, collecting ultrasonic signals and setting up a detection model;
step 2: extracting deep features of ultrasonic defect signals;
step 3: training a cavity convolutional neural network;
step 4: defining the types of defects and classifying the defects;
the step 1 specifically includes:
the device comprises an ultrasonic detection moving device and an ultrasonic probe clamping device, the detected ceramic membrane is placed on the ultrasonic detection moving device, the ultrasonic probe is held by the ultrasonic probe clamping device, the ceramic membrane is adjusted to the optimal position of the ultrasonic probe, so that the ultrasonic probe extends over the whole ceramic membrane surface, a plurality of flat ceramic membranes with scratch, fracture and hole defects are replaced, the flat ceramic membranes with scratch, fracture and hole are detected, and the defect detection signals of the flat ceramic membranes at each time are collected,
establishing a detection model, wherein the scattering effect generated when ultrasonic waves strike the defect depends on the relation between the wavelength lambda and the defect diameter d, and Rayleigh scattering occurs when the following conditions are met:
wherein < means less than one tenth of the wavelength lambda, alpha is ultrasonic attenuation, beta 1 For average grain volume, s is the material optical propagation rate, f the frequency of the incident light, and → is the inverse ratio,
establishing a mathematical model h (t) of the single-pulse ultrasonic echo, as shown in a formula (2):
h(t)＝s(t)+δe(t) (2)
where t=0, 1, …, s (t) represents a defect detection signal, δe (t) represents a noise signal, which is distributed as σ 2 (0, 1) a gaussian random variable composition,takes the reflecting surface as zero, tau is the thickness of the detected ceramic film, beta 2 For echo peak amplitude, f c As the center frequency of the echo, psi is the bandwidth coefficient;
according to the Thompson-Gray measurement model, the transducer occupies V F Echo signal amplitude modeling received by a defect of a region is as follows:
wherein phi is the amplitude of the echo signal, P is the electric power of the propagation cable, i is the wave velocity of the medium, ω is the angular frequency of the ultrasonic wave, δ p To detect density differences between the medium and the defect, U i Representing particle polarization, superscript o represents defect-free field variation, delta c Is the similarity difference of elastic constants, U ij Expressed as the density of the scatterer,
neglecting scatterer density fluctuations, then:
as can be seen from the equation (5), if a defect occurs in the propagation path of the received echo, P > 0, the amplitude Φ of the echo signal decreases as P increases;
the ultrasonic detection mobile device comprises a supporting plate (1), a ceramic membrane placing table (2), four supporting columns (3), a stabilizing plate (4), a Y-direction advancing device (5), an X-direction advancing device (6), four supporting screw caps (7) and a mobile platform (14), wherein the supporting plate (1) and the four supporting columns (3) are connected through threads, the ceramic membrane placing table (2) is matched with each other through the four supporting columns (3) through apertures, the four supporting screw caps (7) and the four supporting columns (3) are matched with each other through threads to support the ceramic membrane placing table (2), the supporting screw caps (7) are in threaded connection with the corresponding supporting columns (3), the supporting screw caps (7) are positioned between the ceramic membrane placing table (2) and the supporting plate (1), the ceramic membrane placing table (2) can be driven to move up and down through left and right rotations of the supporting screw caps (7), the Y-direction advancing device (5) can drive the mobile platform (14) to advance and retreat in the Y direction, further realize the forward and retreat of the probe in the Y direction, the X-direction advancing device (6) can drive the mobile platform (14) to move forward and retreat in the X direction, the X-direction (14) can be driven to move forward and retreat in the X direction (3) through the supporting platform (3) through the supporting plate (3) and the stabilizing plate (13), the ultrasonic probe clamping device comprises a probe clamping supporting plate (8), a clamping block (9), a guide rail (10), a rotary handle (11), a probe positioning hole (12) and a clamping block (13), and the ultrasonic signal acquisition specifically comprises the following steps:
when the defect detection is carried out on the flat ceramic membrane, an ultrasonic probe is placed in a probe positioning hole (12), the ultrasonic probe is held by a clamping block (9) through a rotating handle (11), then a probe clamping support plate (9) is installed on an X-direction advancing device (6), the detected ceramic membrane is placed on a ceramic membrane placing table (2), the ceramic membrane is adjusted to the optimal position by rotating a support nut (7), the ultrasonic probe extends over the whole ceramic membrane surface through the X-direction advancing device (6) and a Y-direction advancing device (5), a plurality of flat ceramic membranes with scratches, cracks and holes are replaced, the flat ceramic membranes with the scratches, the cracks and the holes are detected, and detection signals of the defects of the flat ceramic membranes at each time are collected;
the step 2 is specifically as follows: the acquired ultrasonic signals are one-dimensional time sequence signals, deep features are extracted from the signals acquired in the step 1, the signals are subjected to noise reduction by adopting a continuous wavelet transformation method, useful defect signals of the signals are removed, the denoised time sequence signals are converted into two-dimensional scale images, all the two-dimensional scale images are collected, the features of the signals in a time domain and a frequency domain can be seen from the two-dimensional scale images, the collected two-dimensional scale images are subjected to image graying and ROI segmentation, the images are filtered, edge extraction is performed through a Canny operator, the edges of the features of the two-dimensional scale images are extracted through the Canny operator, the edges of the defects are highlighted, and the accuracy of subsequent model identification is improved,
the preprocessing is to carry out gray processing on the acquired picture, reduce the memory of the picture, accelerate the training and recognition speed of the cavity convolutional neural network, divide the ROI to determine the defect characteristic position of the two-dimensional scale picture, filter the image, reduce the filtering of the image, convert the data after each A scan into the two-dimensional scale picture, not only can provide useful defect signals in depth, but also can be used as the input of the cavity convolutional neural network, and the cavity convolutional neural network can be better recognized to reach the optimal state.
2. The method according to claim 1, wherein step 3 is specifically:
the cavity convolution neural network comprises a convolution layer, a pooling layer, an activation function, a batch normalization layer and an attention mechanism, wherein the cavity convolution is to expand the receptive field of the conventional 3*3 convolution core to the receptive field of the 5*5 convolution core through the change of the cavity, the other expanded cavities are filled with zeros, and the same calculated amount can be obtained by increasing the cavity convolution neural network, so that more useful information can be obtained.
3. The method according to claim 2, wherein step 4 is specifically:
dividing the types of defects of the ceramic membrane into scratches, breaks and holes, establishing a general defect classification formula model, setting a classification task as N, enabling embedding of the classification task to be expressed as x, and evaluating probability of the y-th classification by using a Softmax function, wherein the probability is expressed as shown in a formula (6):
if x belongs to class ζ, the corresponding Softmax loss function is expressed as:
wherein (y) 1 ,…,y k ,…,y N ) Is a classified tag, y k Take a value of 0 or 1, output a summary through the final SoftmaxThe rate can obtain the detection result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111259179.9A CN114088817B (en) | 2021-10-28 | 2021-10-28 | Deep learning flat ceramic membrane ultrasonic defect detection method based on deep features |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111259179.9A CN114088817B (en) | 2021-10-28 | 2021-10-28 | Deep learning flat ceramic membrane ultrasonic defect detection method based on deep features |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114088817A CN114088817A (en) | 2022-02-25 |
| CN114088817B true CN114088817B (en) | 2023-10-24 |
Family
ID=80298090
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111259179.9A Active CN114088817B (en) | 2021-10-28 | 2021-10-28 | Deep learning flat ceramic membrane ultrasonic defect detection method based on deep features |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114088817B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117428291A (en) * | 2023-12-18 | 2024-01-23 | 南京理工大学 | Weld bead fusion width quantification method based on sonogram characteristic analysis |
Citations (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| RU2246724C1 (en) * | 2003-08-25 | 2005-02-20 | Кубланов Владимир Семенович | Method of ultrasonic testing of material quality |
| WO2017063355A1 (en) * | 2015-10-15 | 2017-04-20 | 浙江大学 | Method for automatically identifying defect type of polyethylene electrofusion joint by means of ultrasonic phased array inspection |
| CN107037131A (en) * | 2017-05-04 | 2017-08-11 | 中南大学 | A kind of tiny flaw supersonic detection method theoretical based on the extreme value distribution |
| CN108693251A (en) * | 2018-02-19 | 2018-10-23 | 江苏新时高温材料股份有限公司 | The 3 D detection method of hollow plate type ceramic film deep zone defect is realized based on ultrasonic technique |
| CN108694711A (en) * | 2018-02-19 | 2018-10-23 | 江苏新时高温材料股份有限公司 | Ceramic membrane surface defect two-dimensional detecting method is realized based on machine vision technique |
| CN108764006A (en) * | 2018-02-05 | 2018-11-06 | 北京航空航天大学 | A kind of SAR image object detection method based on deeply study |
| CN108896660A (en) * | 2018-07-09 | 2018-11-27 | 中南大学 | A kind of hexagonal crystal material near surface tiny flaw detection method based on shear wave back scattering |
| CN108961203A (en) * | 2018-02-19 | 2018-12-07 | 江苏新时高温材料股份有限公司 | A kind of three-dimensional rebuilding method of fusion ultrasound and the hollow plate type ceramic film defect of machine vision technique |
| CN109444863A (en) * | 2018-10-23 | 2019-03-08 | 广西民族大学 | A kind of estimation method of the narrowband ultrasonic echo number based on convolutional neural networks |
| CN110045015A (en) * | 2019-04-18 | 2019-07-23 | 河海大学 | A kind of concrete structure Inner Defect Testing method 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 |
| WO2019200753A1 (en) * | 2018-04-17 | 2019-10-24 | 平安科技（深圳）有限公司 | Lesion detection method, device, computer apparatus and storage medium |
| EP3581961A1 (en) * | 2018-06-13 | 2019-12-18 | Technische Universität München | Method and apparatus for ultrasound imaging with improved beamforming |
| CN111795931A (en) * | 2020-07-06 | 2020-10-20 | 武汉理工大学 | Reconstruction extraction method for laser ultrasonic defect detection diffraction echo signal |
| CN112147223A (en) * | 2020-10-14 | 2020-12-29 | 首钢京唐钢铁联合有限责任公司 | Method for detecting internal defects of casting blank |
| WO2021004174A1 (en) * | 2019-07-11 | 2021-01-14 | 暨南大学 | Ultrasound image-based automatic measuring method and device for intrapartum cephalopelvic relationship |
| 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 |
| CN112614091A (en) * | 2020-12-10 | 2021-04-06 | 清华大学 | Ultrasonic multi-section data detection method for congenital heart disease |
| CN113298757A (en) * | 2021-04-29 | 2021-08-24 | 同济大学 | Metal surface defect detection method based on U-NET convolutional neural network |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6753965B2 (en) * | 2001-01-09 | 2004-06-22 | The University Of Hong Kong | Defect detection system for quality assurance using automated visual inspection |
| US20080004527A1 (en) * | 2006-04-05 | 2008-01-03 | Coleman D Jackson | High-resolution ultrasound spectral and wavelet analysis of vascular tissue |
-
2021
- 2021-10-28 CN CN202111259179.9A patent/CN114088817B/en active Active
Patent Citations (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| RU2246724C1 (en) * | 2003-08-25 | 2005-02-20 | Кубланов Владимир Семенович | Method of ultrasonic testing of material quality |
| WO2017063355A1 (en) * | 2015-10-15 | 2017-04-20 | 浙江大学 | Method for automatically identifying defect type of polyethylene electrofusion joint by means of ultrasonic phased array inspection |
| CN107037131A (en) * | 2017-05-04 | 2017-08-11 | 中南大学 | A kind of tiny flaw supersonic detection method theoretical based on the extreme value distribution |
| CN108764006A (en) * | 2018-02-05 | 2018-11-06 | 北京航空航天大学 | A kind of SAR image object detection method based on deeply study |
| CN108693251A (en) * | 2018-02-19 | 2018-10-23 | 江苏新时高温材料股份有限公司 | The 3 D detection method of hollow plate type ceramic film deep zone defect is realized based on ultrasonic technique |
| CN108694711A (en) * | 2018-02-19 | 2018-10-23 | 江苏新时高温材料股份有限公司 | Ceramic membrane surface defect two-dimensional detecting method is realized based on machine vision technique |
| CN108961203A (en) * | 2018-02-19 | 2018-12-07 | 江苏新时高温材料股份有限公司 | A kind of three-dimensional rebuilding method of fusion ultrasound and the hollow plate type ceramic film defect of machine vision technique |
| WO2019200753A1 (en) * | 2018-04-17 | 2019-10-24 | 平安科技（深圳）有限公司 | Lesion detection method, device, computer apparatus and storage medium |
| EP3581961A1 (en) * | 2018-06-13 | 2019-12-18 | Technische Universität München | Method and apparatus for ultrasound imaging with improved beamforming |
| CN108896660A (en) * | 2018-07-09 | 2018-11-27 | 中南大学 | A kind of hexagonal crystal material near surface tiny flaw detection method based on shear wave back scattering |
| CN109444863A (en) * | 2018-10-23 | 2019-03-08 | 广西民族大学 | A kind of estimation method of the narrowband ultrasonic echo number based on convolutional neural networks |
| CN110045015A (en) * | 2019-04-18 | 2019-07-23 | 河海大学 | A kind of concrete structure Inner Defect Testing method 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 |
| WO2021004174A1 (en) * | 2019-07-11 | 2021-01-14 | 暨南大学 | Ultrasound image-based automatic measuring method and device for intrapartum cephalopelvic relationship |
| CN111795931A (en) * | 2020-07-06 | 2020-10-20 | 武汉理工大学 | Reconstruction extraction method for laser ultrasonic defect detection diffraction echo signal |
| 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 |
| CN112147223A (en) * | 2020-10-14 | 2020-12-29 | 首钢京唐钢铁联合有限责任公司 | Method for detecting internal defects of casting blank |
| CN112614091A (en) * | 2020-12-10 | 2021-04-06 | 清华大学 | Ultrasonic multi-section data detection method for congenital heart disease |
| CN113298757A (en) * | 2021-04-29 | 2021-08-24 | 同济大学 | Metal surface defect detection method based on U-NET convolutional neural network |
Non-Patent Citations (4)
| Title |
|---|
| 基于频域特征的图像深度信息提取方法;廖均梅等;《自动化与仪器仪表》;第167-168,171页 * |
| 超声 A 扫描信号建模及其缺陷识别方法研究;李力等;《中国测试》;第14-16,32页 * |
| 超声检测缺陷分类的降噪及特征提取问题研究;刘旭等;《中国矿业大学学报》;第248-251页 * |
| 陶瓷膜表面缺陷的表征与分类研究;孙进等;《徐州工程学院学报》;第76-79页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114088817A (en) | 2022-02-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113888471B (en) | High-efficiency high-resolution defect nondestructive testing method based on convolutional neural network | |
| AU2019101789A4 (en) | Method for quantitatively measuring spatial structure of soil particulate organic matter | |
| CN109507304B (en) | Defect detection method based on ultrasonic flaw detection | |
| CN114088817B (en) | Deep learning flat ceramic membrane ultrasonic defect detection method based on deep features | |
| CN1641504A (en) | Method for controlling quality of industry process especially of laser welding process | |
| CN108693251B (en) | Three-dimensional detection method for realizing deep-layer defects of hollow plate-type ceramic membrane based on ultrasonic technology | |
| CN109165617A (en) | A kind of ultrasonic signal sparse decomposition method and its signal de-noising and defect inspection method | |
| CN111521683A (en) | Material defect ultrasonic three-dimensional imaging method based on multi-array element equal-amplitude synchronous excitation | |
| CN100495018C (en) | Ultrasonic method and device for testing macroscopic cleanness of continuous casting billet | |
| CN107727749A (en) | A kind of ultrasonic quantitative detection method based on wavelet packet fusion feature extraction algorithm | |
| CN109682892B (en) | Signal denoising method based on time-frequency analysis | |
| Giurgiutiu et al. | Comparison of short-time fourier transform and wavelet transform of transient and tone burst wave propagation signals for structural health monitoring | |
| CN109632974B (en) | Echo signal separation method for ultrasonic flaw detection | |
| CN116429700A (en) | Laser electromagnetic ultrasonic defect detection system and laser electromagnetic ultrasonic SAFT imaging detection method for additive titanium alloy | |
| CN115184276A (en) | Towed plankton polarization imaging recorder | |
| CN113409202B (en) | Ultrasonic image restoration method based on point spread function parameter optimization | |
| CN109507291B (en) | Signal preprocessing method | |
| CN112183297B (en) | Ultrasonic phased array signal sparse feature extraction method | |
| CN115078190A (en) | Suspension body on-site laser granularity data processing method and device | |
| CN109711333B (en) | Ultrasonic signal receiving and processing method based on signal section segmentation | |
| Wu et al. | Using ground penetrating radar to classify features within structural timbers | |
| Chen | Pattern recognition in nondestructive evaluation of materials | |
| CN113552218B (en) | Array ultrasonic signal amplitude and phase characteristic weighting-based defect qualitative detection method | |
| Jayasudha et al. | Feature analysis for characterization of phased array images based on hilbert transform | |
| Chen et al. | Texture analysis of UTDR images for enhancement of monitoring and diagnosis of membrane filtration |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |