Metal structure defect multi-source heterogeneous nondestructive testing signal composite characterization method based on deep neural network and attention mechanism

Abstract

The invention discloses a metal structure defect multi-source heterogeneous nondestructive testing signal composite characterization method based on a deep neural network and an attention mechanism. According to the method, a deep neural network model is constructed, and data fusion and defect composite characterization are carried out on multi-source heterogeneous signals obtained by three nondestructive testing technologies of magnetic flux leakage testing, electromagnetic ultrasonic testing and electromagnetic guided wave testing in combination with an attention mechanism. Firstly, through advanced signal processing and data fusion technologies, the data processing and analysis process is simplified, effective processing and interpretation of complex data are achieved, and the detection efficiency and reliability are improved; and then more accurate and comprehensive characterization of the size, type, shape and position of the metal structure defect is realized by comprehensively utilizing information of different signal sources, and innovative technical support is provided for nondestructive testing in the industrial field.

Description

Technical Field

[0001] The present invention relates to the field of nondestructive testing, and in particular to a composite characterization method for multi-source heterogeneous nondestructive testing signals of metal structure defects based on a deep neural network and an attention mechanism. Background Art

[0002] With the widespread application of metal structures in modern industry, non-destructive testing of complex metal structures has become an important means to ensure the safety and stability of structures. Magnetic flux leakage, electromagnetic ultrasound, and electromagnetic guided waves are commonly used non-destructive testing technologies, each with its own advantages in metal structure defect detection. Magnetic flux leakage technology identifies surface and near-surface cracks or corrosion defects by detecting changes in the magnetic field on the metal surface, and has high sensitivity; electromagnetic ultrasound combines the characteristics of electromagnetic fields and ultrasonic waves to accurately locate internal defects over a large range; and electromagnetic guided waves, due to their long-distance propagation capabilities, can effectively detect deep defects in metal structures, especially in large-scale structures. However, due to differences in signal transmission characteristics, defect types, and perception depth among these detection technologies, each one cannot fully and accurately characterize the various defects that may exist in complex metal structures when used alone.

[0003] Traditional nondestructive testing methods typically rely on a single signal source or detection technology, making it difficult to fully capture diverse defect characteristics. This is particularly true in complex metal structures, where deep or minute defects are often inaccurately located. Therefore, efficiently integrating heterogeneous signals from three nondestructive testing technologies—magnetic flux leakage, electromagnetic ultrasound, and electromagnetic guided waves—to achieve more comprehensive and accurate defect characterization has become a major challenge in the field.

[0004] In recent years, the combination of deep neural networks and attention mechanisms has provided new solutions for non-destructive testing. Deep neural networks can automatically extract deep features from complex multi-source heterogeneous signals, while attention mechanisms can dynamically adjust their weights according to the importance of different parts of the signal, thereby focusing on the defect area and improving the accuracy and robustness of defect identification. By performing deep learning and composite representation on the signals of magnetic flux leakage, electromagnetic ultrasound, and electromagnetic guided waves, the limitations of a single technology can be overcome, and all-round perception and precise positioning of defects in complex metal structures can be achieved. However, there is currently a lack of a composite representation method that can simultaneously process the signals of three technical technologies: magnetic flux leakage, electromagnetic ultrasound, and electromagnetic guided waves. In particular, in the defect detection of complex metal structures, how to effectively fuse and extract the complementary information of different signal sources to improve the accuracy and reliability of defect identification remains a technical problem that needs to be solved urgently.

[0005] Therefore, it is of great significance to study a composite characterization method for multi-source heterogeneous non-destructive testing signals of metal structure defects based on deep neural networks and attention mechanism. Summary of the Invention

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: A composite characterization method for multi-source heterogeneous nondestructive testing signals of metal structure defects based on deep neural networks and attention mechanisms includes the following steps: (1) Make standard metal structure defect samples for standard data collection experiments; (2) Use magnetic flux leakage, electromagnetic ultrasonic and guided wave sensors to detect the same metal structure defects of the standard plate, and collect no less than 1,000 data samples; (3) The data set is divided into a training set and a test set; the output data of the magnetic flux leakage, electromagnetic ultrasound, and electromagnetic guided wave sensors are respectively X L ，X EMUS ，X EMAT The output data is used as the input layer of the neural network. The convolutional neural network (CNN) is used to extract features from the input data of each sensor. (4) Aggregate the extracted features into a shared hidden layer and use convolutional neural networks to learn the nonlinear relationship between signals; (5) Adjust the features of each sensor by calculating the attention weight of each sensor data, and the weighted representation of each sensor feature; (6) The trained feature vector Z is obtained by weighted average;

[0007] (7) The feature Z processed by the attention mechanism is input into the subsequent layers of the network for processing; the model outputs defect data through regression and classification tasks respectively.

[0008] Preferably, the feature extraction formula in step (3) is:

[0009] in , , is the eigenvalue of each sensor data after being processed by the convolution layer.

[0010] Preferably, the learning formula in step (4) is:

[0011] in, is the weight matrix of the shared layer, is the bias term, is a nonlinear activation function (such as ReLU or Sigmoid); Preferably, in step (5), the weighted expression formula is: .

[0012] Preferably, in step (6), the weighted average formula is:

[0013] in, , , They are respectively represented as weight coefficients in the attention mechanism.

[0014] Furthermore, the defect data includes the size, location, and type of the defect; Furthermore, defect data includes defect size regression, defect location regression, defect shape regression, and defect type classification; Defect size regression: The defect size is predicted through the regression layer. The defect size regression formula is:

[0015] Defect location regression: The coordinate position of the defect is predicted through the regression layer. The formula for defect location regression is:

[0016] Defect shape regression: The shape of the defect is predicted through the classification or regression layer. The formula for defect shape regression is:

[0017] Defect type classification: The defect type is predicted through the classification layer. The defect type classification formula is: .

[0018] Preferably, due to multi-task learning, the loss function needs to consider the losses of multiple tasks at the same time. The loss function of each task is as follows: Defect size regression loss: Mean Squared Error (MSE):

[0019] Defect location regression loss: Mean Squared Error (MSE):

[0020] Defect shape classification loss: cross entropy loss

[0021] Defect type classification loss: Cross entropy loss:

[0022] Ultimately, the total loss function is the weighted sum of the losses of each task:

[0023] in, , , , is the weight coefficient of each task.

[0024] Due to the adoption of the above technical solution, the present invention has the following technical advancements compared to the prior art: 1. Improved detection accuracy: By comprehensively utilizing information from multi-source heterogeneous signals, this method can improve the accuracy and precision of metal defect detection. Compared with traditional methods, it can more effectively identify and locate various defects in metal components, reducing the possibility of missed detection and misjudgment.

[0025] 2. Comprehensive Characterization of Metal Defects: This method enables comprehensive characterization of the internal structure and defects of metal materials. By comprehensively analyzing data from different signal sources, it can provide more comprehensive and detailed defect information, helping to comprehensively assess the health of metal components.

[0026] 3. Improved detection efficiency: This method simplifies the data processing and analysis process. Through advanced signal processing and data fusion technology, it can improve detection efficiency and reliability. The optimized data processing process helps to quickly and accurately identify metal defects, improving production efficiency and quality control.

[0027] 4. Innovation and Practicality: This method's innovation lies in its comprehensive utilization of multi-source heterogeneous electromagnetic acoustic signals, combined with advanced data analysis techniques, bringing a new solution to the field of metal defect detection. Its practicality lies in its realization of a more reliable and comprehensive metal defect detection method, which is expected to have significant applications and impact in industrial production and materials science. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0029] Figure 1 Schematic diagram of the composite characterization method for multi-source heterogeneous non-destructive testing signals of metal structure defects based on deep neural networks and attention mechanisms Figure 2 This is a structural diagram of the Q235 steel plate inspection object. DETAILED DESCRIPTION

[0030] The following examples may help those skilled in the art to more fully understand the present invention, but shall not limit the present invention in any way. Example

[0031] A composite characterization method for multi-source heterogeneous nondestructive testing signals of metal structure defects based on deep neural networks and attention mechanisms includes the following steps: (1) Make standard metal structure defect samples for standard data collection experiments; (2) Use magnetic flux leakage, electromagnetic ultrasonic and guided wave sensors to detect the same metal structure defects of the standard plate, and collect no less than 1,000 data samples; (3) The data set is divided into a training set and a test set; the output data of the magnetic flux leakage, electromagnetic ultrasound, and electromagnetic guided wave sensors are respectively X L ，X EMUS ，X EMAT The output data is used as the input layer of the neural network. The convolutional neural network (CNN) is used to extract features from the input data of each sensor. The feature extraction formula is:

[0032] in , , is the eigenvalue of each sensor data after being processed by the convolution layer; (4) The extracted features are aggregated into a shared hidden layer, and the nonlinear relationship between the signals is learned using a convolutional neural network; the learning formula is:

[0033] in, is the weight matrix of the shared layer, is the bias term, is a nonlinear activation function (such as ReLU or Sigmoid); (5) The features of each sensor are adjusted by calculating the attention weight of each sensor data. The weighted representation of each sensor feature is as follows:

[0034] (6) The trained feature vector Z is obtained by weighted average. The weighted average formula is:

[0035] in, , , They are respectively represented as weight coefficients in the attention mechanism.

[0036] (7) The feature Z processed by the attention mechanism is input into the subsequent layers of the network for processing; finally, the model outputs the size, location and type of the defect through regression and classification tasks respectively: Defect size regression: Predict the size of the defect through the regression layer:

[0037] Defect location regression: The coordinate location of the defect is predicted through the regression layer:

[0038] Defect shape regression: Predict the shape of the defect (e.g. rectangle, ellipse, etc.) through a classification or regression layer:

[0039] Defect type classification: Predict the type of defect (such as fatigue crack, corrosion, etc.) through the classification layer: .

[0040] Because it is multi-task learning, the loss function needs to consider the losses of multiple tasks at the same time. The loss function of each task is as follows: Defect size regression loss: Mean Squared Error (MSE):

[0041] Defect location regression loss: Mean Squared Error (MSE):

[0042] Defect shape classification loss: cross entropy loss

[0043] Defect type classification loss: Cross entropy loss:

[0044] Ultimately, the total loss function is the weighted sum of the losses of each task:

[0045] in, , , , is the weight coefficient of each task.

[0046] Example 1 First, a standard metal structure defect sample is made for standard data collection experiments, such as Figure 2The following is an example of a standard sample, using Q235 steel plate as the test object. Figure 2 As shown, its dimensions are 1000x500x20mm, and the radii of the defect holes to be detected are ø50x8mm, ø70x8mm and ø100x12mm, respectively. The distances between the defect holes from left to right are 400mm and 300mm, respectively.

[0047] Then, the same metal structure defect of the standard plate is detected using leakage magnetic field, electromagnetic ultrasonic and guided wave sensors respectively, and a data sample of no less than 1,000 pieces is collected.

[0048] according to Figure 1 The requirement is to divide the data set into 80% training set and 20% test set. Figure 1 As shown, the output data of magnetic flux leakage, electromagnetic ultrasound, and electromagnetic guided wave sensors are respectively used as X L ，X EMUS ，X EMAT , and use these data as the input layer of the neural network. For the input data of each sensor, a convolutional neural network (CNN) is used to extract features:

[0049] in , , is the eigenvalue of each sensor data after being processed by the convolution layer.

[0050] The extracted features are aggregated into a shared hidden layer, and a convolutional neural network is used to learn the nonlinear relationship between the signals.

[0051]

[0052] in, is the weight matrix of the shared layer, is the bias term, is a non-linear activation function (such as ReLU or Sigmoid).

[0053] The features of each sensor are adjusted by calculating the attention weight of each sensor data. The weighted representation of each sensor feature is:

[0054] The final trained feature vector Z is obtained by weighted average:

[0055] in, , , They are respectively represented as weight coefficients in the attention mechanism.

[0056] The feature Z processed by the attention mechanism is input into the subsequent layers of the network for processing; finally, the model outputs the size, location, and type of the defect through regression and classification tasks respectively: Defect size regression: Predict the size of the defect through the regression layer:

[0057] Defect location regression: The coordinate location of the defect is predicted through the regression layer:

[0058] Defect shape regression: Predict the shape of the defect (e.g. rectangle, ellipse, etc.) through a classification or regression layer:

[0059] Defect type classification: Predict the type of defect (such as fatigue crack, corrosion, etc.) through the classification layer:

[0060] Because it is multi-task learning, the loss function needs to consider the losses of multiple tasks at the same time. The loss function of each task is as follows: Defect size regression loss: Mean Squared Error (MSE):

[0061] Defect location regression loss: Mean Squared Error (MSE):

[0062] Defect shape classification loss: cross entropy loss

[0063] Defect type classification loss: Cross entropy loss:

[0064] Ultimately, the total loss function is the weighted sum of the losses of each task:

[0065] in, , , , is the weight coefficient of each task.

[0066] While the present invention has been generally described above, modifications and improvements are readily apparent to those skilled in the art. Therefore, modifications and improvements that do not depart from the spirit of the present invention are intended to be within the scope of the present invention.

Claims

1. A composite characterization method for multi-source heterogeneous nondestructive testing signals of metal structural defects based on deep neural networks and attention mechanisms, characterized by: The following steps are involved: (1) Make standard metal structure defect samples for standard data collection experiments; (2) Detect the same metal structure defects of the standard plate using magnetic flux leakage, electromagnetic ultrasound, and guided wave sensors, and collect data samples; (3) Divide the data set into a training set and a test set; the output data of the magnetic flux leakage, electromagnetic ultrasound, and electromagnetic guided wave sensors are X L , X EMUS , X EMAT The output data is used as the input layer of the neural network. The convolutional neural network is used to extract features from the input data of each sensor. (4) Aggregate the extracted features into a shared hidden layer and use convolutional neural networks to learn the nonlinear relationship between signals; (5) Adjust the features of each sensor by calculating the attention weight of each sensor data; (6) The trained feature vector Z is obtained by weighted average; (7) The feature Z processed by the attention mechanism is input into the subsequent layers of the network for processing; The model outputs defect data through regression and classification tasks respectively.

2. The composite characterization method for multi-source heterogeneous nondestructive testing signals of metal structure defects based on a deep neural network and an attention mechanism according to claim 1 is characterized by: The feature extraction formula in step (3) is: Among them F L , F EMUS , F EMAT is the eigenvalue of each sensor data after being processed by the convolution layer.

3. The composite characterization method for multi-source heterogeneous nondestructive testing signals of metal structure defects based on a deep neural network and an attention mechanism according to claim 2 is characterized by: The learning formula in step (4) is: in, is the weight matrix of the shared layer, is the bias term, is a non-linear activation function.

4. The composite characterization method for multi-source heterogeneous nondestructive testing signals of metal structure defects based on a deep neural network and an attention mechanism according to claim 3 is characterized by: In step (5), the weighted representation formula is:

5. The composite characterization method for multi-source heterogeneous nondestructive testing signals of metal structure defects based on a deep neural network and an attention mechanism according to claim 4 is characterized by: In step (6), the weighted average formula is: Z=W L F L ′+W EMUS F E ′ MUS +W EMAT F E ′ MAT (4) Among them, W L , W EMUS , W EMAT They are respectively represented as weight coefficients in the attention mechanism.

6. The composite characterization method for multi-source heterogeneous nondestructive testing signals of metal structure defects based on a deep neural network and an attention mechanism according to claim 5, characterized in that: Defect data includes defect size regression, defect location regression, defect shape regression, and defect type classification; Defect size regression: The defect size is predicted through the regression layer. The defect size regression formula is: D size =f size (Z)(5) Defect location regression: The coordinate position of the defect is predicted through the regression layer. The formula for defect location regression is: D position =f position (Z) (6) Defect shape regression: The shape of the defect is predicted through the classification or regression layer. The formula for defect shape regression is: D shape =f shape (Z) (7) Defect type classification: The defect type is predicted through the classification layer. The defect type classification formula is: D type =arg max(f type (Z)) (8)。 7. The composite characterization method for multi-source heterogeneous nondestructive testing signals of metal structure defects based on a deep neural network and an attention mechanism according to claim 1 is characterized by: The data sample size collected in step (2) shall not be less than 1,000.

8. The composite characterization method for multi-source heterogeneous nondestructive testing signals of metal structure defects based on a deep neural network and an attention mechanism according to any one of claims 1 to 7, characterized in that: The loss function for each task is as follows: Defect size regression loss: Mean Squared Error (MSE): Defect location regression loss: Mean Squared Error (MSE): Defect shape classification loss: cross entropy loss Defect type classification loss: Cross entropy loss: Ultimately, the total loss function is the weighted sum of the losses of each task: L total =λ1L size +λ2L position +λ3L shape +λ4L type (13) Among them, λ1, λ2, λ3, λ3 are the weight coefficients of each task.

Images