A welding structure defect recognition method fusing ray physical prior knowledge

Abstract

The application provides a fusion welding structure defect recognition method of ray physical priori knowledge, and relates to the technical field of nondestructive testing, and comprises the following steps: S1, a bevel geometry topology mask is constructed according to weld geometric parameters; a nonlinear ray attenuation model is established; S2, an enhanced feature map guided by a heat affected zone is obtained; S3, the bevel position mask is passed through a nonlinear modulation module to obtain an enhanced feature map guided by the bevel; S4, adaptive fusion is performed on data-driven branch output and physical guided branch output to generate joint features; S5, a priori probability distribution is determined according to a heat affected zone probability field and a bevel position mask, and a multi-scale variational physical consistency loss function is constructed; S6, a trained defect recognition model is obtained; and S7, a to-be-detected ray image is input into the trained defect recognition model. The application significantly improves the robustness and explainability of the model under complex conditions such as fuzzy defect boundaries and low contrast.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, and more particularly to a method for identifying defects in welded structures that integrates prior knowledge of radiographic physics. Background Technology

[0002] Welded structures are widely used in key industrial fields such as aerospace, shipbuilding, and nuclear energy equipment. The quality of welds directly affects the safety, durability, and reliability of the overall structure (Zhang Xiaobin. Application of Non-destructive Testing in Steel Structure Welding Quality Control [J]. Mechanical Engineering and Automation). Radiographic testing, as a commonly used non-destructive testing method, is widely used in weld quality assessment (Wu Guanhua, et al. Current Status and Research Progress of Radiographic Testing Technology in China [J]. Journal of Instrumentation). With the development of computer technology, automatic defect identification based on deep learning has gradually become a research hotspot.

[0003] However, existing deep learning methods still face core challenges in practical industrial applications. Purely data-driven deep learning models are essentially black-box systems, and their decision-making processes lack interpretability (Ke Shuai, et al. Development and Application of Convolutional Neural Networks in Industrial Defect Detection [J / OL]. Mechanical Research and Application). Furthermore, existing models only utilize image pixel information and fail to fully leverage the physical principles of ray detection, leading to misjudgments in cases of blurred defect boundaries and low contrast, and making it difficult to provide a physical explanation for the detection results. Therefore, a defect identification method that deeply integrates prior physical knowledge with deep learning models is needed to achieve a synergistic improvement in detection accuracy and interpretability. Summary of the Invention

[0004] In view of this, the purpose of this invention is to propose a welding structure defect identification method that integrates prior knowledge of X-ray physics, in order to solve the technical problem that existing models only use image pixel information and fail to make full use of the physical principles of X-ray detection, resulting in the model being prone to misjudgment when the defect boundary is blurred and the contrast is low, and it is difficult to give a physical explanation for the detection results.

[0005] The technical means employed in this invention are as follows: A method for identifying defects in welded structures that integrates prior knowledge of X-ray physics includes the following steps: S1. Obtain welding process parameters and material properties, solve the temperature field distribution based on the nonlinear heat diffusion equation, determine the boundary of the heat-affected zone based on the temperature field and generate the probability field of the heat-affected zone; construct the groove geometric topology mask based on the weld geometric parameters; establish a nonlinear ray attenuation model. S2. The heat-affected zone probability map generated in S1 is passed through a non-linear gated attention module to generate attention weights. The attention weights are then multiplied element-wise with the original feature map and added to obtain the enhanced feature map after heat-affected zone guidance. S3. The bevel position mask generated in S1 is passed through a nonlinear modulation module to obtain the enhanced feature map after bevel guidance; S4. A dual-branch feature encoder is used to perform data-driven branch encoding on the enhanced feature map after the heat-affected zone output from S2 and the enhanced feature map after the bevel output from S3. The nonlinear ray attenuation model established in S1 is physically guided by branch encoding. The data-driven branch output and the physical guided branch output are adaptively fused to generate joint features. The joint features are output by the detection head as the predicted gray value of the defect area, the predicted gradient value of the defect edge, the predicted defect location distribution, and the predicted defect image curvature. S5. Based on the probability field of the heat-affected zone and the bevel position mask generated in S1, determine the prior probability distribution and construct a multi-scale variational physical consistency loss function. S6. Repeat S2 to S5 until the total loss function converges, save the network parameters in the converged state, and obtain the trained defect recognition model. S7. Input the ray image to be detected into the trained defect recognition model obtained in S6, and output the defect type, confidence score and bounding box location information.

[0006] Furthermore, the formula for the probability field of the heat-affected zone described in S1 is as follows:

[0007] in, s ( () is the Sigmoid function. T For temperature field, T th Threshold temperature, α For smoothing parameters.

[0008] Furthermore, the formula for the bevel geometry topology mask described in S1 is as follows:

[0009] in, d ( x , y ) is a pixel ( x , y The shortest distance to the edge of the bevel. s g This is the scale parameter.

[0010] Furthermore, the formula for the nonlinear ray attenuation model described in S1 is as follows:

[0011] in, I 0 represents the incident intensity. I For the emission intensity, m ( d ) represents the thickness-dependent nonlinear attenuation coefficient. or The scattering modulation coefficient is, i The scattering angle is denoted by .

[0012] Furthermore, the formula for the nonlinear gated attention module described in S2 is as follows:

[0013] in, W 1 and W 2 is the learnable weight matrix. b 1 and b 2 represents the bias term, and σ is the Sigmoid activation function; The formula for the enhanced feature map guided by the heat-affected zone is as follows:

[0014] Where ⊙ represents element-wise multiplication. α For learnable parameters, F This is the input feature map.

[0015] Furthermore, the formula for the enhanced feature map after bevel guidance described in S3 is as follows:

[0016] in, β tanh( is a learnable parameter) ) is the hyperbolic tangent activation function. F in This is the input feature map. M groove A mask for the bevel location.

[0017] Furthermore, the formula for adaptive fusion to generate joint features described in S4 is as follows:

[0018] in, F data For data-driven branch output, F phys For physical guidance branch output, c and d These are learnable fusion parameters.

[0019] Furthermore, the loss function described in S5 includes grayscale consistency loss, edge gradient structure similarity loss, spatial distribution relative entropy loss, and local curvature regularization loss; The formula for grayscale consistency loss is as follows:

[0020] in, Let be the grayscale value of the defect region at the i-th pixel predicted by the model. The theoretical grayscale value is calculated based on the ray attenuation model. N This represents the total number of pixels.

[0021] The formula for the edge gradient structure similarity loss is as follows:

[0022] in, To predict the gradient vector of the image at the defect edge for the model, SSIM is the edge theory gradient vector calculated based on the ray attenuation model, and SSIM is the structural similarity index function.

[0023] The formula for the relative entropy loss of the spatial distribution is as follows:

[0024] in, This represents the probability distribution of defect locations predicted by the model. Let be the prior probability distribution.

[0025] The formula for the local curvature regularization loss is as follows:

[0026] in, I This is a grayscale image of the defect area.

[0027] Furthermore, the spatial distribution relative entropy loss mentioned in S5 is the relative entropy between the prior probability distribution and the predicted defect location distribution, wherein the prior probability distribution is:

[0028] in, This represents the probability value of the heat-affected zone. This represents the mask value at the bevel location.

[0029] Furthermore, the formula for the multi-scale variational physical consistency loss function is as follows:

[0030] in, L det To detect the loss, l 1. l 2. l 3. l 4 represents the balancing weight.

[0031] Compared with the prior art, the present invention has the following advantages: This invention deeply integrates a nonlinear ray attenuation physical model with prior knowledge of welding processes, ensuring that the defect features learned by the model are constrained by physical laws rather than entirely dependent on data-driven approaches. Compared to purely data-driven black-box models, the decision-making process of this invention is traceable to the physical field of the welding thermal cycle and the ray attenuation physical model, providing physical interpretability of the detection results. This effectively solves the problem of deep learning models being difficult to interpret due to their black-box nature in industrial nondestructive testing.

[0032] This invention constructs a probability field for the heat-affected zone and a geometric topological mask for the bevel, guiding the model to focus on defect-prone areas during the feature extraction stage and reducing the interference of background noise. At the same time, the multi-scale variational physical consistency loss function constrains the model's prediction results within the range allowed by the physical laws of ray attenuation, effectively suppressing the misjudgments and omissions that are prone to occur in pure data-driven models under complex conditions such as blurred defect boundaries and low contrast, and significantly improving the defect localization accuracy and classification accuracy.

[0033] This invention addresses practical engineering challenges in X-ray image inspection of welded structures, such as blurred defect boundaries and low contrast between defects and background. Local curvature regularization loss constrains the predicted image curvature to ensure the smoothness and geometric rationality of defect boundaries, making the inspection results more consistent with the physical morphological characteristics of actual welding defects. It exhibits stronger adaptability to X-ray inspection images under complex on-site conditions and can be directly deployed in intelligent welding quality inspection processes in key industrial fields such as aerospace, shipbuilding, and nuclear energy equipment, demonstrating clear engineering application prospects and promotional value. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a schematic diagram of the nonlinear gating attention module structure of the present invention.

[0036] Figure 2 This is a diagram of the non-Euclidean dual-branch encoder architecture of the present invention.

[0037] Figure 3This is a flowchart illustrating the calculation of the multi-scale variational physical consistency loss function of this invention.

[0038] Figure 4 This is a schematic diagram of the defect identification results of the welded structure radiographic image according to the present invention. Detailed Implementation

[0039] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0040] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0041] This invention provides a method for identifying welded structural defects by integrating prior knowledge of X-ray physics. Based on welding process parameters and material properties, it generates a probability field for the heat-affected zone and a geometric topological mask for the bevel. It embeds nonlinear physical priors into a neural network structure and applies multi-scale variational physical consistency constraints, achieving a synergistic improvement in detection accuracy and interpretability. The method includes the following steps: S1. Obtain welding process parameters and material properties; solve for the temperature field distribution based on the nonlinear heat diffusion equation; determine the boundary of the heat-affected zone and generate the probability field of the heat-affected zone based on the temperature field; construct a groove geometric topology mask based on the weld geometric parameters; establish a nonlinear ray attenuation model; specifically: Welding process parameters and material properties are obtained. Based on welding thermal cycle simulation, the nonlinear thermal diffusion equation is solved using the finite element method, where the temperature-dependent thermal conductivity is expressed in the Arrhenius form. Based on the obtained temperature field distribution, the boundary of the heat-affected zone (HAZ) is determined, and a HAZ distribution probability map is generated. Weld geometric parameters are obtained, and a bevel geometric topology mask is constructed based on these parameters, with high weight values ​​assigned to the bevel edge region and low weight values ​​assigned to non-bevel regions. A nonlinear ray attenuation physical model is established, introducing a scattering modulation term for subsequent calculation of physical consistency loss.

[0042] Welding process parameters and material properties are obtained based on the nonlinear thermal diffusion equation: (1) in, T For temperature field, k ( T ( ) represents temperature-dependent thermal conductivity. Q ( r , t () represents the heat source term. Solve for the temperature field distribution, determine the boundary of the heat-affected zone based on the temperature field, and generate the probability field of the heat-affected zone: (2) in, s ( () is the Sigmoid function. T th Threshold temperature, α The smoothing parameter is used. The probability map represents the prior probability that each pixel is located in a defect-prone area. A bevel geometry topology mask is constructed based on the weld geometry parameters: (3) in, d ( x , y ) is a pixel ( x , y The shortest distance to the edge of the bevel. s g Let be the scale parameter. Establish a nonlinear ray attenuation model: (4) in, I 0 represents the incident intensity. I For the emission intensity, m ( d ) represents the thickness-dependent nonlinear attenuation coefficient. or The scattering modulation coefficient is, i The scattering angle; S2. The heat-affected zone probability map generated in S1 is passed through a non-linear gated attention module to generate attention weights. The attention weights are then multiplied element-wise with the original feature map and added to obtain the enhanced feature map after heat-affected zone guidance. Specifically: like Figure 1 As shown, the probability map of the heat-affected zone is mapped to the feature map space through a convolutional layer. Attention weights are then generated by a nonlinear gated attention module. These attention weights are multiplied element-wise with the original feature map and added to the original feature map to obtain an enhanced feature map. An object detection network is used as the baseline model. This module is embedded in a pre-defined layer of the backbone network and applies enhanced weights to defect-prone areas during feature extraction, guiding the model to focus on defect features within the heat-affected zone.

[0043] The probability map of the heat-affected zone is processed through a non-linear gated attention module: (5) in, W 1 and W 2 is the learnable weight matrix. b 1 and b 2 represents the bias term. The enhanced feature map is obtained by performing element-wise multiplication of the attention weights with the original feature map and then summing the results. (6) Where ⊙ represents element-wise multiplication. α These are learnable parameters. During the backbone network feature extraction process, enhancement weights are applied to defect-prone areas to guide the model to focus on defect features within the heat-affected zone. S3. The bevel position mask generated in S1 is passed through a nonlinear modulation module to obtain the enhanced feature map after bevel guidance; specifically: A bevel geometry guidance module is embedded in the backbone network. The bevel location mask size is aligned with the feature map size. The bevel location mask is transformed using a hyperbolic tangent function through a nonlinear modulation module. The modulated mask is then multiplied element-wise with the feature map, and finally added to the original feature map through a residual connection. Learnable parameters are initialized to preset values ​​and adjusted during training. For incomplete penetration defects, this module assigns higher attention weights to the bevel edge region of the model.

[0044] The bevel location mask is passed through a nonlinear modulation module: (7) in, β tanh( is a learnable parameter) () is the hyperbolic tangent activation function. Higher feature extraction weights are applied to the bevel edge region to enhance the model's ability to detect incomplete penetration defects; S4. A dual-branch feature encoder is used to perform data-driven branch encoding on the enhanced feature map after heat-affected zone guidance output from S2 and the enhanced feature map after bevel guidance output from S3. Physically guided branch encoding is performed on the nonlinear ray attenuation model established in S1. The data-driven branch output and the physically guided branch output are adaptively fused to generate joint features. The joint features are then output by the detection head (composed of four parallel fully connected layers) as the predicted defect region grayscale value, predicted defect edge gradient value, predicted defect location distribution, and predicted defect image curvature. Specifically: A dual-branch feature encoder is constructed, where the data-driven branch employs a graph neural network to build the image feature map into a graph structure. Defect features with spatial topological structure are extracted through graph convolution operations, such as... Figure 2 As shown, the physics-guided branch employs a physical information neural network, taking the material's attenuation coefficient, thickness distribution, and geometric information as input, and outputting a physical feature vector. Feature fusion uses a combination of adaptive weighted fusion and element-wise multiplication to generate joint features. The fusion weights are learnable parameters initialized to preset values.

[0045] A dual-branch feature encoder is employed. The data-driven branch learns defect features in the image using a graph neural network, while the physics-guided branch constructs a physical encoder network using a physical information neural network. The inputs are material attenuation coefficients, thickness distribution, and geometric information, and the output is a defect feature representation conforming to a nonlinear ray attenuation law. Then, the two branches are adaptively fused to generate joint features. (8) in, F data For data-driven branch output, F phys For physical guidance branch output, c and d For learnable fusion parameters; S5. Based on the probability field of the heat-affected zone and the bevel location mask generated in S1, determine the prior probability distribution and construct a multi-scale variational physical consistency loss function; specifically: A multi-scale variational physical consistency loss function is constructed, including gray-level consistency loss, edge gradient structure similarity loss, spatial distribution relative entropy loss, and local curvature regularization loss, such as... Figure 3As shown. Gray-level consistency loss is measured by the mean square error between the actual gray-level value predicted by the model for the defect region and the theoretical gray-level value calculated based on the ray attenuation model, with weights set to preset values. Edge gradient structure similarity loss is measured by the structural similarity between the actual gradient value and the theoretical gradient value of the defect edge, with weights set to preset values. Spatial distribution relative entropy loss is measured by the relative entropy between the model-predicted defect location and the prior probability distribution determined by the prior map of the heat-affected zone and the bevel position mask, with weights set to preset values. Local curvature regularization loss is measured by the squared integral of the Laplacian operator of the predicted image, with weights set to preset values. The total loss function is a weighted sum of the detection loss and the above four physical consistency losses.

[0046] The total physical consistency loss function is: (9) Gray-scale consistency loss is measured by the mean square error between the actual gray-scale value of the defect area predicted by the calculation model and the theoretical gray-scale value calculated based on the nonlinear ray attenuation model. (10) Edge gradient structural similarity loss is measured by calculating the structural similarity between the actual gradient values ​​and the theoretical gradient values ​​at the defect edges: (11) The spatial distribution relative entropy loss is measured by calculating the relative entropy between the predicted defect location and the prior probability distribution determined by the probability field of the heat-affected zone and the bevel location mask. (12) The prior probability distribution is: (13) The local curvature regularization loss is measured by calculating the squared integral of the Laplacian operator over the predicted image: (14) The total loss function is a weighted sum of the detection loss and the physical consistency loss: (15) in, L det To detect the loss, l 1. l 2. l 3. l 4 represents the balancing weight.

[0047] S6. Repeat S2 to S5 until the total loss function converges, save the network parameters in the converged state, and obtain the trained defect recognition model. S7. Input the image of the ray to be detected into the trained defect recognition model obtained in S6, and output the defect type, confidence score, and bounding box location information. For example... Figure 4 As shown, the model automatically marks the defect detection results on the input welded structure ray image: the identified defect type is crack, the confidence score is 0.94, and the defect area is accurately marked with a red bounding box.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying defects in welded structures by integrating prior knowledge of X-ray physics, characterized in that, Includes the following steps: S1. Obtain welding process parameters and material properties, solve the temperature field distribution based on the nonlinear heat diffusion equation, determine the boundary of the heat-affected zone based on the temperature field and generate the probability field of the heat-affected zone; construct the groove geometric topology mask based on the weld geometric parameters; establish a nonlinear ray attenuation model. S2. The heat-affected zone probability map generated in S1 is passed through a non-linear gated attention module to generate attention weights. The attention weights are then multiplied element-wise with the original feature map and added to obtain the enhanced feature map after heat-affected zone guidance. S3. The bevel position mask generated in S1 is passed through a nonlinear modulation module to obtain the enhanced feature map after bevel guidance; S4. A dual-branch feature encoder is used to perform data-driven branch encoding on the enhanced feature map after the heat-affected zone output from S2 and the enhanced feature map after the bevel output from S3. The nonlinear ray attenuation model established in S1 is physically guided by branch encoding. The data-driven branch output and the physical guided branch output are adaptively fused to generate joint features. The joint features are output by the detection head as the predicted gray value of the defect area, the predicted gradient value of the defect edge, the predicted defect location distribution, and the predicted defect image curvature. S5. Based on the probability field of the heat-affected zone and the bevel position mask generated in S1, determine the prior probability distribution and construct a multi-scale variational physical consistency loss function. S6. Repeat S2 to S5 until the total loss function converges, save the network parameters in the converged state, and obtain the trained defect recognition model. S7. Input the ray image to be detected into the trained defect recognition model obtained in S6, and output the defect type, confidence score and bounding box location information.

2. The method for identifying welded structural defects by incorporating prior knowledge of radiographic physics as described in claim 1, characterized in that, The formula for the probability field of the thermally affected zone described in S1 is as follows: in, σ ( () is the Sigmoid function. T For temperature field, T th Threshold temperature, α For smoothing parameters.

3. The method for identifying welded structural defects by incorporating prior knowledge of radiographic physics as described in claim 1, characterized in that, The formula for the bevel geometry topology mask described in S1 is as follows: in, d ( x , y ) is a pixel ( x , y The shortest distance to the edge of the bevel. σ g This is the scale parameter.

4. The method for identifying welded structural defects by incorporating prior knowledge of radiographic physics as described in claim 1, characterized in that, The formula for the nonlinear ray attenuation model described in S1 is as follows: in, I 0 represents the incident intensity. I For the emission intensity, μ ( d ) represents the thickness-dependent nonlinear attenuation coefficient. η The scattering modulation coefficient is, θ The scattering angle is denoted by .

5. The method for identifying welded structural defects by incorporating prior knowledge of radiographic physics as described in claim 1, characterized in that, The formula for the nonlinear gated attention module described in S2 is as follows: in, W 1 and W 2 is the learnable weight matrix. b 1 and b 2 represents the bias term, and σ is the Sigmoid activation function; The formula for the enhanced feature map guided by the heat-affected zone is as follows: Where ⊙ represents element-wise multiplication. α For learnable parameters, F This is the input feature map.

6. The method for identifying welded structural defects by incorporating prior knowledge of radiographic physics as described in claim 1, characterized in that, The formula for the enhanced feature map after beveling described in S3 is as follows: in, β tanh( is a learnable parameter) ) is the hyperbolic tangent activation function. F in For the input feature map, M groove A mask for the bevel location.

7. The method for identifying welded structural defects by incorporating prior knowledge of radiographic physics as described in claim 1, characterized in that, The formula for adaptive fusion to generate joint features described in S4 is as follows: in, F data For data-driven branch output, F phys For physical guidance branch output, γ and δ These are learnable fusion parameters.

8. The method for identifying welded structural defects by incorporating prior knowledge of radiographic physics as described in claim 1, characterized in that, The loss functions described in S5 include gray-level consistency loss, edge gradient structure similarity loss, spatial distribution relative entropy loss, and local curvature regularization loss; The formula for grayscale consistency loss is as follows: in, Let be the grayscale value of the defect region at the i-th pixel predicted by the model. The theoretical grayscale value is calculated based on the ray attenuation model. N Total number of pixels; The formula for the edge gradient structure similarity loss is as follows: in, To predict the gradient vector of the image at the defect edge for the model, SSIM is the edge theory gradient vector calculated based on the ray attenuation model, and SSIM is the structural similarity index function. The formula for the relative entropy loss of the spatial distribution is as follows: in, This represents the probability distribution of defect locations predicted by the model. This represents the prior probability distribution; The formula for the local curvature regularization loss is as follows: in, I This is a grayscale image of the defect area.

9. The method for identifying welded structural defects by incorporating prior knowledge of radiographic physics as described in claim 8, characterized in that, The spatial distribution relative entropy loss mentioned in S5 is the relative entropy between the prior probability distribution and the predicted defect location distribution. The prior probability distribution is: in, This represents the probability value of the heat-affected zone. This represents the mask value at the bevel location.

10. The method for identifying welded structural defects by incorporating prior knowledge of radiographic physics as described in claim 8, characterized in that, The formula for the multi-scale variational physical consistency loss function is as follows: in, L det To detect the loss, λ 1. λ 2. λ 3. λ 4 represents the balancing weight.

Images