CN112668527A - Ultrasonic guided wave semi-supervised imaging detection method - Google Patents
Ultrasonic guided wave semi-supervised imaging detection method Download PDFInfo
- Publication number
- CN112668527A CN112668527A CN202011642474.8A CN202011642474A CN112668527A CN 112668527 A CN112668527 A CN 112668527A CN 202011642474 A CN202011642474 A CN 202011642474A CN 112668527 A CN112668527 A CN 112668527A
- Authority
- CN
- China
- Prior art keywords
- signal
- damage
- network
- imaging
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000003384 imaging method Methods 0.000 title claims abstract description 59
- 238000001514 detection method Methods 0.000 title claims abstract description 25
- 238000012549 training Methods 0.000 claims abstract description 43
- 238000012360 testing method Methods 0.000 claims abstract description 16
- 230000009467 reduction Effects 0.000 claims abstract description 12
- 230000002159 abnormal effect Effects 0.000 claims abstract description 10
- 238000000034 method Methods 0.000 claims abstract description 10
- 238000005516 engineering process Methods 0.000 claims abstract description 7
- 230000006870 function Effects 0.000 claims description 21
- 238000012544 monitoring process Methods 0.000 claims description 15
- 238000013507 mapping Methods 0.000 claims description 12
- 230000005540 biological transmission Effects 0.000 claims description 10
- 238000010606 normalization Methods 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 6
- 230000005284 excitation Effects 0.000 claims description 6
- 238000011176 pooling Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 230000036962 time dependent Effects 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 5
- 238000013527 convolutional neural network Methods 0.000 claims description 5
- 230000003902 lesion Effects 0.000 claims description 5
- 239000000284 extract Substances 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 3
- 230000008859 change Effects 0.000 claims description 3
- 210000002569 neuron Anatomy 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 239000002131 composite material Substances 0.000 abstract description 8
- 238000012512 characterization method Methods 0.000 abstract description 2
- 238000002474 experimental method Methods 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 9
- 238000000605 extraction Methods 0.000 description 5
- 238000005070 sampling Methods 0.000 description 3
- 229920000049 Carbon (fiber) Polymers 0.000 description 2
- 239000004917 carbon fiber Substances 0.000 description 2
- 230000036541 health Effects 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- VNWKTOKETHGBQD-UHFFFAOYSA-N methane Chemical compound C VNWKTOKETHGBQD-UHFFFAOYSA-N 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000009659 non-destructive testing Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000002604 ultrasonography Methods 0.000 description 1
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses an ultrasonic guided wave semi-supervised imaging detection method. The method comprises the following steps: standardizing signals collected from different sensors to ensure the consistency of training samples; obtaining a dimensionality reduction representation by using a convolutional coding network, and ensuring dimensionality reduction quality through a reconstruction error; adding signal interference in the dimensionality reduction representation of the normal signal, and reconstructing a simulated damage signal; constructing a twin network by using a semi-supervised learning framework, training the twin network by using an abnormal signal, and obtaining a damage index of the twin network by comparing a normal state signal and a test signal in the twin network; and visually describing the damage state by using a probability imaging technology. Experiments on the composite material plate show that the damage index of the structure can be effectively obtained, and damage imaging characterization is realized by combining the time coefficient.
Description
Technical Field
The invention belongs to the field of nondestructive testing, and particularly relates to an ultrasonic guided wave semi-supervised imaging detection method.
Background
The ultrasonic guided wave imaging can characterize the damage of the structure through imaging characteristics and time difference, and becomes a promising structural health monitoring tool. Most imaging algorithms today extract and select imaging features from the first wave and scatter signals manually. Such feature selection may reduce the reliability of the features, thereby significantly reducing the performance of the monitoring model.
The deep learning-based ultrasonic guided wave detection has great potential in the aspect of feature extraction, but still has some problems to be solved in the aspect of imaging feature extraction. First, semi-supervised feature extraction without pattern separation and selection remains a challenge. Especially in the case of unknown material parameters, the traditional feature extraction based on discrete features is difficult to realize. In addition, the received guided wave modes at specific excitation frequencies all contain damage related information and should be comprehensively used. When using multi-sensor monitoring, it is also necessary to balance the effects of different sensor amplitudes.
Disclosure of Invention
The present invention is intended to solve at least the above-mentioned technical problems. Therefore, the invention aims to provide an automatic extraction algorithm of imaging characteristics in multi-sensor structural health monitoring.
The object of the present invention is achieved by at least one of the following aspects.
An ultrasonic guided wave semi-supervised imaging detection method comprises the following steps:
s1, standardizing signals collected from different sensors to ensure the consistency of training samples;
s2, obtaining dimensionality reduction representation by using a convolutional coding network, and ensuring dimensionality reduction quality through reconstruction errors;
s3, adding signal interference in the dimensionality reduction representation of the normal signal, and reconstructing a simulated damage signal;
s4, constructing a twin network by using a semi-supervised learning framework, training the twin network by using an abnormal signal, and comparing a normal state signal and a test signal in the twin network to obtain a damage index of the twin network;
and S5, visually describing the damage state by using a probability imaging technology.
Further, in step S1, T of the maximum value is set for the standard input signalmax% set as threshold for signal acquisition; the threshold interception position is aligned with the nearest channel, so that more complete wave packets are intercepted; all samples used for training and testing are intercepted in the same mode, and then normalization processing is carried out on each training sample, so that the difference between different sensors caused by a deep neural network is avoided;
signal normalization means that all signal samples used for training and testing are truncated as:
Si＝min[where(Xi>Xmax×Tmax)]； (9)
Si＝Si-min(Si)； (10)
wherein XiFor signal samples received by the i-th sensor, XmaxIs the maximum value of the training set, SiThe starting point of the intercepted signal is; each signal sample will be normalized to [ -1,1 ] according to the following equation]：
wherein XjFor the jth element, N, in the training samplefFor the normalization factor, max (x) and min (x) are the maximum and minimum values of the training samples, respectively.
Further, in step S2, the convolutional encoding network extracts the hidden layer representation of the signal by convolution and pooling, performs signal reconstruction by upsampling and deconvolution, and sets the convolution kernel size to m1Setting the number of convolution kernels to n1(ii) a The specific convolution calculation is activated by using a Leaky ReLU function which can generate non-zero slope so as to keep the invariance of the waveform characteristics:
wherein ,Yl kTo an activation value, Xl kIs a weighted sum of the ith neurons of the kth layer, PreluIs the slope parameter of the activation function.
Further, in step S3, the disturbance in the hidden layer may change the reconstructed signal, so as to simulate the damage state signal; the number of convolution kernels of the last convolution layer in the convolutional coding network is set to 1, and the hidden representation of the layer is RHIs a vector; given a magnification factor RdThe ratio in the hidden representation is RdPart of the maximum values of (A) will be added with random disturbance, which is obtained by normal distribution of random numbers each time the random disturbance is added, and the random disturbance DaObey a normal distribution of N (0, 1); in order to ensure the proportion of the disturbance, when the disturbance D is randomaIs more than [0.1,1 ]]Will carry on the value again to it during interval scope:
wherein ,
for hidden representation, the ratio is RdThe qth hidden representation element of the partial maximum of (a).
Further, in step S4, the twin network includes two convolutional neural networks with shared parameters, the input signal pair is mapped to the high-dimensional space through the same convolutional network, the signal difference of the signal in the high-dimensional space is obtained, and the structural state is identified according to the mapping distance represented by the high-dimensional space, which is specifically as follows:
for input signal pair [ X ] composed of normal signal and test signaln,Xt]Mapping function M implemented by convolutional neural network in twin networkw(X) mapping the input signal to a high dimensional space; then mapping the signal Mw(Xn) and Mw(Xt) Is calculated by the Euclidean distance EIs the output of the twin network:
E(Mw(Xn),Mw(Xt))＝||Mw(Xn),Mw(Xt)||。 (14)
furthermore, semi-supervised training is realized in the training process of the twin network, namely only signals in a normal state are collected without collecting real damage signals; using normal state input pairs [ X ]n1,Xn2]And simulated damage input pair [ X ]n,Xs]Network training is carried out and network parameters are updated; setting the label of the normal input pair as 0, setting the label of the simulated damage pair as 1, wherein the number of the labels of the normal input pair is the same as that of the simulated damage pair; binary cross entropy was used as a loss function:
wherein 
Is a label of the input signal pair, YkIs the output of the twin network, n is the batch size in the training process, Xn1,Xn2For any two of the normal signals, XnFor simulating normal state signal samples in the lesion input pairs, XsIs a sample of a signal that simulates the state of damage.
Further, in step S5, the detection signals of different sensors are standardized, and the damage index of the detection signal of each sensor can be predicted through the twin network, so as to evaluate the structural state; carrying out probability imaging by using damage indexes of different sensors and displaying a structural state; the probability imaging technology is specifically as follows:
assigning damage probability to imaging points near the transmission path through a multi-sensor monitoring system to improve imaging performance, wherein each excitation corresponds to (Ns-1) transmission paths for a monitoring system with Ns sensors; acquiring two imaging characteristics, namely a Damage Index (DI) and a time-dependent imaging weight, wherein the damage index can be obtained by evaluating a test signal through a twin network, and the time-dependent imaging weight can be obtained by thresholding a time domain signal to obtain a corresponding time parameter; the damage probability of a discretized plate can be calculated from these two features:
wherein ,Pr(x, y) is the estimated damage probability of the r-th path, NPNumber of transmission paths for all sensors, WrAnd (x, y) is the linear attenuation imaging weight, and x and y are two-dimensional coordinates of pixel points in the imaging picture.
Furthermore, the convolution coding network and the twin network both adopt convolution network structures; wherein the convolutional coding network is used to reduce information redundancy and try to find a reasonable way to model the abnormal signal; the abnormal signal that needs to be simulated can show a difference from the normal signal, but need not be exactly similar to the real lesion signal.
Further, the encoder of the convolutional coding network has a plurality of convolutional layers, the input length of the input layer of the encoder selects the size of the pooling layer closest to twice the head wave length, and the input length adopts 2 exponential power; kernel size selects the odd kernel size closest to one quarter of the sine wave period; the slope parameter of the Leaky ReLU function is set to 0.2, which is the default setting for the function.
Further, the training of the convolutional coding network adopts a mean square error function; the twin network training adopts binary cross entropy; all networks use an optimizer with a learning rate of 0.001.
Compared with the prior art, the beneficial results of the invention comprise:
the invention can solve the problems of feature selection and multi-sensor imbalance and visually present the damage. According to the practical requirements of monitoring application, the invention aims to train a monitoring model only by using monitoring signals collected under normal conditions, and effectively identify the abnormal state of the structure. Experiments on the composite material plate show that the damage index of the structure can be effectively obtained, and damage imaging characterization is realized by combining the time coefficient.
Drawings
FIG. 1 is a flow chart of an ultrasonic guided wave semi-supervised imaging detection method in an embodiment of the present invention;
FIG. 2 is a schematic diagram of an experimental platform in an embodiment of the present invention, wherein FIG. 2a is a schematic diagram of a sensor installation position, and FIG. 2b is a schematic diagram of wave velocity-direction characteristics of an example material;
FIG. 3 is a schematic diagram of a convolutional self-coding network structure in an embodiment of the present invention;
FIG. 4 is a schematic diagram of a twin network in an embodiment of the invention;
FIG. 5 is a schematic diagram of semi-supervised imaging features extracted in an embodiment of the present invention;
fig. 6 is a schematic diagram of a semi-supervised imaging detection result in an embodiment of the present invention, where fig. 6a is a schematic diagram of an original imaging detection result, and fig. 6b is a schematic diagram of a 90% threshold damage result.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the following embodiments and accompanying drawings.
Example (b):
an ultrasonic guided wave semi-supervised imaging detection method is shown in fig. 1, and comprises the following steps:
s1, standardizing signals collected from different sensors to ensure the consistency of training samples;
in this embodiment, in order to input a standard signal, 5% of the maximum value is set as a threshold value for signal capture, and for example, the damage detection of the carbon fiber composite board is adopted, and the transducers are all circular PZT with the diameter of 10 × 1 mm; the threshold interception position is aligned with the nearest channel, so that more complete wave packets are intercepted; all samples used for training and testing are intercepted in the same mode, and then normalization processing is carried out on each training sample, so that the difference between different sensors caused by a deep neural network is avoided;
in this embodiment, the excitation signal is determined as a 5-cycle sinusoidal signal at 200 KHz. The sampling rate was set to 10MHz and the sampling length was 5K. Based on the detected object and the sensor arrangement as shown in fig. 2a, the ultrasound propagation velocities in different directions in the monitoring structure are shown in fig. 2 b. 16 PZT sensors are mounted in a 400mm diameter ring at 22.5 degree intervals. The PZT-1 sensor is used as an excitation sensor, and other sensors are used as receiving sensors. Five damage grade conditions of 5mm, 10mm, 15mm and 20mm are adopted for data acquisition. In each case, 30 samples from 15 PZT sensors were obtained in the dataset.
Signal normalization means that all signal samples used for training and testing are truncated as:
Si＝min[where(Xi>Xmax×Tmax)]； (1)
Si＝Si-min(Si)； (2)
wherein XiFor signal samples received by the i-th sensor, XmaxIs the maximum value of the training set, SiThe starting point of the intercepted signal is; each signal sample will be normalized to [ -1,1 ] according to the following equation]：
wherein XjFor the jth element, N, in the training samplefFor the normalization factor, max (x) and min (x) are the maximum and minimum values of the training samples, respectively.
S2, obtaining dimensionality reduction representation by using a convolutional coding network, and ensuring dimensionality reduction quality through reconstruction errors;
the convolution neural network extracts hidden layer representation of signals through convolution and pooling, realizes signal reconstruction through up-sampling and deconvolution, and enables convolution kernel to be m in size1Set to 3, convolution kernel n1The number is set to 8; the specific convolution calculation is activated by using a Leaky ReLU function which can generate non-zero slope so as to keep the invariance of the waveform characteristics:
wherein ,Yl kTo an activation value, Xl kIs a weighted sum of the ith neurons of the kth layer, PreluIs the slope parameter of the activation function.
As shown in fig. 3, in the present embodiment, the convolutional coding network adopts a convolutional network structure. The convolutional coding network is a one-dimensional convolutional coding network, and the encoder mainly comprises five convolutional layers, five pooling layers and a flat layer. The decoder is constructed symmetrically to the encoder, reconstructing the signal by deconvolution. The encoder of the convolutional encoding network has a plurality of convolutional layers, and the input length of the input layer of the encoder is set to 1024; kernel size was set to 13, selecting the odd kernel size closest to one quarter of the sine wave period; the slope parameter of the Leaky ReLU function is set to be 0.2, and the slope parameter is the default setting of the function; the convolutional coding network is trained by using a mean square error function and using an optimizer with a learning rate of 0.001.
S3, adding signal interference in the dimensionality reduction representation of the normal signal, and reconstructing a simulated damage signal;
disturbance in the hidden layer can change the reconstructed signal, so that a damage state signal is simulated; the number of convolution kernels of the last convolution layer in the convolutional coding network is set to 1, and the hidden representation of the layer is RHIs a vector; given a magnification factor RdThe ratio in the hidden representation is RdPart of the maximum values of (A) will be added with random disturbance, which is obtained by normal distribution of random numbers each time the random disturbance is added, and the random disturbance DaObey a normal distribution of N (0, 1); in order to ensure the proportion of the disturbance, when the disturbance D is randomaIs more than [0.1,1 ]]Will carry on the value again to it during interval scope:
wherein ,
for hidden representation, the ratio is RdThe qth hidden representation element of the partial maximum of (a).
S4, constructing a twin network by using a semi-supervised learning framework, training the twin network by using an abnormal signal, and comparing a normal state signal and a test signal in the twin network to obtain a damage index of the twin network;
the twin network comprises two convolution neural networks with shared parameters, the input signal pair is mapped to a high-dimensional space through the same convolution network, the signal difference of the signal in the high-dimensional space is obtained, and the structural state is identified according to the mapping distance represented by the high-dimensional space, and the method specifically comprises the following steps:
for input signal pair [ X ] composed of normal signal and test signaln,Xt]Mapping function M implemented by convolutional neural network in twin networkw(X) mapping the input signal to a high dimensional space; then mapping the signal Mw(Xn) and Mw(Xt) The euclidean distance E of (a) is calculated as the output of the twin network:
E(Mw(Xn),Mw(Xt))＝||Mw(Xn),Mw(Xt)||。 (4)
semi-supervised training is realized in the training process of the twin network, namely only signals in a normal state are acquired, and real damage signals do not need to be acquired; using normal state input pairs [ X ]n1,Xn2]And simulated damage input pair [ X ]n,Xs]Network training is carried out and network parameters are updated; setting the label of the normal input pair as 0, setting the label of the simulated damage pair as 1, wherein the number of the labels of the normal input pair is the same as that of the simulated damage pair; binary cross entropy was used as a loss function:
wherein 
Is a label of the input signal pair, YkIs the output of the twin network, n is the batch size in the training process, Xn1,Xn2For any two of the normal signals, XnFor simulating normal state signal samples in the lesion input pairs, XsIs a sample of a signal that simulates the state of damage.
As shown in fig. 4, in this embodiment, the twin network adopts a convolutional network structure, the twin network adopts 4 convolutional layers, and finally, the values of the normal and abnormal state signals in the dense layer are compared to obtain the euclidean distance. The input length of the twin network is set to 1024 and the input signal will contain several modes of the supervisory signal. Twin network training employs a binary cross entropy, using an adam optimizer with a learning rate of 0.001.
And S5, visually describing the damage state by using a probability imaging technology.
The detection signals of different sensors are standardized, and the damage index of the detection signal of each sensor can be predicted through the twin network, so that the structural state is evaluated; carrying out probability imaging by using damage indexes of different sensors and displaying a structural state; the probability imaging technology is specifically as follows:
the imaging performance is improved by assigning a damage probability to imaging points near the transmission path by a multi-sensor monitoring system, for which there are Ns sensors, each excitation corresponds to (Ns-1) transmission paths. With the multi-sensor monitoring system shown in fig. 2a, there are 15 imaging paths, and two characteristics of imaging, namely, a Damage Index (DI) and a time-dependent imaging weight, are obtained, the damage index can be obtained by evaluating a test signal through a twin network, and the time-dependent imaging weight can be obtained by thresholding a time domain signal to obtain a corresponding time parameter; the damage probability of a discretized plate can be calculated from these two features:
wherein ,Pr(x, y) is the estimated damage probability of the r-th path, NPNumber of transmission paths for all sensors, WrAnd (x, y) is the linear attenuation imaging weight, and x and y are two-dimensional coordinates of pixel points in the imaging picture.
In this embodiment, the structure of the composite plate is relatively complex, and the wave velocity is different in some propagation directions from others. In addition, the multilayer structure brings more guided wave modes, and the wave overlapping is more serious. In order to acquire the structural state image, a multi-sensor monitoring system needs to be designed, as shown in fig. 2a, to collect more information. The damage index of the actual damage signal of the composite material plate is shown in fig. 5, and the damage indexes of the fourth sensor pzt4, the ninth sensor pzt9, the tenth sensor pzt10 and the fourteenth sensor pzt14 are obviously larger than those of other sensors. Thus, damage is more likely to occur on the transmission paths associated with these sensors. The damage index also increases with increasing degree of damage, which fully accounts for the damage recognition capability of the twin network.
By calculating the damage index and the time coefficient and setting the scale coefficient as 10, a damage image of the carbon fiber composite plate structure can be obtained, and the position with the highest total damage probability is the position where damage is most likely to occur. The original image of 20mm damage is shown in fig. 6a, and the 90% threshold damage image is shown in fig. 6b, and it can be seen from the damage image that the imaging shape after velocity correction is closer to a rectangle, which is caused by the difference of the composite plate guided wave velocity. The damage threshold was set to 90% of the maximum damage value, leaving only the area in fig. 6b that is greater than the damage threshold, and the damage could be clearly shown, demonstrating that the present invention is able to detect damage in composite materials well, providing an effective maintenance strategy.
It should be noted that the above-mentioned embodiment is only an example of the present invention, and is not intended to limit the scope of the present invention. Workers skilled in the art will recognize that changes may be made in form and detail without departing from the spirit and scope of the disclosure. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and some equivalent modifications and variations of the present invention should be covered by the protection scope of the claims of the present invention. Furthermore, although specific terms are employed herein, they are used in a generic and descriptive sense only and not for purposes of limitation.
Claims (10)
1. An ultrasonic guided wave semi-supervised imaging detection method is characterized by comprising the following steps:
s1, standardizing signals collected from different sensors to ensure the consistency of training samples;
s2, obtaining dimensionality reduction representation by using a convolutional coding network, and ensuring dimensionality reduction quality through reconstruction errors;
s3, adding signal interference in the dimensionality reduction representation of the normal signal, and reconstructing a simulated damage signal;
s4, constructing a twin network by using a semi-supervised learning framework, training the twin network by using an abnormal signal, and comparing a normal state signal and a test signal in the twin network to obtain a damage index of the twin network;
and S5, visually describing the damage state by using a probability imaging technology.
2. The method as claimed in claim 1, wherein in step S1, the maximum value T is used to normalize the input signalmax% set as threshold for signal acquisition; the threshold interception position is aligned with the nearest channel, so that more complete wave packets are intercepted; all samples used for training and testing are intercepted in the same mode, and then normalization processing is carried out on each training sample, so that the difference between different sensors caused by a deep neural network is avoided;
signal normalization means that all signal samples used for training and testing are truncated as:
Si＝min[where(Xi>Xmax×Tmax)]； (1)
Si＝Si-min(Si)； (2)
wherein XiFor signal samples received by the ith sensor，XmaxIs the maximum value of the training set, SiThe starting point of the intercepted signal is; each signal sample will be normalized to [ -1,1 ] according to the following equation]：
wherein XjFor the jth element, N, in the training samplefFor the normalization factor, max (x) and min (x) are the maximum and minimum values of the training samples, respectively.
3. The ultrasonic guided wave semi-supervised imaging detection method according to claim 2, characterized in that: in step S2, the convolutional encoding network extracts the hidden layer representation of the signal by convolution and pooling, performs signal reconstruction by upsampling and deconvolution, and sets the convolution kernel size to m1Setting the number of convolution kernels to n1(ii) a The specific convolution calculation is activated by using a Leaky ReLU function which can generate non-zero slope so as to keep the invariance of the waveform characteristics:
wherein ,Yl kTo an activation value, Xl kIs a weighted sum of the ith neurons of the kth layer, PreluIs the slope parameter of the activation function.
4. The ultrasonic guided wave semi-supervised imaging detection method according to claim 3, characterized in that: in step S3, the disturbance in the hidden layer may change the reconstructed signal, thereby simulating the damage state signal; the number of convolution kernels of the last convolution layer in the convolutional coding network is set to 1, and the hidden representation of the layer is RHIs a vector; given a magnification factor RdThe ratio in the hidden representation is RdThe partial maximum value of (A) will add random disturbance, and each time the random disturbance is added, the random disturbance is added by normal distributionMachine number acquisition, random disturbance DaObey a normal distribution of N (0, 1); in order to ensure the proportion of the disturbance, when the disturbance D is randomaIs more than [0.1,1 ]]Will carry on the value again to it during interval scope:
wherein ,
for hidden representation, the ratio is RdThe qth hidden representation element of the partial maximum of (a).
5. The ultrasonic guided wave semi-supervised imaging detection method according to claim 4, wherein in step S4, the twin network includes two parameter-shared convolutional neural networks, the input signal pair is mapped to a high-dimensional space through the same convolutional network, the signal difference of the signal in the high-dimensional space is obtained, and the structural state is identified according to the mapping distance represented by the high-dimensional space, which is as follows:
for input signal pair [ X ] composed of normal signal and test signaln,Xt]Mapping function M implemented by convolutional neural network in twin networkw(X) mapping the input signal to a high dimensional space; then mapping the signal Mw(Xn) and Mw(Xt) The euclidean distance E of (a) is calculated as the output of the twin network:
E(Mw(Xn),Mw(Xt))＝||Mw(Xn),Mw(Xt)|| (6)。
6. the ultrasonic guided wave semi-supervised imaging detection method of claim 5, wherein semi-supervised training is realized in the training process of the twin network, namely only signals in a normal state are acquired without acquiring real damage signals; using normal state input pairs [ X ]n1,Xn2]And simulated damage input pair [ X ]n,Xs]Network training is carried out and network parameters are updated; setting the label of the normal input pair as 0, setting the label of the simulated damage pair as 1, wherein the number of the labels of the normal input pair is the same as that of the simulated damage pair; binary cross entropy was used as a loss function:
wherein 
Is a label of the input signal pair, YkIs the output of the twin network, n is the batch size in the training process, Xn1,Xn2For any two of the normal signals, XnFor simulating normal state signal samples in the lesion input pairs, XsIs a sample of a signal that simulates the state of damage.
7. The method according to claim 1, wherein in step S5, the detection signals of different sensors are standardized, and the damage indicators of the detection signals of the sensors can be predicted through the twin network, so as to evaluate the structural status; carrying out probability imaging by using damage indexes of different sensors and displaying a structural state; the probability imaging technology is specifically as follows:
assigning damage probability to imaging points near the transmission path through a multi-sensor monitoring system to improve imaging performance, wherein each excitation corresponds to (Ns-1) transmission paths for a monitoring system with Ns sensors; acquiring two imaging characteristics, namely a Damage Index (DI) and a time-dependent imaging weight, wherein the damage index can be obtained by evaluating a test signal through a twin network, and the time-dependent imaging weight can be obtained by thresholding a time domain signal to obtain a corresponding time parameter; the damage probability of a discretized plate can be calculated from these two features:
wherein ,Pr(x, y) is the estimated damage probability of the r-th path, NPNumber of transmission paths for all sensors, WrAnd (x, y) is the linear attenuation imaging weight, and x and y are two-dimensional coordinates of pixel points in the imaging picture.
8. The ultrasonic guided wave semi-supervised imaging detection method according to claim 1, characterized in that:
the convolutional coding network and the twin network both adopt convolutional network structures; wherein the convolutional coding network is used to reduce information redundancy and try to find a reasonable way to model the abnormal signal; the abnormal signal that needs to be simulated can show a difference from the normal signal, but need not be exactly similar to the real lesion signal.
9. The ultrasonic guided wave semi-supervised imaging detection method according to claim 1, characterized in that: the encoder of the convolutional coding network is provided with a plurality of convolutional layers, the input length of the input layer of the encoder selects the size of a pooling layer closest to twice of the head wave length, and the input length adopts 2 exponential power; kernel size selects the odd kernel size closest to one quarter of the sine wave period; the slope parameter of the Leaky ReLU function is set to 0.2, which is the default setting for the function.
10. The ultrasonic guided wave semi-supervised imaging detection method according to any one of claims 1 to 9, characterized in that: training a convolutional coding network by adopting a mean square error function; the twin network training adopts binary cross entropy; all networks use an optimizer with a learning rate of 0.001.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011642474.8A CN112668527B (en) | 2020-12-31 | 2020-12-31 | Ultrasonic guided wave semi-supervised imaging detection method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011642474.8A CN112668527B (en) | 2020-12-31 | 2020-12-31 | Ultrasonic guided wave semi-supervised imaging detection method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112668527A true CN112668527A (en) | 2021-04-16 |
| CN112668527B CN112668527B (en) | 2023-04-25 |
Family
ID=75412342
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011642474.8A Active CN112668527B (en) | 2020-12-31 | 2020-12-31 | Ultrasonic guided wave semi-supervised imaging detection method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112668527B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113848252A (en) * | 2021-09-28 | 2021-12-28 | 天津大学 | Corrosion imaging method based on ultrasonic guided wave and convolutional neural network |
| CN114199992A (en) * | 2021-12-02 | 2022-03-18 | 国家石油天然气管网集团有限公司 | Oil storage tank wall corrosion detection method and system |
| CN114414659A (en) * | 2022-01-21 | 2022-04-29 | 山东大学 | Non-linear ultrasonic guided wave non-parametric damage identification method and system based on frequency fusion |
| CN115128166A (en) * | 2022-06-17 | 2022-09-30 | 广东电网有限责任公司广州供电局 | Cable aluminum sheath corrosion damage imaging method and device based on twin network and ultrasonic guided wave |
| CN116934758A (en) * | 2023-09-18 | 2023-10-24 | 南通华隆微电子股份有限公司 | Semiconductor detection method and system based on convolution model |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109523525A (en) * | 2018-11-07 | 2019-03-26 | 广州大学 | Malign lung nodules recognition methods, device, equipment and the storage medium of image co-registration |
| CN109885482A (en) * | 2019-01-16 | 2019-06-14 | 重庆大学 | Software Defects Predict Methods based on the study of few sample data |
| CN110567558A (en) * | 2019-08-28 | 2019-12-13 | 华南理工大学 | Ultrasonic guided wave detection method based on deep convolution characteristics |
| WO2020156348A1 (en) * | 2019-01-31 | 2020-08-06 | 青岛理工大学 | Structural damage identification method based on ensemble empirical mode decomposition and convolution neural network |
-
2020
- 2020-12-31 CN CN202011642474.8A patent/CN112668527B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109523525A (en) * | 2018-11-07 | 2019-03-26 | 广州大学 | Malign lung nodules recognition methods, device, equipment and the storage medium of image co-registration |
| CN109885482A (en) * | 2019-01-16 | 2019-06-14 | 重庆大学 | Software Defects Predict Methods based on the study of few sample data |
| WO2020156348A1 (en) * | 2019-01-31 | 2020-08-06 | 青岛理工大学 | Structural damage identification method based on ensemble empirical mode decomposition and convolution neural network |
| CN110567558A (en) * | 2019-08-28 | 2019-12-13 | 华南理工大学 | Ultrasonic guided wave detection method based on deep convolution characteristics |
Non-Patent Citations (1)
| Title |
|---|
| YUAN LIU .ET.AL: "A novel velocity anisotropy probability imaging method using ultrasonic guided waves for composite plates", 《MEASUREMENT》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113848252A (en) * | 2021-09-28 | 2021-12-28 | 天津大学 | Corrosion imaging method based on ultrasonic guided wave and convolutional neural network |
| CN113848252B (en) * | 2021-09-28 | 2022-06-24 | 天津大学 | Corrosion imaging method based on ultrasonic guided wave and convolutional neural network |
| CN114199992A (en) * | 2021-12-02 | 2022-03-18 | 国家石油天然气管网集团有限公司 | Oil storage tank wall corrosion detection method and system |
| CN114199992B (en) * | 2021-12-02 | 2024-04-05 | 国家石油天然气管网集团有限公司 | Method and system for detecting corrosion of tank wall of oil storage tank |
| CN114414659A (en) * | 2022-01-21 | 2022-04-29 | 山东大学 | Non-linear ultrasonic guided wave non-parametric damage identification method and system based on frequency fusion |
| CN114414659B (en) * | 2022-01-21 | 2023-12-29 | 山东大学 | Nonlinear ultrasonic guided wave parameter-free damage identification method and system based on frequency fusion |
| CN115128166A (en) * | 2022-06-17 | 2022-09-30 | 广东电网有限责任公司广州供电局 | Cable aluminum sheath corrosion damage imaging method and device based on twin network and ultrasonic guided wave |
| CN115128166B (en) * | 2022-06-17 | 2024-07-12 | 广东电网有限责任公司广州供电局 | Cable aluminum sheath corrosion damage imaging method and device based on twin network and ultrasonic guided wave |
| CN116934758A (en) * | 2023-09-18 | 2023-10-24 | 南通华隆微电子股份有限公司 | Semiconductor detection method and system based on convolution model |
| CN116934758B (en) * | 2023-09-18 | 2023-11-24 | 南通华隆微电子股份有限公司 | Semiconductor detection method and system based on convolution model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112668527B (en) | 2023-04-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112668527A (en) | Ultrasonic guided wave semi-supervised imaging detection method | |
| CN112232400B (en) | Stainless steel weld ultrasonic defect detection method based on depth feature fusion | |
| Meng et al. | Ultrasonic signal classification and imaging system for composite materials via deep convolutional neural networks | |
| Oliveira et al. | Ultrasound-based identification of damage in wind turbine blades using novelty detection | |
| CN108732465B (en) | Power distribution network fault positioning method based on wavelet transformation and CNN | |
| KR102209401B1 (en) | Apparatus and method for fault diagnosis of gearbox using cnn | |
| CN109993105A (en) | A kind of improved adaptive sparse sampling Fault Classification | |
| CN106841403A (en) | A kind of acoustics glass defect detection method based on neutral net | |
| Wang et al. | Prediction model of natural gas pipeline crack evolution based on optimized DCNN-LSTM | |
| CN114445386B (en) | PVC pipe quality detection and evaluation method and system based on artificial intelligence | |
| CN109649432B (en) | System and method for monitoring integrity of steel rail of cloud platform based on guided wave technology | |
| CN102800070B (en) | Multi-modality image fusion method based on region and human eye contrast sensitivity characteristic | |
| CN111695465B (en) | Pipe network fault diagnosis and positioning method and system based on pressure wave mode identification | |
| CN112946081B (en) | Ultrasonic imaging method based on intelligent extraction and fusion of defect multiple features | |
| CN117540174B (en) | Building structure multi-source heterogeneous data intelligent analysis system and method based on neural network | |
| CN115824519A (en) | Valve leakage fault comprehensive diagnosis method based on multi-sensor information fusion | |
| JP7492567B2 (en) | Cable aluminum sheath corrosion damage imaging method and device based on twin network and ultrasonic guided wave | |
| CN110702792A (en) | Alloy tissue ultrasonic detection classification method based on deep learning | |
| Acciani et al. | Angular and axial evaluation of superficial defects on non-accessible pipes by wavelet transform and neural network-based classification | |
| CN117764980A (en) | Automatic identification and measurement method for defects of composite material based on infrared multi-feature fusion | |
| Yu et al. | Intelligent damage classification for tensile membrane structure based on continuous wavelet transform and improved ResNet50 | |
| CN115235877A (en) | Corrosion state monitoring system and stress corrosion state evaluation method and device | |
| CN110033846B (en) | Obstetrical clinical auxiliary sequential labor device, control method and computer program | |
| CN116773678A (en) | Intelligent maintenance method and system for subway rail | |
| CN116541771A (en) | Unbalanced sample bearing fault diagnosis method based on multi-scale feature fusion |
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 |