A method for three-dimensional reconstruction of internal corrosion defects of a rotating electromagnetic field structure

Abstract

The application provides a method for three-dimensional reconstruction of internal corrosion defects of a rotating electromagnetic field structure, which comprises the following steps: firstly, a sample library of internal corrosion defects of the rotating electromagnetic field structure is established, the sample library comprises magnetic field images of multiple frequencies and three-dimensional images of internal corrosion defects; then, a three-dimensional reconstruction network of the internal corrosion defects is established according to physical characteristics of the magnetic field images of multiple frequencies, a loss function embedded with physical information is constructed, the optimal parameters of the network are obtained through training, and finally, the network is deployed to a rotating electromagnetic field detection system to realize three-dimensional reconstruction and accurate evaluation of the internal corrosion defects. The method realizes accurate reconstruction of internal corrosion defects of a structure under a small sample condition through a physical and data dual driving mode, makes up for the problems that internal corrosion defects of a structure cannot be detected and reconstructed by a traditional rotating electromagnetic field, and provides accurate data support for safety evaluation and maintenance decision of the structure.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing signal processing, and in particular to a method for three-dimensional reconstruction of internal corrosion defects in a rotating electromagnetic field structure. Background Technology

[0002] Critical structures in industrial sectors, such as aircraft skin and spent fuel pool cladding, are composed of non-ferromagnetic thin plates. These structures are highly susceptible to corrosion due to corrosive media and alternating stress. Compared to surface corrosion, internal corrosion is more concealed and poses a greater threat. Therefore, regular inspection of internal corrosion and accurate acquisition of its three-dimensional morphology are crucial for safety assessments and maintenance decisions for critical structures.

[0003] Rotating electromagnetic field is a novel electromagnetic nondestructive testing technology that overcomes the directional limitations of traditional AC electromagnetic field and eddy current testing technologies. However, it is mainly applicable to the detection of surface defects, and it faces challenges in detecting and reconstructing internal corrosion defects. Summary of the Invention

[0004] The problem to be solved by this invention is to propose a three-dimensional reconstruction method for internal corrosion defects in rotating electromagnetic field structures. Through a dual-drive approach of physics and data, it achieves accurate reconstruction of internal corrosion defects in structures with small samples, which makes up for the problems of undetectable and difficult reconstruction of internal corrosion defects in traditional rotating electromagnetic field structures, and provides accurate data support for structural safety assessment and maintenance decisions.

[0005] This invention provides a method for three-dimensional reconstruction of surface corrosion defects in a rotating electromagnetic field structure, the steps of which include:

[0006] Step 1: Establish a 3D reconstruction sample library of corrosion defects inside a rotating electromagnetic field structure. The sample library includes magnetic field images at U frequencies and actual 3D reconstruction images of corrosion defects, where U is the number of frequencies of the magnetic field images, not less than 4 and not more than 10. According to the skin effect, the penetration depth of the induced current in the tested structure is related to the frequency; the lower the frequency, the deeper the penetration. Therefore, by combining multiple different frequencies, magnetic field images with different penetration depths can be obtained, thus enabling more accurate acquisition of the 3D morphology information of corrosion defects. The data in the sample library are obtained from experiments and simulations; the higher the frequency, the worse the consistency between simulation and experiment.

[0007] Step 2: Establish a 3D reconstruction network for internal corrosion defects. The network input is a magnetic field image with U frequencies, and the output is a predicted 3D reconstruction image of internal corrosion defects. The network structure is as follows:

[0008] (1) Perform two convolution operations on the input U magnetic field images, and then pass them through a residual block to obtain U magnetic field matrices. Since feature fusion and extraction are required for the U magnetic field images of different frequencies, in order to avoid gradient explosion and vanishing due to too many network layers, a basic structure based on residual networks is chosen to realize the three-dimensional reconstruction of corrosion defects inside the structure. The biggest advantage of residual networks over other neural networks is that they introduce identity mappings, and solve the accuracy degradation problem caused by too many network layers by calculating residuals.

[0009] (2) Since the higher the frequency, the less information the magnetic field image contains, and the greater the difference between the simulation and experimental images, after two convolutions and one residual operation, the two magnetic field matrices with the highest frequencies are superimposed to form a fused magnetic field matrix. The above method can solve the problem by combining the low-frequency magnetic field image with the high-frequency magnetic field image while retaining the features of the high-frequency magnetic field image alone, thereby minimizing the inconsistency between the simulation and experimental data and obtaining richer internal corrosion information.

[0010] (3) The multi-frequency fused magnetic field matrix is ​​used to obtain a segmentation matrix through the Deeplab V3 network, and then the segmentation matrix is ​​mapped to the gray value of the response to form a three-dimensional reconstruction image of the predicted internal corrosion defects of the structure.

[0011] (4) Since the lower the frequency, the richer the information about the depth direction of the defects contained in the magnetic field image, the more sufficient the feature extraction needs to be, that is, the more residual blocks it passes through individually. Therefore, according to the method in (3) above, the fused magnetic field matrix needs to be superimposed with the low-frequency magnetic field matrix in sequence to form the final multi-frequency fused magnetic field matrix. Through this method, the defect information in the magnetic field images of different frequencies can be extracted to the maximum extent while saving computing resources, thereby improving the accuracy of the three-dimensional reconstruction of corrosion defects inside the structure.

[0012] (5) The multi-frequency fused magnetic field matrix is ​​used to obtain a segmentation matrix through the Deeplab V3 network, and then the segmentation matrix is ​​mapped to the corresponding gray value to form a three-dimensional reconstruction image of the predicted internal corrosion defects of the structure.

[0013] Step 3: Establish the loss function for the 3D reconstruction network of internal corrosion defects. The loss function is a crucial component of artificial neural networks, primarily used to calculate the difference between predicted and true values, and then optimize the network through error backpropagation. Conventional image segmentation loss functions do not consider the 3D features and evaluation metrics of internal corrosion defects. Embedding physical information into the loss function allows the network to optimize in the direction constrained by the physical information, improving the accuracy of the 3D reconstruction of internal corrosion defects. The final constructed loss function embedding physical information is as follows:

[0014]

[0015] Wherein, λ1, λ2 and λ3 are weighting coefficients;

[0016] Currently, the most widely used loss function in the field of image segmentation is the Softmax cross-entropy loss function. Its expression is:

[0017]

[0018] In the formula, N represents the number of categories, and P i This represents the probability distribution of the prediction results for the i-th category. Let represent the probability distribution of the true result for the i-th category;

[0019] For the 3D reconstruction of internal corrosion defects, the single Softmax cross-entropy loss function does not consider the volume and depth information of the internal corrosion, and cannot fully meet the accuracy requirements of 3D reconstruction. Therefore, a new physical information loss function needs to be constructed. The image segmentation task established in this method uses categories representing the depth of the defect, which differs from conventional image segmentation tasks. Therefore, the closer the network's classification result is to the true category, the better the network's performance is considered. For example, if the category of a pixel is 30, the closer the classification result is to 30, the better the network is considered. Therefore, it is necessary to calculate the difference between each pixel in the true and predicted 3D images of internal corrosion defects, and to focus on the magnitude of this difference; hence, an L1 loss function is required. The expression is:

[0020]

[0021] In the formula, x is the magnetic field image, z is noise, y is the actual three-dimensional reconstruction image of internal corrosion defects, and f(x,z) is the predicted three-dimensional reconstruction image of internal corrosion defects.

[0022] Volume is an important indicator for assessing internal corrosion of a structure. In order to make the predicted internal corrosion volume of the structure closer to the true value, it is necessary to calculate the difference between the actual internal corrosion volume and the predicted volume output by the network, and then incorporate it into the loss function to make the corrosion volume predicted by the network more accurate. The expression is:

[0023]

[0024] In the formula, sum represents summation, and |·| represents taking the absolute value;

[0025] Step 4: Use the magnetic field image sample library of internal corrosion defects of the rotating electromagnetic field structure established by S1 to train the three-dimensional reconstruction network of internal corrosion defects of the structure established by S2 and S3, and take the optimal solution;

[0026] Step 5: Deploy the trained network into the internal corrosion defect detection system of the rotating electromagnetic field.

[0027] Furthermore, the process of constructing the real 3D reconstructed image of the defect is as follows:

[0028] Draw an M×M grayscale image, where the pixel size is 0.5 mm and the thickness of the test block is H. Map the depth of each internal corrosion defect to the grayscale value of the grayscale image. The formula for calculating the grayscale value is P=(HD) / H×255, where P is the pixel value and D is the depth of a point in the internal corrosion defect.

[0029] M represents the length and width of the grayscale image.

[0030] Furthermore, the number and value of the frequencies of the magnetic field images are determined by the material and thickness of the structure under test. In order to obtain corrosion defect information of the full thickness of the structure, the skin depth of the lowest frequency of the magnetic field image must be greater than the thickness of the structure under test.

[0031] Furthermore, the residual block consists of three 1×1 convolutions and one 3×3 convolution.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] (1) This invention proposes a novel three-dimensional reconstruction method for internal corrosion defects of a structure based on a rotating electromagnetic field, which realizes the detection and high-precision reconstruction of internal corrosion defects of the structure, and provides data support for the safety assessment and maintenance decision of the structure.

[0034] (2) To address the difficulty in extracting features from multi-frequency magnetic field images, high- and low-frequency magnetic field images are fused under the guidance of physical information. This can maximize the extraction of defect information from magnetic field images of different frequencies while saving computational resources, thereby improving the accuracy of three-dimensional reconstruction of corrosion defects inside the structure.

[0035] (3) To address the issue of limited samples of internal surface corrosion defects, a loss function embedding physical information is proposed. The main evaluation indicators for internal corrosion defects are the depth and volume of internal corrosion. To incorporate the depth of internal corrosion into the loss function, a... The loss function calculates the difference between each pixel in the actual and predicted 3D images of corrosion defects, focusing on the magnitude of the difference. To incorporate the volume of corrosion into the loss function, a... The loss function focuses on calculating the difference between the actual volume of internal corrosion and the volume predicted by the network. By incorporating evaluation metrics into the loss function, the required sample size can be minimized, resulting in high-precision reconstructions of internal corrosion defects. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the three-dimensional reconstruction of internal corrosion defects in the structure as described in this application embodiment.

[0037] Figure 2 Magnetic field images of multiple frequencies in the embodiments of this application

[0038] Figure 3 Three-dimensional reconstruction image of internal corrosion defects in the structure in the embodiments of this application.

[0039] Figure 4 This application provides a three-dimensional reconstruction network of internal corrosion defects in the structure. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are all within the scope of protection of this invention.

[0041] This invention provides a method for three-dimensional reconstruction of internal corrosion defects in a rotating electromagnetic field structure, the steps of which include:

[0042] The first step is to establish a three-dimensional reconstruction sample library of corrosion defects inside rotating electromagnetic field structures, such as... Figure 2 , 3 As shown, the sample library includes magnetic field images at 5 frequencies and real three-dimensional reconstruction images of internal corrosion defects in the structure. The thickness of the tested component is 5 mm, the material is aluminum, and the frequencies of the magnetic field images are 200 Hz, 400 Hz, 600 Hz, 800 Hz and 1000 Hz.

[0043] Step 2: Establish a three-dimensional reconstruction network of internal corrosion defects in the structure, such as... Figure 4 As shown. The network input is a magnetic field image at 5 frequencies, and the output is a 3D reconstructed image of corrosion defects inside the structure. The network structure is as follows:

[0044] First, two convolution operations are performed on the input magnetic field images of different frequencies. The first convolution uses a 3×3 kernel with 16 layers, and the second uses a 1×1 kernel with 32 layers. Since higher frequencies contain less information in the magnetic field image, the difference between simulated and experimental images for deep defects is greater. Therefore, after the two convolution operations, the 1000 Hz and 800 Hz magnetic field images are each connected to a residual block for feature extraction, and then superimposed. The 600 Hz magnetic field image is superimposed with the fused magnetic field matrix formed by the 1000 Hz and 800 Hz images after feature extraction from two residual blocks. Similarly, the 400 Hz image is superimposed with the fused magnetic field matrix formed by the above frequencies after feature extraction from three residual blocks, and the 200 Hz image is superimposed with the fused magnetic field matrix formed by the above frequencies after feature extraction from four residual blocks. Because the 200 Hz image contains the richest information in the defect depth direction, it also has the most residual blocks for feature extraction. The multi-frequency fused magnetic field matrix is ​​used to obtain a segmentation matrix through the Deeplab V3 network. Then, the segmentation matrix is ​​mapped to the gray values ​​of the response to form a three-dimensional reconstruction image of the predicted internal corrosion defects of the structure.

[0045] Step 3: Establish the loss function for the 3D reconstruction network of internal corrosion defects. The loss function is a crucial component of artificial neural networks, primarily used to calculate the difference between predicted and true values, and then optimize the network through error backpropagation. Conventional image segmentation loss functions do not consider the 3D features and evaluation metrics of internal corrosion defects. Embedding physical information into the loss function allows the network to optimize in the direction constrained by the physical information, improving the accuracy of the 3D reconstruction of internal corrosion. The final constructed loss function embedding physical information is as follows:

[0046]

[0047] λ1, λ2, and λ3 are weight coefficients, set to 1, 10, and 100 respectively. The weight coefficients are allocated based on the importance of each loss term to the 3D reconstruction task of the defect, guiding the network to prioritize the optimization of key physical indicators through differentiated weights. Volume loss has the highest weight (100) because volume is the core physical indicator for evaluating internal corrosion defects and has the greatest impact on the reliability of the reconstruction results, requiring careful constraint. The loss weight is set to 10, which constrains the pixel-level reconstruction error of the defect depth, and has a secondary impact on the reconstruction accuracy of the depth dimension. Softmax cross-entropy loss has the lowest weight (1), as it is a conventional image segmentation loss, mainly assisting in category classification, and has a weaker constraint on 3D physical information, hence its lowest weight. This setting allows the network to prioritize the optimization of key physical indicators, taking into account errors in different dimensions, and improving the overall reconstruction accuracy.

[0048] Currently, the most widely used loss function in image segmentation is the Softmax cross-entropy loss function.

[0049] Softmax cross-entropy loss is a commonly used loss function for deep learning multi-class classification tasks, composed of Softmax activation and cross-entropy loss. Softmax normalizes the model's original output into probabilities for each class, satisfying the requirement of a non-negative distribution that sums to 1; cross-entropy measures the difference between the predicted probability distribution and the true label distribution, with a smaller loss indicating more accurate predictions. Its theoretical basis is information-theoretic cross-entropy and maximum likelihood estimation, equivalent to minimizing the negative log-likelihood. Gradient calculation is simple, and training is stable and converges quickly. Its expression is:

[0050]

[0051] In the formula, N represents the number of categories, and P i This represents the probability distribution of the prediction results for the i-th category. Let represent the probability distribution of the true result for the i-th category;

[0052] For the 3D reconstruction of internal corrosion defects, the single Softmax cross-entropy loss function does not consider the volume and depth information of the internal corrosion, and cannot fully meet the accuracy requirements of the 3D reconstruction. Therefore, a new physical information loss function needs to be constructed. The image segmentation task established in this method uses categories representing the depth of the defect, which differs from conventional image segmentation tasks. Therefore, the closer the network's classification result is to the true category, the better the network's performance is considered. For example, if the category of a pixel is 30, the closer the classification result is to 30, the better the network is considered. Therefore, it is necessary to calculate the difference between each pixel in the real and generated 3D corrosion defect images, and to focus on the magnitude of the difference; hence, an L1 loss function is required. The expression is:

[0053]

[0054] In the formula, x represents the magnetic field images at five frequencies, z represents noise, y represents the actual 3D reconstructed image of internal corrosion defects, and f(x,z) represents the predicted 3D reconstructed image of internal corrosion defects. Noise z is an unavoidable interference in actual detection, mainly including environmental noise during magnetic field image acquisition and instrument noise. Introducing this type of noise during training can simulate real-world conditions, improve the model's robustness to detection interference, and make the reconstruction results more consistent with actual scenarios.

[0055] Volume is an important indicator for assessing internal corrosion of a structure. In order to make the predicted internal surface corrosion volume closer to the true value, it is necessary to calculate the difference between the actual internal corrosion volume and the predicted volume output by the network, and then incorporate it into the loss function to make the corrosion volume predicted by the network more accurate. The expression is:

[0056]

[0057] In the formula, sum represents summation, and |·| represents taking the absolute value;

[0058] The two new constraints added to the loss function address the shortcomings of conventional image segmentation losses. First, there's the defect depth information constraint: in this task, the image segmentation category directly represents the defect depth. The loss function, by calculating the pixel-level difference between the real and predicted 3D reconstructed images, essentially constrains the reconstruction error of the defect depth, making the network focus more on the reconstruction accuracy of the depth dimension. Second, there's the erosion volume information constraint: volume is a key physical indicator for evaluating erosion defects. The volume loss function calculates the volume difference between the real and predicted erosion defects, guiding the network to optimize the predicted erosion volume to be closer to the real value. These two types of physical information constraints, combined with the conventional Softmax cross-entropy loss, allow the network to optimize according to physical information constraints, improving the accuracy of 3D reconstruction.

[0059] Step 4: Use the magnetic field image sample library of internal corrosion defects of rotating electromagnetic field structure established by S1 to train the three-dimensional reconstruction network of internal corrosion defects of structure established by S2 and S3. Set the batch size to 50, the training rounds to 300, select Adam optimizer for network training, set the learning rate to 1e-4, and take the optimal solution.

[0060] Step 5: Deploy the trained network into the internal corrosion defect detection system of the rotating electromagnetic field.

[0061] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for three-dimensional reconstruction of internal corrosion defects in a rotating electromagnetic field structure, characterized in that, S1: Establish a three-dimensional reconstruction sample library of internal corrosion defects in rotating electromagnetic field structures. The sample library includes magnetic field images of U frequencies and real three-dimensional reconstruction images of internal corrosion defects in the structure. U is the number of frequencies of the magnetic field images, which is not less than 4 and not more than 10. The data in the sample library are obtained from simulation and experiment. S2: Establish a three-dimensional reconstruction network for internal corrosion defects. The network input is a magnetic field image with U frequencies, and the output is a predicted three-dimensional reconstruction image of internal corrosion defects. The network structure is as follows: First, perform two convolution operations on the input U magnetic field images, and then pass them through a residual block to obtain U magnetic field matrices; The two magnetic field matrices with the highest frequencies are superimposed to form a fused magnetic field matrix; The fused magnetic field matrix and the magnetic field matrix with the third highest frequency are both passed through a residual block, and then the two are superimposed to obtain a new fused magnetic field matrix; Following the above method, the low-frequency magnetic field matrix is ​​sequentially superimposed on the fused magnetic field matrix to form the final multi-frequency fused magnetic field matrix; The multi-frequency fused magnetic field matrix is ​​used to obtain a segmentation matrix through the Deeplab V3 network. Then, the segmentation matrix is ​​mapped to the corresponding gray values ​​to form a three-dimensional reconstruction image of the predicted internal corrosion defects of the structure. S3: The loss function for establishing the 3D reconstruction network of internal corrosion defects in the structure is: ； Wherein, λ1, λ2 and λ3 are weighting coefficients; The expression is: ； In the formula, N represents the number of categories, and P i This represents the probability distribution of the prediction results for the i-th category. Let represent the probability distribution of the true result for the i-th category; The expression is: ； In the formula, x is the magnetic field image at U frequencies, z is noise, y is the actual three-dimensional reconstruction image of internal corrosion defects, and f(x,z) is the predicted three-dimensional reconstruction image of internal corrosion defects. The expression is: ； In the formula, sum represents summation, |·| represents taking the absolute value, and H is the thickness of the structure being measured; S4: The three-dimensional reconstruction network of internal corrosion defects of the structure established in S2 and S3 is trained using the magnetic field image sample library of internal corrosion defects of the rotating electromagnetic field structure established in S1, and the optimal solution is obtained. S5: Deploy the trained network into the internal corrosion defect detection system of the rotating electromagnetic field.

2. The method for three-dimensional reconstruction of internal corrosion defects in a rotating electromagnetic field structure according to claim 1, characterized in that, The process of constructing the three-dimensional reconstruction image of the actual internal corrosion defects of the structure is as follows: Draw an M×M grayscale image with a pixel size of 0.5 mm. Map the depth of each internal corrosion defect to the grayscale value of the grayscale image. The formula for calculating the grayscale value is P=(HD) / H×255, where P is the pixel value and D is the depth of a point in the internal corrosion defect. M represents the length and width of the grayscale image.

3. The method for three-dimensional reconstruction of internal corrosion defects in a rotating electromagnetic field structure according to claim 1, characterized in that, The number and value of the frequencies of the magnetic field images are determined by the material and thickness of the structure being measured, and the skin depth of the lowest frequency of the magnetic field images must be greater than the thickness of the structure being measured.

4. The method for three-dimensional reconstruction of internal corrosion defects in a rotating electromagnetic field structure according to claim 1, characterized in that, The residual block consists of three 1×1 convolutions and one 3×3 convolution.

Images