Magnetostriction-based multi-physical field defect detection method and system for pressure-bearing equipment

Abstract

The invention provides a magnetostriction-based multi-physical-field defect detection method and system for pressure-bearing equipment, and relates to the technical field of detection, and the method comprises the steps: constructing a reconfigurable magnetic field excitation array and a multi-physical-quantity sensing network through an intelligent excitation and sensing device, and dynamically adjusting the excitation phase to improve the detection precision. According to the layered coupling calculation scheme, a coupling relationship among a magnetic field, a stress field and a temperature field is established, and multi-physical field data calculation is carried out in combination with physical constraint conditions. After multi-level wavelet packet decomposition and noise reduction, the deep neural network is utilized to extract features and generate a defect feature dictionary, so that efficient defect positioning and size evaluation are realized, and the accuracy and reliability of detection are remarkably improved.

Description

Technical Field

[0001] The present invention relates to detection technology, and in particular to a multi-physical field defect detection method and system for pressure-bearing equipment based on magnetostriction. Background Art

[0002] During the operation of pressure-bearing equipment, various defects may occur inside the equipment, such as cracks, corrosion, and fatigue. These defects will seriously affect the safety and reliability of the equipment. Traditional defect detection methods mainly rely on technologies such as ultrasonic, ray, and magnetic particle. These methods have certain limitations in detection accuracy and efficiency. Defects and deficiencies of the prior art: First, traditional detection methods have poor adaptability to pressure-bearing equipment with complex geometries, making it difficult to achieve full coverage, resulting in some defects not being discovered in time. Second, in the multi-physical field coupling analysis of the prior art, the influence of temperature changes on material properties is often ignored, resulting in a reduction in the accuracy of detection results. Finally, in the feature extraction process of existing deep learning models, the effective integration of physical constraints is lacking, which may lead to inaccurate identification of defect features and affect subsequent defect location and size evaluation. Summary of the Invention

[0003] Embodiments of the present invention provide a multi-physical field defect detection method and system for pressure-bearing equipment based on magnetostriction, which can solve the problems in the prior art.

[0004] In the first aspect of the embodiments of the present invention, A multi-physical field defect detection method for pressure-bearing equipment based on magnetostriction is provided, including: Construct an intelligent excitation and sensing device, where the intelligent excitation and sensing device includes a reconfigurable magnetic field excitation array and a multi-physical quantity sensing network; the reconfigurable magnetic field excitation array uses electromagnetic coils arranged in a matrix layout to form a focused magnetic field through adaptive phase modulation, and the adaptive phase modulation dynamically adjusts the excitation phases of adjacent coils based on the geometric characteristics of the area to be detected and the magnetic field response signals collected in real time; the multi-physical quantity sensing network includes a double-layer orthogonally arranged GMR sensor array, a temperature sensor array, and a stress sensor array for collecting magnetic field signals, temperature signals, and stress signals; Construct a hierarchical coupling calculation scheme, and establish a coupling relationship between the magnetic field, stress field, and temperature field based on the magnetic field signals, temperature signals, and stress signals; the magnetic field and stress field are coupled through a magnetostriction constitutive equation including a stress-induced magnetization term and a magnetostrictive strain term, and the temperature field is coupled with the stress field by establishing a piecewise linearized mapping relationship of the influence of temperature on the magnetization characteristics and mechanical characteristics of the material; introduce magnetic field divergence constraints, mechanical equilibrium constraints, and heat conduction boundary constraints as physical constraint conditions, and use an adaptive grid meshing method based on the physical field gradient for calculation to obtain multi-physical field coupling data; Denoise the multi - physical - field coupling data by using multi - level wavelet packet decomposition based on the energy entropy criterion; input the denoised data into a deep neural network. The loss function of the deep neural network introduces field quantity conservation constraints and constitutive relation constraints, and uses a spatio - temporal feature fusion method with an attention mechanism to fuse the multi - scale features extracted by the deep neural network to generate a defect feature dictionary; perform defect location based on the defect feature dictionary, and evaluate the defect size through a deep regression network introducing physical interpretability constraints.

[0005] In an alternative implementation, Construct a hierarchical coupling calculation scheme, and establish the coupling relationship between the magnetic field, stress field, and temperature field based on the magnetic field signal, temperature signal, and stress signal; the magnetic field and stress field are coupled through the magnetostrictive constitutive equation including the stress - induced magnetization term and magnetostrictive strain term, and the temperature field is coupled with the stress field by establishing a piece - wise linearized mapping relationship of the influence of temperature on the magnetization characteristics and mechanical properties of materials; introduce the magnetic field divergence constraint, mechanical equilibrium constraint, and heat conduction boundary constraint as physical constraint conditions, and use an adaptive grid meshing method based on the physical field gradient for calculation. The steps to obtain multi - physical - field coupling data include: Establish a hierarchical coupling framework with the stress field as the core, and construct a coupling strength evaluation function by introducing a coupling weight coefficient based on the physical field action intensity. The coupling weight coefficient is determined according to the characteristic intensities of the magnetic field, temperature field, and stress field, and the characteristic intensities are characterized by the gradients and energy densities of each physical field. The coupling strength evaluation function is used to dynamically adjust the iterative strategy; Under the hierarchical coupling framework, construct a strong coupling layer of the magnetic field - stress field, and use an improved Jiles - Atherton model including the stress - induced magnetization term and magnetostrictive strain term to describe the bidirectional coupling relationship. Modify the magnetostrictive strain by introducing a non - linear magnetization intensity dependence term, and the stress - induced magnetization term considers the influence of the effective magnetic field and stress sensitivity coefficient; establish a weak coupling layer of the temperature field - stress field, and use a piece - wise linearized method to describe the influence of temperature on the magnetization characteristics and mechanical properties of materials. Introduce an improved Hermite interpolation function as a smooth transition function at the temperature critical point, and the improved Hermite interpolation function includes a normalized temperature variable and endpoint derivative values; Construct a physical constraint condition system based on the hierarchical coupling framework, including magnetic field divergence constraint, mechanical equilibrium constraint, and heat conduction boundary constraint. A coupling source term is introduced in the heat conduction boundary constraint, and the coupling source term includes hysteresis loss, magneto-elastic coupling loss, and eddy current loss. Construct a grid density function based on the physical field gradient information, divide grid cells based on the grid density function, perform a weighted combination of the output value of the coupling strength evaluation function and the grid density function to form a composite density function, use the composite density function to determine the local refinement coefficient of the grid cell, and dynamically adjust the size of the grid cell according to the local refinement coefficient.

[0006] In an alternative embodiment, The steps of constructing a grid density function based on the physical field gradient information, dividing grid cells based on the grid density function, performing a weighted combination of the output value of the coupling strength evaluation function and the grid density function to form a composite density function, using the composite density function to determine the local refinement coefficient of the grid cell, and dynamically adjusting the size of the grid cell according to the local refinement coefficient include: Construct a grid density function based on the physical field gradient information. The grid density function determines the weight coefficients of each physical field through the maximum variation of the magnetic field strength gradient, stress gradient, and temperature gradient within the computational domain. Perform a weighted combination of the output value of the coupling strength evaluation function and the grid density function to form a composite density function. The balance coefficient of the weighted combination is given by an exponentially decaying function determined by the output value of the coupling strength evaluation function. Use the composite density function to determine the local refinement coefficient of the grid cell, and introduce a smoothing factor to smooth the local refinement coefficient of the grid cell. The smoothing factor performs a weighted combination of the currently calculated local refinement coefficient and the local refinement coefficient of the previous iteration. The value of the smoothing factor is positively correlated with the deformation rate of the grid cell. When the deformation rate of the grid cell is greater than the preset deformation rate threshold, the weight coefficient of the historical information increases proportionally. When the deformation rate of the grid cell is less than the preset deformation rate threshold, the weight coefficient of the current information increases accordingly. Dynamically adjust the size of the grid cell according to the local refinement coefficient. When the size ratio of adjacent grid cells exceeds the set size ratio threshold or the shape quality index of the grid cell is less than the set quality threshold, use the Laplacian smoothing method for local remeshing. During the dynamic adjustment of the grid cell size, determine the time step based on the ratio of the local deformation speed of the grid cell to the characteristic size of the grid cell. The time step decreases as the deformation degree of the grid cell increases. Construct a grid cell optimization index based on strain energy density, and mark the corresponding grid cells as areas to be removed when the grid cell optimization index is less than the set optimization threshold; trigger a grid cell local reconstruction mechanism in the high-coupling area, including identifying nodes in the high-stress area, reconstructing local grid cells based on the Delaunay criterion, and using the spline interpolation method to restore physical field information.

[0007] In an alternative embodiment, Input the denoised data into a deep neural network. The loss function of the deep neural network introduces field quantity conservation constraints and constitutive relation constraints. The steps of generating a defect feature dictionary by using a spatio-temporal feature fusion method with an attention mechanism to fuse the multi-scale features extracted by the deep neural network include: Input the denoised data into a deep neural network, and construct a composite loss function including a data fitting loss term, a field quantity conservation constraint loss term, and a constitutive relation constraint loss term. The data fitting loss term is calculated by the mean square error between the network prediction output and the actual labeled data. The field quantity conservation constraint loss term is calculated based on the square difference between the divergence of the stress tensor and the body force. The constitutive relation constraint loss term is calculated based on the square difference between the stress tensor and the product of the elastic tensor and the strain tensor. The composite loss function combines the data fitting loss term, the field quantity conservation constraint loss term, and the constitutive relation constraint loss term through a weighting coefficient; Fuse the multi-scale features extracted by the deep neural network to obtain fused features, and calculate feature weights using a temporal attention mechanism and a spatial attention mechanism. The temporal attention weight is obtained through the product operation of a learnable parameter matrix with a query matrix and a key-value matrix. The spatial attention weight is obtained through concatenation and convolution operations on the average pooling result and the maximum pooling result of the feature map; Construct a hierarchical-structured defect feature dictionary through sparse coding based on the fused features, and perform iterative update using an online dictionary learning algorithm. The defect feature dictionary includes a top-level dictionary representing macroscopic features and a bottom-level dictionary representing local features. The dictionary update process includes: dynamic learning rate adjustment based on the sparsity of defect features, redundancy evaluation based on the similarity of dictionary atoms, and update weight allocation based on sample difficulty; the top-level dictionary is optimized based on stress field continuity constraints and strain energy density constraints to obtain macroscopic features that satisfy mechanical laws, and the bottom-level dictionary is optimized through material constitutive relation constraints and stress tensor divergence constraints to obtain local detail features. A two-way constraint transfer mechanism is established between the top-level dictionary and the bottom-level dictionary. The optimization direction of the bottom-level dictionary is guided by the stress field continuity constraints and strain energy density constraints based on the top-level dictionary, and the feature expression of the top-level dictionary is supplemented by the local features of the bottom-level dictionary; the dictionary update uses the gradient descent method to optimize the composite loss function so that each basis vector in the defect feature dictionary corresponds to different types of defect features.

[0008] In an alternative embodiment, The steps of combining the data fitting loss term, the field quantity conservation constraint loss term, and the constitutive relation constraint loss term by the composite loss function through a weighting coefficient include: Construct a confidence evaluation function for the data fitting loss term based on the spatial gradient consistency between the predicted value and the true value, construct a confidence evaluation function for the field quantity conservation constraint loss term based on the local balance degree of the stress field, and construct a confidence evaluation function for the constitutive relation constraint loss term based on the parameter sensitivity of the elastic tensor; The confidence evaluation function all adopts an adaptive scaling factor, and the adaptive scaling factor is determined by local gradient statistics and material parameter sensitivity; Calculate the dynamic weighting coefficient of each loss term based on the confidence evaluation function; Construct a stress-strain field gradient constraint term based on the local gradient statistics; Construct a physical coupling mechanism between loss terms based on the dynamic weighting coefficient, define a local consistency measure of the stress-strain field in the form of volume integral, establish the gradient coupling strength between field quantity conservation and constitutive relations, and the gradient coupling strength is calculated based on the inner product of the gradient fields; Construct a mutual feedback modulation function according to the gradient coupling strength, and integrate the mutual feedback modulation function into the composite loss function; Establish a constraint satisfaction evaluation mechanism with adaptive characteristic scale based on the mutual feedback modulation function, and the constraint satisfaction evaluation mechanism is calculated based on the local balance degree of the stress field and the deviation degree of the stress-strain relationship; Construct a constraint violation penalty term, and the magnitude of the constraint violation penalty term is positively correlated with the difference between the constraint satisfaction degree and the constraint satisfaction degree threshold; Design a gradient modulation factor based on the constraint satisfaction degree, and the gradient modulation factor increases as the constraint satisfaction degree decreases; Integrate the constraint violation penalty term and the stress-strain gradient constraint term into the composite loss function, and use the modulated gradient descent method to update the optimization parameters. The learning rate during the parameter update process is dynamically adjusted according to the constraint satisfaction degree and the gradient norm; When the gradient value is less than the preset gradient threshold and the constraint satisfaction degree is greater than the preset constraint satisfaction degree threshold, it is determined that the optimization converges.

[0009] In an alternative embodiment, The steps of fusing the multi-scale features extracted by the deep neural network to obtain the fused features, and calculating the feature weights using the temporal attention mechanism and the spatial attention mechanism, where the temporal attention weights are obtained through the product operation of the learnable parameter matrix and the query matrix and the key-value matrix, and the spatial attention weights are obtained through the concatenation and convolution operations of the average pooling result and the maximum pooling result of the feature map include: Construct a time-series feature matrix of the stress field, introduce a learnable physical constraint parameter matrix to map the original features to a feature space that satisfies the continuity of the stress field, and calculate a query matrix, a key matrix, and a value matrix based on the physical constraint parameter matrix; combine the product result of the query matrix and the key matrix with the stress field continuity constraint term to calculate the time-series attention score, where the stress field continuity constraint term is calculated through the divergence of the stress tensor; multiply the time-series attention score by the value matrix to obtain the time-series feature output; Perform average pooling and max pooling operations on the time-series feature output respectively to obtain two feature maps, combine the concatenation result of the feature maps with the strain energy density constraint term to calculate the spatial attention weight, where the strain energy density constraint term is calculated by integrating the inner product of the stress tensor and the strain tensor over the volume, and the spatial attention weight is normalized through the sigmoid function; Fuse the features at different scale levels, multiply each layer of features by the corresponding time-series feature output and spatial attention weight respectively and sum them to obtain the fused features; Update the physical constraint parameter matrix by optimizing the composite loss function, and substitute the updated physical constraint parameter matrix into the calculation formulas of the stress field continuity constraint term and the strain energy density constraint term; the update result of the physical constraint parameter matrix is used for the calculation of the next round of feature mapping, forming an iterative process of constraint optimization.

[0010] In an optional implementation manner, The steps of defect location based on the defect feature dictionary and evaluating the defect size through a deep regression network introducing physical interpretability constraints include: Perform sparse decomposition on the feature map of the test sample, decompose the feature map into the sum of the product of the defect feature dictionary and the sparse coefficient and the residual term, and introduce the stress field continuity constraint calculated by the divergence of the stress tensor in the sparse decomposition process; based on the non-zero element distribution of the sparse coefficient, map the macroscopic features of the top-level dictionary and the local features of the bottom-level dictionary and perform an element-wise product operation to obtain the defect probability map; Extract the initial candidate region set based on the probability threshold for the defect probability map, calculate the strain energy density of each region in the initial candidate region set, where the strain energy density is obtained by integrating the stress tensor and the strain tensor over the volume; combine the strain energy density with the region area and perimeter to calculate the candidate region score, and screen out the final defect location result based on the preset region score threshold; Construct a deep regression network with an encoder-decoder structure, where the encoder consists of multiple residual blocks with dilated convolutions, and the decoder adopts a progressive upsampling structure; input the original feature map, the defect region mask, and the defect location result into the deep regression network; A feature extraction module, a geometric parameter regression module, and a physical interpretability constraint module are set in the depth regression network. The feature extraction module extracts multi-scale features through residual blocks and performs feature fusion using an attention mechanism. The geometric parameter regression module outputs the length, width, and depth parameters of the defect through independent fully connected layers. The physical interpretability constraint module calculates the stress distribution and strain distribution based on the predicted geometric parameters; A multi-task loss function including a geometric parameter regression loss term, a stress field constraint loss term, a strain field constraint loss term, and an energy conservation constraint loss term is constructed. The stress field constraint loss term is calculated based on the divergence of the stress tensor, the strain field constraint loss term is calculated based on the gradient of the strain tensor, and the energy conservation constraint loss term is calculated based on the difference between the system energy calculated from the original feature map and the system energy calculated after reconstructing the predicted defect geometric parameters. The depth regression network is trained by optimizing the multi-task loss function to obtain a defect size evaluation result that satisfies the physical interpretability constraint. The defect size evaluation result includes the length, width, and depth parameters of the defect.

[0011] In the second aspect of the embodiments of the present invention, A multi-physical field defect detection system for pressure-bearing equipment based on magnetostriction is provided, including: A first unit for constructing an intelligent excitation and sensing device. The intelligent excitation and sensing device includes a reconfigurable magnetic field excitation array and a multi-physical quantity sensing network. The reconfigurable magnetic field excitation array uses an electromagnetic coil with a matrix layout to form a focused magnetic field through adaptive phase modulation. The adaptive phase modulation dynamically adjusts the excitation phases of adjacent coils based on the geometric characteristics of the area to be detected and the magnetic field response signals collected in real time. The multi-physical quantity sensing network includes a double-layer orthogonally arranged GMR sensor array, a temperature sensor array, and a stress sensor array for collecting magnetic field signals, temperature signals, and stress signals; A second unit for constructing a hierarchical coupling calculation scheme and establishing a coupling relationship between the magnetic field, stress field, and temperature field based on the magnetic field signals, temperature signals, and stress signals. The magnetic field and stress field are coupled through a magnetostriction constitutive equation including a stress-induced magnetization term and a magnetostrictive strain term. The temperature field is coupled with the stress field by establishing a piecewise linearized mapping relationship of the influence of temperature on the magnetization characteristics and mechanical characteristics of the material. Magnetic field divergence constraint, mechanical equilibrium constraint, and heat conduction boundary constraint are introduced as physical constraint conditions, and an adaptive grid meshing method based on the gradient of the physical field is used for calculation to obtain multi-physical field coupling data; A third unit is configured to perform noise reduction on the multi - physical - field coupling data by using multi - level wavelet packet decomposition based on the energy entropy criterion; input the denoised data into a deep neural network. The loss function of the deep neural network introduces field quantity conservation constraints and constitutive relation constraints, and uses a spatio - temporal feature fusion method with an attention mechanism to fuse the multi - scale features extracted by the deep neural network to generate a defect feature dictionary; perform defect localization based on the defect feature dictionary, and evaluate the defect size through a deep regression network with physical interpretability constraints introduced.

[0012] In the third aspect of the embodiments of the present invention, a kind of electronic device is provided, including: a processor; a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0013] In the fourth aspect of the embodiments of the present invention, a computer - readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0014] By constructing an intelligent excitation and sensing device, the present invention can realize multi - physical - field defect detection of pressure - bearing equipment, improve the detection accuracy and sensitivity, and ensure the safety and reliability of the equipment; the introduction of the hierarchical coupling calculation scheme of the present invention enables the effective establishment of the coupling relationship between the magnetic field, stress field and temperature field, can more comprehensively reflect the state of the equipment under different working conditions, and enhances the accuracy of the detection results; the present invention uses deep neural network and multi - level wavelet packet decomposition technology for data processing, can effectively reduce noise interference, improve the ability to extract defect features, thereby realizing accurate defect localization and size evaluation, and improving the detection efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a schematic flow chart of the multi - physical - field defect detection method based on magnetostriction for pressure - bearing equipment in the embodiments of the present invention; Figure 2 is a schematic structural diagram of the multi - physical - field defect detection system based on magnetostriction for pressure - bearing equipment in the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0017] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0018] Figure 1 The flow chart of the multi-physical field defect detection method for pressure-bearing equipment based on magnetostriction in the embodiments of the present invention is shown as Figure 1 shown, and the method includes: S1. Construct an intelligent excitation and sensing device, where the intelligent excitation and sensing device includes a reconfigurable magnetic field excitation array and a multi-physical quantity sensing network; the reconfigurable magnetic field excitation array uses electromagnetic coils with a matrix layout to form a focused magnetic field through adaptive phase modulation, and the adaptive phase modulation dynamically adjusts the excitation phases of adjacent coils based on the geometric characteristics of the area to be detected and the magnetoresponse signals collected in real time; the multi-physical quantity sensing network includes a double-layer orthogonally arranged GMR sensor array, a temperature sensor array, and a stress sensor array for collecting magnetic field signals, temperature signals, and stress signals; S2. Construct a hierarchical coupling calculation scheme, and establish the coupling relationship between the magnetic field, stress field, and temperature field based on the magnetic field signal, temperature signal, and stress signal; the magnetic field and stress field are coupled through a magnetostriction constitutive equation including a stress-induced magnetization term and a magnetostrain term, and the temperature field is coupled with the stress field by establishing a piecewise linearized mapping relationship of the influence of temperature on the magnetization characteristics and mechanical characteristics of materials; introduce magnetic field divergence constraints, mechanical equilibrium constraints, and heat conduction boundary constraints as physical constraint conditions, and use an adaptive mesh generation method based on the gradient of the physical field for calculation to obtain multi-physical field coupling data; S3. Use multi-level wavelet packet decomposition based on the energy entropy criterion to denoise the multi-physical field coupling data; input the denoised data into a deep neural network, introduce field quantity conservation constraints and constitutive relationship constraints into the loss function of the deep neural network, and use a spatio-temporal feature fusion method with an attention mechanism to fuse the multi-scale features extracted by the deep neural network to generate a defect feature dictionary; perform defect localization based on the defect feature dictionary, and evaluate the defect size through a deep regression network introducing physical interpretability constraints.

[0019] In an alternative embodiment, Construct a hierarchical coupling calculation scheme, and establish the coupling relationship of the magnetic field, stress field, and temperature field based on the magnetic field signal, temperature signal, and stress signal; the magnetic field and stress field are coupled through the magnetostrictive constitutive equation including the stress-induced magnetization term and magnetostrictive strain term, and the temperature field is coupled with the stress field by establishing a piecewise linearized mapping relationship of the influence of temperature on the magnetization characteristics and mechanical characteristics of the material; introducing the magnetic field divergence constraint, mechanical equilibrium constraint, and heat conduction boundary constraint as physical constraint conditions, and using the adaptive grid meshing method based on the physical field gradient for calculation. The steps to obtain the multi-physical field coupling data include: Establish a hierarchical coupling framework with the stress field as the core, and construct a coupling strength evaluation function by introducing a coupling weight coefficient based on the action intensity of the physical field. The coupling weight coefficient is determined according to the characteristic intensities of the magnetic field, temperature field, and stress field, and the characteristic intensities are characterized by the gradients and energy densities of each physical field. The coupling strength evaluation function is used to dynamically adjust the iterative strategy; Under the hierarchical coupling framework, construct a strong coupling layer of the magnetic field-stress field, and use an improved Jiles-Atherton model including the stress-induced magnetization term and magnetostrictive strain term to describe the bidirectional coupling relationship. Modify the magnetostrictive strain by introducing a non-linear magnetization intensity dependence term, and the stress-induced magnetization term considers the influence of the effective magnetic field and stress sensitivity coefficient; establish a weak coupling layer of the temperature field-stress field, and use the piecewise linearization method to describe the influence of temperature on the magnetization characteristics and mechanical characteristics of the material. Introduce an improved Hermite interpolation function as a smooth transition function at the temperature critical point, and the improved Hermite interpolation function includes a normalized temperature variable and endpoint derivative values; Construct a physical constraint condition system based on the hierarchical coupling framework, including the magnetic field divergence constraint, mechanical equilibrium constraint, and heat conduction boundary constraint. A coupling source term is introduced in the heat conduction boundary constraint, and the coupling source term includes hysteresis loss, magneto-elastic coupling loss, and eddy current loss; construct a grid density function according to the physical field gradient information, divide grid cells based on the grid density function, perform a weighted combination of the output value of the coupling strength evaluation function and the grid density function to form a composite density function, use the composite density function to determine the local refinement coefficient of the grid cells, and dynamically adjust the size of the grid cells according to the local refinement coefficient.

[0020] Exemplarily, the multi-physical field coupling detection technology based on the magnetostrictive effect realizes in-service detection of pressure-bearing equipment by constructing an intelligent excitation and sensing device. This intelligent excitation and sensing device adopts a matrix layout design and includes two core components: a reconfigurable magnetic field excitation array and a multi-physical quantity sensing network. The reconfigurable magnetic field excitation array is composed of 24×24 excitation coils. The outer diameter of each coil is 10 mm, the inner diameter is 6 mm, and the number of turns is 200. The coil array forms a focused magnetic field through adaptive phase modulation technology. The phase difference between adjacent coils can be continuously adjusted within the range of 0-360 degrees. In practical applications, a phase difference of 90 degrees is usually taken to obtain the best focusing effect.

[0021] The multi-physical quantity sensing network adopts a double-layer orthogonal layout. An array of 16×16 GMR sensors is arranged on the upper layer, and the sensor spacing is 15 mm. An array of 8×8 temperature sensors and stress sensors is arranged on the lower layer, and the sensor spacing is 30 mm. The sensitivity of the GMR sensor is 1.5×10⁻ 4 V / (A / m), and the measuring range is ±1600 A / m; the measuring range of the temperature sensor is -40°C to 150°C, and the accuracy is ±0.1°C; the measuring range of the stress sensor is 0-100 MPa, and the resolution is 0.1 MPa.

[0022] The layered coupling calculation scheme takes the stress field as the core and establishes a three-layer coupling framework. The first layer is the strong coupling layer of the magnetic field-stress field, and an improved Jiles-Atherton model is used to describe the bidirectional coupling relationship. This model introduces a non-linear magnetization intensity dependence term on the basis of the traditional Jiles-Atherton model. Among them, the stress-induced magnetization term is corrected by the effective magnetic field coefficient and the stress sensitivity coefficient. The value range of the effective magnetic field coefficient is 0.6-0.9, and the stress sensitivity coefficient range is 0.2-0.5. The correction of the magnetostrictive strain term adopts a piecewise function form. In the low magnetization intensity region (less than 0.5 T), a linear relationship is used, and in the high magnetization intensity region (greater than 0.5 T), a quadratic function relationship is used.

[0023] The second layer is the weak coupling layer of the temperature field-stress field. The mapping relationship between temperature and material properties is established by the piecewise linearization method. In the range from room temperature to 200°C, the magnetization characteristics of the material decrease by 2% for every 10°C increase, and the mechanical characteristics decrease by 1% for every 10°C increase. An improved Hermite interpolation function is used to achieve a smooth transition at the temperature critical point. This interpolation function includes a normalized temperature variable and the derivative values at two endpoints to ensure the continuity of the first derivative at the critical point.

[0024] The third layer is the physical constraint condition system, including magnetic field divergence constraint, mechanical equilibrium constraint, and heat conduction boundary constraint. The coupled source term introduced in the heat conduction boundary constraint considers three types of losses: hysteresis loss accounts for about 60% of the total loss, magneto-elastic coupling loss accounts for about 25%, and eddy current loss accounts for about 15%. The grid density function constructed based on the physical field gradient information adopts an adaptive method, and grid encryption is triggered when the local physical field gradient exceeds 1.5 times the average value. The local encryption coefficient range of the grid cells is from 1 to 4, that is, the grid size of the maximum encryption area is 1 / 4 of the reference grid size.

[0025] The present invention effectively handles the interaction between multiple physical fields through a hierarchical coupling framework. In particular, the strong coupling relationship between the magnetic field and the stress field is accurately described; the use of an adaptive grid meshing method significantly improves the calculation efficiency, and more refined calculation results can be obtained in key areas; the heat conduction boundary constraint introducing the coupled source term realizes the accurate characterization of the energy conversion process, providing a reliable data basis for subsequent defect detection.

[0026] In an alternative embodiment, The steps of constructing a grid density function according to the physical field gradient information, dividing grid cells based on the grid density function, performing a weighted combination of the output value of the coupling strength evaluation function and the grid density function to form a composite density function, determining the local encryption coefficient of the grid cells using the composite density function, and dynamically adjusting the size of the grid cells according to the local encryption coefficient include: Construct a grid density function according to the physical field gradient information, and the grid density function determines the weight coefficients of each physical field through the maximum variation of the magnetic field strength gradient, stress gradient, and temperature gradient in the computational domain; Perform a weighted combination of the output value of the coupling strength evaluation function and the grid density function to form a composite density function, and the balance coefficient of the weighted combination is given by an exponentially decaying function determined by the output value of the coupling strength evaluation function; Use the composite density function to determine the local encryption coefficient of the grid cells, and introduce a smoothing factor to smooth the local encryption coefficient of the grid cells. The smoothing factor performs a weighted combination of the currently calculated local encryption coefficient and the local encryption coefficient of the previous iteration. The value of the smoothing factor is positively correlated with the deformation rate of the grid cells. When the deformation rate of the grid cells is greater than the preset deformation rate threshold, the weight coefficient of the historical information increases proportionally. When the deformation rate of the grid cells is less than the preset deformation rate threshold, the weight coefficient of the current information increases accordingly; Dynamically adjust the size of grid cells according to the local encryption coefficient. When the size ratio of adjacent grid cells exceeds the set size ratio threshold or the shape quality index of the grid cell is less than the set quality threshold, use the Laplacian smoothing method for local remeshing; during the dynamic adjustment of the size of the grid cell, determine the time step based on the ratio of the local deformation speed of the grid cell to the characteristic size of the grid cell, and the time step decreases as the deformation degree of the grid cell increases; Construct an optimization index of grid cells based on strain energy density. When the optimization index of the grid cell is less than the set optimization threshold, mark the corresponding grid cell as the area to be removed; trigger the local grid cell reconstruction mechanism in the high-coupling area, including identifying the nodes in the high-stress area, reconstructing the local grid cells based on the Delaunay criterion, and using the spline interpolation method to restore the physical field information.

[0027] Exemplarily, in the process of constructing the grid density function according to the physical field gradient information, first, it is necessary to collect the gradient information of magnetic field strength, stress, and temperature within the computational domain. This information will be used to determine the weight coefficients of each physical field. By analyzing the maximum change amount of these gradients, a grid density function reflecting the characteristics of the physical field can be constructed. The construction of this function is to better reflect the change characteristics of the physical field in the subsequent grid division, so as to achieve a finer grid division.

[0028] After constructing the grid density function, next, it is necessary to perform a weighted combination of the output value of the coupling strength evaluation function and the grid density function to form a composite density function. The balance coefficient of the weighted combination is given by an exponentially decaying function determined by the output value of the coupling strength evaluation function. This process ensures that the coupling relationship between different physical fields can be effectively reflected in the grid division, thereby improving the calculation accuracy.

[0029] Through the composite density function, the local encryption coefficient of the grid cell can be determined. To improve the stability of the local encryption coefficient, a smoothing factor is introduced for smoothing. The role of the smoothing factor is to perform a weighted combination of the currently calculated local encryption coefficient and the local encryption coefficient of the previous iteration. The value of the smoothing factor is positively correlated with the deformation rate of the grid cell. When the deformation rate of the grid cell exceeds the preset threshold, the weight coefficient of the historical information will increase proportionally to ensure that the historical information can have a greater impact on the current calculation in the case of large deformation.

[0030] After determining the local encryption coefficient, next, dynamically adjust the size of the grid cell. When the size ratio of adjacent grid cells exceeds the set threshold, or the shape quality index of the grid cell is less than the set quality threshold, use the Laplacian smoothing method for local remeshing. This process ensures the quality of the grid and the stability of the calculation.

[0031] During the dynamic adjustment of the grid cell size, the time step is determined based on the ratio of the local deformation speed of the grid cell to the characteristic size of the grid cell. As the deformation degree of the grid cell increases, the time step will be correspondingly reduced to more accurately capture the changes in the physical field.

[0032] Finally, a grid cell optimization index based on strain energy density is constructed. When this index is less than the set optimization threshold, the corresponding grid cell is marked as the area to be removed. In the high-coupling region, a local grid cell reconstruction mechanism is triggered, including identifying the nodes in the high-stress region, reconstructing the local grid cells based on the Delaunay criterion, and using the spline interpolation method to restore the physical field information.

[0033] The present invention improves the accuracy of grid division, making the calculation results more reliable. By dynamically adjusting the size of grid cells, the use of computing resources is optimized, and the computing efficiency is improved. By introducing a smoothing factor and an optimization index, the stability of the grid is enhanced, and the errors in the calculation process are reduced.

[0034] In an alternative embodiment, Input the denoised data into a deep neural network. The loss function of the deep neural network introduces the field quantity conservation constraint and the constitutive relation constraint. The steps of generating a defect feature dictionary by fusing the multi-scale features extracted by the deep neural network using the spatio-temporal feature fusion method with an attention mechanism include: Input the denoised data into a deep neural network, and construct a composite loss function including a data fitting loss term, a field quantity conservation constraint loss term, and a constitutive relation constraint loss term. The data fitting loss term is calculated by the mean square error between the network prediction output and the actual labeled data. The field quantity conservation constraint loss term is calculated based on the square difference between the divergence of the stress tensor and the body force. The constitutive relation constraint loss term is calculated based on the square difference between the stress tensor and the product of the elastic tensor and the strain tensor. The composite loss function combines the data fitting loss term, the field quantity conservation constraint loss term, and the constitutive relation constraint loss term through a weighting coefficient; Fuse the multi-scale features extracted by the deep neural network to obtain a fused feature, and calculate the feature weights using a temporal attention mechanism and a spatial attention mechanism. The temporal attention weight is obtained through the product operation of a learnable parameter matrix with a query matrix and a key-value matrix. The spatial attention weight is obtained through concatenation and convolution operations on the average pooling result and the maximum pooling result of the feature map; Construct a defect feature dictionary with a hierarchical structure through sparse coding based on the fusion features, and use an online dictionary learning algorithm for iterative update. The defect feature dictionary includes a top-level dictionary representing macroscopic features and a bottom-level dictionary representing local features. The dictionary update process includes: dynamic learning rate adjustment based on the sparsity of defect features, redundancy evaluation based on the similarity of dictionary atoms, and update weight allocation based on sample difficulty; the top-level dictionary is optimized based on stress field continuity constraints and strain energy density constraints to obtain macroscopic features that satisfy mechanical laws, and the bottom-level dictionary is optimized through material constitutive relation constraints and stress tensor divergence constraints to obtain local detail features. A two-way constraint transfer mechanism is established between the top-level dictionary and the bottom-level dictionary. The optimization direction of the bottom-level dictionary is guided by the stress field continuity constraints and strain energy density constraints based on the top-level dictionary, and the feature expression of the top-level dictionary is supplemented by the local features of the bottom-level dictionary; the dictionary update uses the gradient descent method to optimize the composite loss function, so that each basis vector in the defect feature dictionary corresponds to defect features of different categories.

[0035] Exemplarily, the design of the deep neural network adopts an encoder-decoder structure. The encoder contains 5 convolutional blocks, and each convolutional block consists of two 3×3 convolutional layers and one 2×2 max-pooling layer. The decoder adopts a symmetric structure and contains 5 deconvolutional blocks. Each deconvolutional block contains a 2×2 upsampling layer and two 3×3 convolutional layers. Skip connections are set between the encoder and the decoder to retain high-resolution feature information.

[0036] The construction of the composite loss function integrates three types of loss terms. The data fitting loss term adopts the mean square error form and is obtained by calculating the difference between the network prediction output and the actual labeled data. In practical applications, for a 100×100 pixel defect image, the mean square error between the prediction output and the labeled data is usually controlled within 0.05. The field quantity conservation constraint loss term is calculated based on the square difference between the stress tensor divergence and the body force, where the body force includes gravity and electromagnetic force, and the typical value range is between 0.01 and 0.1. The constitutive relation constraint loss term is obtained by calculating the square difference between the stress tensor and the product of the elastic tensor and the strain tensor, and the constraint strength coefficient is set to 0.1.

[0037] The spatio-temporal feature fusion adopts a dual attention mechanism. In the temporal attention mechanism, the dimension of the learnable parameter matrix is 64×64, and the dimensions of the query matrix and the key-value matrix are both 64×128. The attention weights are obtained through matrix multiplication operations, and the weights are normalized by softmax. The spatial attention mechanism performs global average pooling and max pooling on the feature map respectively to obtain two feature maps with 256 channels, and the spatial attention weights are obtained by splicing them and passing through a 7×7 convolutional layer.

[0038] The defect feature dictionary adopts a hierarchical structure, including a top-level dictionary and a bottom-level dictionary. The top-level dictionary contains 256 basis vectors, mainly representing macroscopic features; the bottom-level dictionary contains 512 basis vectors, used to describe local detail features. During the dictionary learning process, the initial value of the dynamic learning rate is set to 0.01, and it decreases at a decay rate of 0.95 as the number of training rounds increases. The dictionary atom similarity threshold is set to 0.85, and when the cosine similarity of two basis vectors exceeds this threshold, the redundancy evaluation mechanism is triggered. The sample difficulty is determined based on the reconstruction error, and samples with larger errors obtain higher update weights.

[0039] During the optimization process of the top-level dictionary, stress field continuity constraints and strain energy density constraints are introduced. The stress field continuity constraint requires that the maximum value of the divergence of the stress tensor does not exceed the preset threshold of 0.1 MPa / mm. The strain energy density constraint ensures that the reconstruction result satisfies the principle of energy conservation, and the typical constraint strength is 0.05. The optimization of the bottom-level dictionary is achieved through material constitutive relation constraints and stress tensor divergence constraints, where the tolerance of the constitutive relation constraint is set to 2%, and the tolerance of the stress tensor divergence constraint is set to 5%.

[0040] The two-way constraint transfer mechanism is implemented through an iterative method. The physical constraints of the top-level dictionary are transferred to the bottom-level dictionary through a projection matrix, and the dimension of the projection matrix is 256×512. The local features of the bottom-level dictionary supplement the top-level dictionary through a feature aggregation method. The aggregation process uses attention weighted average, and the weights are determined by feature similarity. In actual training, the number of iterations is usually set to 50 rounds, and each round of update uses the mini-batch gradient descent method with a batch size of 32.

[0041] The present invention significantly improves the physical rationality of feature extraction by introducing a composite loss function with physical constraints; adopts a spatio-temporal feature fusion method with a dual attention mechanism to improve the comprehensiveness of feature expression; the hierarchical dictionary structure combined with the two-way constraint transfer mechanism realizes the effective fusion of macroscopic features and local details, and has good generalization ability, showing stable detection performance on pressure-bearing equipment of different models.

[0042] In an alternative embodiment, The steps of combining the data fitting loss term, the field quantity conservation constraint loss term, and the constitutive relation constraint loss term by the composite loss function through a weighting coefficient include: Construct a confidence evaluation function for the data fitting loss term based on the spatial gradient consistency between the predicted value and the true value, construct a confidence evaluation function for the field quantity conservation constraint loss term based on the local balance degree of the stress field, and construct a confidence evaluation function for the constitutive relation constraint loss term based on the parameter sensitivity of the elastic tensor; the confidence evaluation functions all adopt an adaptive scaling factor, and the adaptive scaling factor is determined by local gradient statistics and material parameter sensitivity; calculate the dynamic weighting coefficients of each loss term based on the confidence evaluation function; construct a stress-strain field gradient constraint term based on the local gradient statistics; Construct a physical coupling mechanism between loss terms based on the dynamic weighting coefficients, define the local consistency measure of the stress-strain field through the volume integral form, establish the gradient coupling strength between the field quantity conservation and the constitutive relation, and the gradient coupling strength is calculated based on the inner product of the gradient fields; construct a mutual feedback modulation function according to the gradient coupling strength, and integrate the mutual feedback modulation function into the composite loss function; Establish a constraint satisfaction evaluation mechanism with adaptive characteristic scale based on the mutual feedback modulation function, and the constraint satisfaction evaluation mechanism is calculated based on the local balance degree of the stress field and the deviation degree of the stress-strain relationship; construct a constraint violation penalty term, and the magnitude of the constraint violation penalty term is positively correlated with the difference between the constraint satisfaction and the constraint satisfaction threshold; design a gradient modulation factor based on the constraint satisfaction, and the gradient modulation factor increases as the constraint satisfaction decreases; Integrate the constraint violation penalty term and the stress-strain gradient constraint term into the composite loss function, and use the modulated gradient descent method to update the optimization parameters. The learning rate during the parameter update process is dynamically adjusted according to the constraint satisfaction and the gradient norm; when the gradient value is less than the preset gradient threshold and the constraint satisfaction is greater than the preset constraint satisfaction threshold, it is determined that the optimization converges.

[0043] Exemplarily, the construction of the composite loss function first requires designing a confidence evaluation mechanism for each loss term. The confidence evaluation of the data fitting loss term is based on the spatial gradient consistency between the predicted value and the true value, and is achieved by calculating the cosine similarity of the gradient vectors of the two. When the included angle of the gradient directions is less than 30 degrees, a high confidence is given. When the included angle is between 30 and 60 degrees, the confidence decreases linearly. When it is greater than 60 degrees, the confidence decreases significantly. The confidence evaluation of the field quantity conservation constraint loss term is determined by calculating the balance residual of the local stress field. When the residual value is less than 0.1 MPa, it corresponds to a high confidence region. When the residual value is greater than 1 MPa, it corresponds to a low confidence region. The confidence evaluation of the constitutive relation constraint loss term is based on the parameter sensitivity of the elastic tensor. When the change rate of the elastic modulus is less than 5%, it has a high confidence.

[0044] The design of the adaptive scaling factor adopts a multi-scale strategy. In terms of local gradient statistics, three feature scale windows are set, which are 5×5, 15×15, and 45×45 pixels respectively, and the gradient mean and standard deviation at each scale are calculated. The sensitivity of material parameters is obtained through parameter perturbation experiments, and the perturbation range is set to ±10% of the nominal value. The adaptive scaling factor is finally determined by the weighted combination of gradient statistics and parameter sensitivity, and the weight ratio is 7:3.

[0045] The calculation of the dynamic weighting coefficient adopts a soft threshold mechanism. For the data fitting loss term, when the confidence level is greater than 0.8, the weight coefficient is 1.0; when the confidence level is between 0.5 and 0.8, the weight linearly decays; when the confidence level is less than 0.5, the weight drops to 0.2. The weight adjustment of the field quantity conservation constraint loss term and the constitutive relation constraint loss term follows a similar rule, but the critical point positions are different. In addition, the stress-strain field gradient constraint term constructed based on local gradient statistics adopts an adaptive weight, and the weight size is positively correlated with the gradient amplitude.

[0046] The physical coupling mechanism is realized by establishing the gradient coupling strength between field quantity conservation and constitutive relations. The calculation of the gradient coupling strength adopts a sliding window method, the window size is 16×16 pixels, and the step size is 8 pixels. Within each window, the inner product of the stress field and strain field gradient vectors is calculated, and the larger the inner product value, the higher the coupling strength. The mutual feedback modulation function adopts the sigmoid form and tends to saturate when the coupling strength is greater than 0.7.

[0047] The constraint satisfaction evaluation mechanism comprehensively considers the local balance degree of the stress field and the deviation degree of the stress-strain relationship. The local balance degree is characterized by the root mean square error of the stress tensor divergence, and the error threshold is set to 0.5 MPa / mm. The deviation degree of the stress-strain relationship is evaluated based on the relative error of the constitutive equation, and the allowable error range is 3%. The constraint violation penalty term adopts an exponential form, and the penalty mechanism is activated when the constraint satisfaction is lower than the threshold of 0.8. The design of the gradient modulation factor makes the gradient gain increase by 20% for every 0.1 decrease in the constraint satisfaction.

[0048] The optimization process adopts the modulated gradient descent method, and the initial learning rate is set to 0.01. The dynamic adjustment rule of the learning rate is: when the constraint satisfaction is lower than 0.7, the learning rate is reduced to 80% of the current value; when the gradient norm is greater than the preset threshold of 2.0, the learning rate is reduced to 70% of the current value. The optimization convergence criterion is set as: the gradient norm is less than 0.01 and the constraint satisfaction is greater than 0.95.

[0049] By introducing a dynamic weighting mechanism based on confidence evaluation, the present invention significantly improves the adaptability of the loss function to different physical constraints, effectively enhances the degree of constraint satisfaction; adopts a multi-scale adaptive scaling strategy, enhances the model's expression ability on features of different scales, and improves the accuracy of feature extraction; a gradient modulation method based on a physical coupling mechanism significantly improves the convergence performance of the optimization process, speeds up the convergence speed, and the finally obtained solution has better physical rationality, and the continuity of the stress field is significantly improved. The present invention realizes the effective integration of physical constraints and deep learning, and provides a new research idea for mechanics feature extraction based on deep learning.

[0050] In an alternative embodiment, The steps of fusing the multi-scale features extracted by the deep neural network to obtain a fused feature, and calculating the feature weights by using a temporal attention mechanism and a spatial attention mechanism, wherein the temporal attention weight is obtained by the product operation of a learnable parameter matrix and a query matrix and a key-value matrix, and the spatial attention weight is obtained by concatenating and convolving the average pooling result and the max pooling result of the feature map include: Construct a stress field time series feature matrix, introduce a learnable physical constraint parameter matrix to map the original features to a feature space that satisfies the continuity of the stress field, and calculate the query matrix, key matrix and value matrix based on the physical constraint parameter matrix; combine the product result of the query matrix and the key matrix with the stress field continuity constraint term to calculate the temporal attention score, and the stress field continuity constraint term is calculated by the divergence of the stress tensor; multiply the temporal attention score by the value matrix to obtain the temporal feature output; Perform average pooling and max pooling operations on the temporal feature output respectively to obtain two feature maps, combine the concatenation result of the feature maps with the strain energy density constraint term to calculate the spatial attention weight, and the strain energy density constraint term is calculated by integrating the inner product of the stress tensor and the strain tensor over the volume, and the spatial attention weight is normalized by the sigmoid function; Fuse the features of different scale levels, multiply each layer of features by the corresponding temporal feature output and spatial attention weight respectively and sum to obtain the fused feature; Update the physical constraint parameter matrix by optimizing the composite loss function, and substitute the updated physical constraint parameter matrix into the calculation formulas of the stress field continuity constraint term and the strain energy density constraint term; the update result of the physical constraint parameter matrix is used for the calculation of the next round of feature mapping, forming an iterative process of constraint optimization.

[0051] Exemplarily, the construction of the stress field time-series feature matrix adopts the sliding time window method. The window size is set to 16 time steps, and the step size is 8 time steps. Each time step contains a feature vector of 64×64×32 dimensions. The dimension of the physical constraint parameter matrix is 32×32, and it is initialized with a normal distribution, with a mean of 0 and a standard deviation of 0.02. The feature mapping process is realized through the linear transformation of the physical constraint parameter matrix and the original features, and the output dimension remains 64×64×32.

[0052] The calculation of the query matrix, key matrix, and value matrix adopts three independent linear transformations. The dimension of the transformation matrix is 32×32 for all, and the initialization method is Xavier initialization. The calculation of the time-series attention score combines the product result of the query matrix and the key matrix and the stress field continuity constraint. The stress field continuity constraint is obtained by calculating the divergence of the stress tensor, and a higher weight is given when the divergence value is less than 0.1 MPa / mm. The time-series attention score is normalized by the softmax function to ensure that the sum of the weights of each time step is 1.

[0053] The time-series feature output undergoes processing in the spatial dimension and adopts two parallel branches. The average pooling branch uses a pooling kernel of 7×7 with a step size of 1; the max pooling branch adopts the same configuration. The outputs of the two pooling branches are concatenated to obtain a feature map with 128 channels. The strain energy density constraint is obtained by calculating the volume integral of the stress tensor and the strain tensor, and the constraint strength increases with the increase of the strain energy density. The spatial attention weight is calculated using a 7×7 convolutional kernel and is normalized to the 0-1 interval by the sigmoid function.

[0054] The feature fusion process considers five different scale levels. The dimension of the first-level feature is 64×64×32, and the dimension of the fifth-level feature is 4×4×512. The features at each level are first multiplied by the corresponding time-series feature output, and then multiplied by the spatial attention weight. The fusion weights are adopted in an adaptive manner and are dynamically adjusted according to the importance of the semantic information of the feature levels. The weight coefficients of the high-level features are relatively large.

[0055] The optimization of the physical constraint parameter matrix adopts the Adam optimizer, and the initial learning rate is set to 0.001. In each training batch, first calculate the gradient of the composite loss function, and then update the physical constraint parameter matrix. When the change rate of the loss function for five consecutive training epochs is less than 1%, the learning rate is reduced to 80% of the original. During the optimization process, the weight coefficient of the stress field continuity constraint term is set to 0.3, and the weight coefficient of the strain energy density constraint term is set to 0.2.

[0056] During the iterative optimization process, an early stopping mechanism is set. When the loss value on the validation set has not improved for 10 consecutive rounds, the training stops. The update of the physical constraint parameter matrix takes into account the momentum term, and the momentum coefficient is set to 0.9. The updated parameter matrix is used for the next round of feature mapping calculation, ensuring the continuity of constraint optimization. The optimization convergence criteria include: the relative change rate of the loss function value is less than 0.1%, and the satisfaction degrees of the stress field continuity constraint and the strain energy density constraint both reach the preset thresholds.

[0057] By introducing the spatio-temporal attention mechanism with physical constraints, the present invention improves the physical rationality of feature extraction, making the extracted features more in line with the mechanical laws; adopts the multi-scale feature fusion strategy to enhance the model's expression ability for defect features of different scales and improve the comprehensiveness of feature extraction; based on the iterative optimization mechanism of the physical constraint parameter matrix, realizes the organic combination of constraint conditions and deep learning, and significantly improves the generalization performance of the model. The present invention provides a new research idea for the application of deep learning in the field of mechanics and has important theoretical value and engineering application prospects.

[0058] In an alternative embodiment, The steps of defect localization based on the defect feature dictionary and evaluating the defect size through a deep regression network introducing physical interpretability constraints include: Perform sparse decomposition on the feature map of the sample to be tested, decompose the feature map into the sum of the product of the defect feature dictionary and the sparse coefficient and the residual term. The stress field continuity constraint calculated by the divergence of the stress tensor is introduced in the sparse decomposition process; based on the non-zero element distribution of the sparse coefficient, map the macroscopic features of the top-level dictionary and the local features of the bottom-level dictionary and perform element-wise product operations to obtain the defect probability map; Extract the initial candidate region set based on the probability threshold for the defect probability map, calculate the strain energy density of each region in the initial candidate region set, and the strain energy density is obtained through the volume integral of the stress tensor and the strain tensor; combine the strain energy density with the region area and perimeter to calculate the candidate region score, and screen out the final defect localization result based on the preset region score threshold; Construct a deep regression network with an encoder and a decoder structure. The encoder is composed of multiple residual blocks with dilated convolutions, and the decoder adopts a progressive upsampling structure; input the original feature map, the defect region mask, and the defect localization result into the deep regression network; A feature extraction module, a geometric parameter regression module, and a physical interpretability constraint module are set in the depth regression network. The feature extraction module extracts multi-scale features through residual blocks and performs feature fusion using an attention mechanism. The geometric parameter regression module outputs the length, width, and depth parameters of the defect through independent fully connected layers. The physical interpretability constraint module calculates the stress distribution and strain distribution based on the predicted geometric parameters. A multi-task loss function including a geometric parameter regression loss term, a stress field constraint loss term, a strain field constraint loss term, and an energy conservation constraint loss term is constructed. The stress field constraint loss term is calculated based on the divergence of the stress tensor. The strain field constraint loss term is calculated based on the gradient of the strain tensor. The energy conservation constraint loss term is calculated based on the difference between the system energy calculated from the original feature map and the system energy calculated after reconstructing the predicted defect geometric parameters. The depth regression network is trained by optimizing the multi-task loss function to obtain a defect size evaluation result that satisfies the physical interpretability constraint. The defect size evaluation result includes the length, width, and depth parameters of the defect.

[0059] Exemplarily, the sparse decomposition of the feature map uses the iterative shrinkage threshold algorithm to decompose the 256×256-dimensional feature map into the product form of a defect feature dictionary and sparse coefficients. The defect feature dictionary contains 768 basis vectors, including 256 top-level dictionary vectors and 512 bottom-level dictionary vectors. The stress field continuity constraint requires that the divergence of the stress tensor is less than 0.1 MPa / mm. The proportion of non-zero elements of the sparse coefficients is usually controlled within 10% to ensure the sparsity of the solution.

[0060] In the process of generating the defect probability map, the features of the top-level dictionary and the bottom-level dictionary are first mapped and aligned. The mapping is implemented using a 1×1 convolutional layer to unify the feature dimensions, and the number of output channels is set to 128. The mapped features are subjected to an element-wise product operation and normalized by the sigmoid function to obtain probability values. For a 100×100 pixel detection area, the probability threshold is typically set to 0.75.

[0061] The initial candidate regions are screened using the connected component analysis method. The strain energy density of each candidate region is calculated, and the calculation process considers the integral value of the stress tensor and the strain tensor within the region. The region score is calculated by comprehensively considering three factors: strain energy density, region area, and perimeter, with a weight ratio of 5:3:2. The region score threshold is set to 0.8, and regions with a score higher than this threshold are determined as the final defect positions.

[0062] The encoder of the depth regression network contains 5 residual blocks, and each residual block contains 3 convolutional layers with dilated convolutions, and the dilation rates are 1, 2, and 4 respectively. The decoder adopts 3 progressive upsampling modules, and each module doubles the size of the feature map. The network input includes an original feature map of 256×256×3, a defect area mask of 256×256×1, and a defect location result.

[0063] The feature extraction module uses a multi-scale feature extraction strategy. The number of channels in the residual block increases from 64 to 512, and the size of the feature map decreases from 256×256 to 16×16. The attention mechanism adopts a combination of channel attention and spatial attention. The channel attention weights are calculated through global average pooling and max pooling, and the spatial attention weights are obtained through a 7×7 convolutional layer.

[0064] The geometric parameter regression module contains three parallel fully connected layers, which respectively output the length, width, and depth parameters of the defect. A global average pooling layer is added before the fully connected layer to reduce the feature dimension. The physically interpretable constraint module constructs a finite element model based on the predicted geometric parameters and calculates the stress distribution and strain distribution.

[0065] The multi-task loss function is constructed in a weighted combination manner. The geometric parameter regression loss term uses smooth L1 loss, and the weight coefficient is 0.4. The stress field constraint loss term is calculated based on the divergence of the stress tensor, and the weight coefficient is 0.2. The strain field constraint loss term is obtained by calculating the gradient of the strain tensor, and the weight coefficient is 0.2. The energy conservation constraint loss term compares the difference between the energy of the original system and the reconstructed system, and the weight coefficient is 0.2.

[0066] The network is trained using the Adam optimizer, with an initial learning rate of 0.001, which decays to 90% of the original every 50 epochs. The training batch size is set to 32, and the total number of training rounds is 200. When the loss value on the validation set does not decrease for 10 consecutive rounds, the early stopping mechanism is triggered. The weights of each loss term are dynamically adjusted during the training process and adaptively adjusted according to the performance of the validation set.

[0067] By introducing a sparse decomposition method with physical constraints, the present invention improves the accuracy and reliability of defect location; adopts a multi-task learning framework combined with physically interpretable constraints to enhance the physical rationality of the defect size evaluation results; and realizes the unified optimization of defect detection and size evaluation based on the end-to-end training method of the depth regression network, significantly improving the overall performance of the system. The present invention provides a new technical route for the intelligent development of the non-destructive testing field and has important engineering application value.

[0068] Figure 2 For the structural schematic diagram of the multi-physical field defect detection system based on magnetostriction of the pressure-bearing equipment in the embodiment of the present invention, as Figure 2 shown, the system includes: The first unit is used to construct an intelligent excitation and sensing device, and the intelligent excitation and sensing device includes a reconfigurable magnetic field excitation array and a multi-physical quantity sensing network; the reconfigurable magnetic field excitation array adopts electromagnetic coils with a matrix layout to form a focused magnetic field through adaptive phase modulation, and the adaptive phase modulation dynamically adjusts the excitation phases of adjacent coils based on the geometric features of the area to be detected and the magnetoresponse signals collected in real time; the multi-physical quantity sensing network includes a double-layer orthogonally arranged GMR sensor array, a temperature sensor array, and a stress sensor array for collecting magnetic field signals, temperature signals, and stress signals; The second unit is used to construct a hierarchical coupling calculation scheme, and establish the coupling relationship between the magnetic field, stress field, and temperature field based on the magnetic field signal, temperature signal, and stress signal; the magnetic field and stress field are coupled through a magnetostrictive constitutive equation including a stress-induced magnetization term and a magnetostrictive strain term, and the temperature field is coupled with the stress field by establishing a piecewise linearized mapping relationship of the influence of temperature on the magnetization characteristics and mechanical characteristics of materials; introduce magnetic field divergence constraints, mechanical equilibrium constraints, and heat conduction boundary constraints as physical constraint conditions, and use an adaptive grid meshing method based on the gradient of the physical field for calculation to obtain multi-physical field coupling data; The third unit is used to denoise the multi-physical field coupling data by using multi-level wavelet packet decomposition based on the energy entropy criterion; input the denoised data into a deep neural network, introduce field quantity conservation constraints and constitutive relationship constraints into the loss function of the deep neural network, and use a spatio-temporal feature fusion method with an attention mechanism to fuse the multi-scale features extracted by the deep neural network to generate a defect feature dictionary; perform defect localization based on the defect feature dictionary, and evaluate the defect size through a deep regression network introducing physical interpretability constraints.

[0069] In the third aspect of the embodiments of the present invention, A kind of electronic device is provided, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0070] In the fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0071] The present invention can be a method, device, system, and / or computer program product. The computer program product can include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements 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 multi-physical field defect detection method for pressure-bearing equipment based on magnetostriction, characterized in that: include: Constructing an intelligent excitation and sensing device, wherein the intelligent excitation and sensing device includes a reconfigurable magnetic field excitation array and a multi-physical quantity sensing network; The reconfigurable magnetic field excitation array uses electromagnetic coils arranged in a matrix, and forms a focusing magnetic field through adaptive phase modulation. The adaptive phase modulation dynamically adjusts the excitation phase of adjacent coils based on the geometric characteristics of the area to be detected and the magnetic field response signal collected in real time; the multi-physical quantity sensing network includes a double-layer orthogonally arranged GMR sensor array, a temperature sensor array and a stress sensor array, which are used to collect magnetic field signals, temperature signals and stress signals; A hierarchical coupling calculation scheme is constructed to establish a coupling relationship between the magnetic field, stress field and temperature field based on the magnetic field signal, temperature signal and stress signal; the magnetic field and stress field are coupled through a magnetostrictive constitutive equation containing a stress-induced magnetization term and a magnetostrictive strain term, and the temperature field is coupled with the stress field by establishing a piecewise linearized mapping relationship of the influence of temperature on the magnetic properties and mechanical properties of the material; magnetic field divergence constraints, mechanical equilibrium constraints and heat conduction boundary constraints are introduced as physical constraints, and an adaptive meshing method based on physical field gradients is used for calculation to obtain multi-physical field coupling data; The multi-physics field coupling data is denoised by using multi-level wavelet packet decomposition based on the energy entropy criterion; the denoised data is input into a deep neural network, the loss function of the deep neural network introduces field quantity conservation constraints and constitutive relationship constraints, and the multi-scale features extracted by the deep neural network are fused by using a spatiotemporal feature fusion method of an attention mechanism to generate a defect feature dictionary; Defect localization is performed based on the defect feature dictionary, and the defect size is evaluated by a deep regression network that introduces physical interpretability constraints.

2. The method according to claim 1, characterized in that A hierarchical coupling calculation scheme is constructed, and a coupling relationship between the magnetic field, stress field and temperature field is established based on the magnetic field signal, temperature signal and stress signal; the magnetic field and stress field are coupled through a magnetostrictive constitutive equation containing a stress-induced magnetization term and a magnetostrictive strain term, and the temperature field is coupled with the stress field by establishing a piecewise linearized mapping relationship of the influence of temperature on the magnetic properties and mechanical properties of the material; a magnetic field divergence constraint, a mechanical equilibrium constraint and a heat conduction boundary constraint are introduced as physical constraint conditions, and an adaptive meshing method based on a physical field gradient is used for calculation. The steps of obtaining multi-physical field coupling data include: A hierarchical coupling framework with stress field as the core is established, and a coupling strength evaluation function is constructed by introducing a coupling weight coefficient based on the strength of the physical field. The coupling weight coefficient is determined according to the characteristic strength of the magnetic field, temperature field and stress field. The characteristic strength is characterized by the gradient and energy density of each physical field. The coupling strength evaluation function is used to dynamically adjust the iteration strategy. In the layered coupling framework, a magnetic field-stress field intensity coupling layer is constructed, and an improved Jiles-Atherton model including stress-induced magnetization terms and magneto-strain terms is used to describe the bidirectional coupling relationship. The magneto-strain is corrected by introducing a nonlinear magnetization intensity dependency term, and the stress-induced magnetization term takes into account the influence of the effective magnetic field and the stress sensitivity coefficient; a temperature field-stress field weak coupling layer is established, and a piecewise linearization method is used to describe the influence of temperature on the magnetization characteristics and mechanical properties of the material. An improved Hermite interpolation function is introduced as a smooth transition function at the temperature critical point, and the improved Hermite interpolation function includes a normalized temperature variable and an endpoint derivative value; A physical constraint condition system is constructed based on the hierarchical coupling framework, including magnetic field divergence constraints, mechanical equilibrium constraints and heat conduction boundary constraints. A coupling source term is introduced into the heat conduction boundary constraint, and the coupling source term includes hysteresis loss, magnetoelastic coupling loss and eddy current loss. A grid density function is constructed according to physical field gradient information, and grid units are divided based on the grid density function. The output value of the coupling strength evaluation function is weightedly combined with the grid density function to form a composite density function. The composite density function is used to determine the local encryption coefficient of the grid unit, and the size of the grid unit is dynamically adjusted according to the local encryption coefficient.

3. The method according to claim 2, characterized in that The steps of constructing a grid density function according to the physical field gradient information, dividing the grid units based on the grid density function, weightedly combining the output value of the coupling strength evaluation function with the grid density function to form a composite density function, using the composite density function to determine the local encryption coefficient of the grid unit, and dynamically adjusting the size of the grid unit according to the local encryption coefficient include: Constructing a grid density function according to the physical field gradient information, wherein the grid density function determines the weight coefficient of each physical field through the maximum variation of the magnetic field intensity gradient, the stress gradient and the temperature gradient in the calculation domain; Performing a weighted combination of the output value of the coupling strength evaluation function and the grid density function to form a composite density function, wherein a balance coefficient of the weighted combination is given by an exponential decay function determined by the output value of the coupling strength evaluation function; The local encryption coefficient of the grid unit is determined by using the composite density function, and a smoothing factor is introduced to smooth the local encryption coefficient of the grid unit. The smoothing factor performs a weighted combination of the currently calculated local encryption coefficient and the local encryption coefficient of the previous iteration. The value of the smoothing factor is positively correlated with the deformation rate of the grid unit. When the deformation rate of the grid unit is greater than a preset deformation rate threshold, the weight coefficient of the historical information increases proportionally. When the deformation rate of the grid unit is less than the preset deformation rate threshold, the weight coefficient of the current information increases accordingly. The size of the grid unit is dynamically adjusted according to the local encryption coefficient, and when the size ratio of adjacent grid units exceeds a set size ratio threshold or the shape quality index of the grid unit is less than a set quality threshold, a Laplacian smoothing method is used to perform local re-division; in the process of dynamic adjustment of the size of the grid unit, a time step is determined based on the ratio of the local deformation speed of the grid unit to the characteristic size of the grid unit, and the time step decreases as the deformation degree of the grid unit increases; A grid unit optimization index based on strain energy density is constructed. When the grid unit optimization index is less than the set optimization threshold, the corresponding grid unit is marked as an area to be removed. The local reconstruction mechanism of the grid unit is triggered in the high-coupling area, including identifying the high-stress area nodes, reconstructing the local grid units based on the Delaunay criterion, and using the spline interpolation method to restore the physical field information.

4. The method according to claim 1, characterized in that The denoised data is input into a deep neural network, the loss function of the deep neural network introduces field quantity conservation constraints and constitutive relationship constraints, and the spatiotemporal feature fusion method of the attention mechanism is used to fuse the multi-scale features extracted by the deep neural network. The steps of generating a defect feature dictionary include: Input the denoised data into a deep neural network, and construct a composite loss function including a data fitting loss term, a field quantity conservation constraint loss term, and a constitutive relationship constraint loss term, wherein the data fitting loss term is calculated by the mean square error between the network prediction output and the actual labeled data, the field quantity conservation constraint loss term is calculated based on the square difference between the stress tensor divergence and the body force, and the constitutive relationship constraint loss term is calculated based on the square difference between the stress tensor and the product of the elastic tensor and the strain tensor, and the composite loss function combines the data fitting loss term, the field quantity conservation constraint loss term, and the constitutive relationship constraint loss term through a weighting coefficient; The multi-scale features extracted by the deep neural network are fused to obtain fused features, and the feature weights are calculated using a temporal attention mechanism and a spatial attention mechanism, wherein the temporal attention weight is obtained by multiplying a learnable parameter matrix with a query matrix and a key value matrix, and the spatial attention weight is obtained by concatenating and convolving the average pooling result and the maximum pooling result of the feature map; Based on the fusion features, a defect feature dictionary with a hierarchical structure is constructed through sparse coding, and an online dictionary learning algorithm is used for iterative updating. The defect feature dictionary includes a top-level dictionary for characterizing macroscopic features and a bottom-level dictionary for characterizing local features. The dictionary updating process includes: dynamic learning rate adjustment based on defect feature sparsity, redundancy evaluation based on dictionary atomic similarity, and update weight allocation based on sample difficulty; the top-level dictionary is optimized based on stress field continuity constraints and strain energy density constraints to obtain macroscopic features that meet mechanical laws, and the bottom-level dictionary is optimized through material constitutive relationship constraints and stress tensor divergence constraints to obtain local detail features. A two-way constraint transmission mechanism is established between the top-level dictionary and the bottom-level dictionary, and the optimization direction of the bottom-level dictionary is guided by the stress field continuity constraints and strain energy density constraints based on the top-level dictionary, and the feature expression of the top-level dictionary is supplemented by the local features of the bottom-level dictionary; the dictionary update uses a gradient descent method to optimize the composite loss function so that each basis vector in the defect feature dictionary corresponds to a defect feature of a different category.

5. The method according to claim 4, characterized in that The step of combining the data fitting loss term, the field quantity conservation constraint loss term and the constitutive relationship constraint loss term by the composite loss function through weighted coefficients includes: A confidence evaluation function for the data fitting loss term is constructed based on the spatial gradient consistency between the predicted value and the true value, a confidence evaluation function for the field quantity conservation constraint loss term is constructed based on the local balance degree of the stress field, and a confidence evaluation function for the constitutive relationship constraint loss term is constructed based on the parameter sensitivity of the elastic tensor; the confidence evaluation functions all use an adaptive scaling factor, which is determined by local gradient statistics and material parameter sensitivity; the dynamic weighting coefficient of each loss term is calculated based on the confidence evaluation function; and the stress-strain field gradient constraint term is constructed based on the local gradient statistics; Based on the dynamic weighting coefficient, a physical coupling mechanism between loss terms is constructed, a local consistency measure of the stress-strain field is defined in the form of a volume integral, and a gradient coupling strength of field quantity conservation and constitutive relationship is established, wherein the gradient coupling strength is calculated based on the inner product of the gradient field; a mutual feedback modulation function is constructed according to the gradient coupling strength, and the mutual feedback modulation function is integrated into the composite loss function; A characteristic scale adaptive constraint satisfaction evaluation mechanism is established based on the mutual feedback modulation function, and the constraint satisfaction evaluation mechanism is calculated based on the local equilibrium degree of the stress field and the degree of deviation of the stress-strain relationship; a constraint violation penalty term is constructed, and the size of the constraint violation penalty term is positively correlated with the difference between the constraint satisfaction and the constraint satisfaction threshold; a gradient modulation factor based on the constraint satisfaction is designed, and the gradient modulation factor increases as the constraint satisfaction decreases; The constraint violation penalty term and the stress-strain gradient constraint term are integrated into the composite loss function, and the modulated gradient descent method is used to update the optimization parameters. The learning rate in the parameter updating process is dynamically adjusted according to the constraint satisfaction and the gradient norm. The optimization convergence is determined when the gradient value is less than the preset gradient threshold and the constraint satisfaction is greater than the preset constraint satisfaction threshold.

6. The method according to claim 4, characterized in that The multi-scale features extracted by the deep neural network are fused to obtain fused features, and the feature weights are calculated using a temporal attention mechanism and a spatial attention mechanism, wherein the temporal attention weights are obtained by multiplying a learnable parameter matrix with a query matrix and a key value matrix, and the spatial attention weights are obtained by concatenating and convolving the average pooling results and the maximum pooling results of the feature graphs, including the following steps: Construct a stress field temporal feature matrix, introduce a learnable physical constraint parameter matrix to map the original features to a feature space that satisfies the continuity of the stress field, calculate a query matrix, a key matrix, and a value matrix based on the physical constraint parameter matrix; combine the product of the query matrix and the key matrix with the stress field continuity constraint term to calculate the temporal attention score, and the stress field continuity constraint term is calculated by the stress tensor divergence; multiply the temporal attention score by the value matrix to obtain a temporal feature output; The time series feature output is subjected to average pooling and maximum pooling operations to obtain two feature maps, and the splicing results of the feature maps are combined with the strain energy density constraint term to calculate the spatial attention weight, the strain energy density constraint term is calculated by integrating the inner product of the stress tensor and the strain tensor over the volume, and the spatial attention weight is normalized by the sigmoid function; The features of different scale levels are fused, and the features of each layer are multiplied by the corresponding temporal feature output and spatial attention weight and summed to obtain the fused features; The physical constraint parameter matrix is ​​updated by optimizing the composite loss function, and the updated physical constraint parameter matrix is ​​substituted into the calculation formula of the stress field continuity constraint term and the strain energy density constraint term; the updated result of the physical constraint parameter matrix is ​​used for the calculation of the next round of feature mapping, forming an iterative process of constraint optimization.

7. The method according to claim 1, characterized in that The steps of locating defects based on the defect feature dictionary and evaluating defect sizes by introducing a deep regression network with physical interpretability constraints include: Performing sparse decomposition on the feature map of the sample to be tested, decomposing the feature map into the product of the defect feature dictionary and the sparse coefficient and the sum of the residual term, and introducing the stress field continuity constraint of the stress tensor divergence calculation in the sparse decomposition process; based on the non-zero element distribution of the sparse coefficient, mapping the macro features of the top dictionary with the local features of the bottom dictionary and performing element-by-element product operation to obtain a defect probability map; Extracting an initial candidate region set from the defect probability map based on a probability threshold, calculating the strain energy density of each region in the initial candidate region set, wherein the strain energy density is obtained by volume integration of the stress tensor and the strain tensor; combining the strain energy density with the region area and perimeter to calculate the candidate region score, and screening out the final defect location result based on a preset region score threshold; Constructing a deep regression network with an encoder and a decoder structure, wherein the encoder is composed of a plurality of residual blocks with dilated convolutions, and the decoder adopts a progressive upsampling structure; inputting the original feature map, the defect area mask and the defect location result into the deep regression network; A feature extraction module, a geometric parameter regression module and a physical interpretability constraint module are set in the deep regression network, the feature extraction module extracts multi-scale features through a residual block and uses an attention mechanism to perform feature fusion, the geometric parameter regression module outputs the length, width and depth parameters of the defect through an independent fully connected layer, and the physical interpretability constraint module calculates stress distribution and strain distribution based on the predicted geometric parameters; A multi-task loss function is constructed, which includes a geometric parameter regression loss term, a stress field constraint loss term, a strain field constraint loss term and an energy conservation constraint loss term, wherein the stress field constraint loss term is calculated based on the stress tensor divergence, the strain field constraint loss term is calculated based on the strain tensor gradient, and the energy conservation constraint loss term is calculated based on the difference between the system energy calculated based on the original feature map and the system energy calculated after the predicted defect geometric parameters are reconstructed; the deep regression network is trained by optimizing the multi-task loss function to obtain a defect size assessment result that meets the physical interpretability constraint, and the defect size assessment result includes the length, width and depth parameters of the defect.

8. A multi-physical field defect detection system for pressure-bearing equipment based on magnetostriction, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to construct an intelligent excitation and sensing device, wherein the intelligent excitation and sensing device includes a reconfigurable magnetic field excitation array and a multi-physical quantity sensing network; The reconfigurable magnetic field excitation array uses electromagnetic coils arranged in a matrix, and forms a focusing magnetic field through adaptive phase modulation. The adaptive phase modulation dynamically adjusts the excitation phase of adjacent coils based on the geometric characteristics of the area to be detected and the magnetic field response signal collected in real time; the multi-physical quantity sensing network includes a double-layer orthogonally arranged GMR sensor array, a temperature sensor array and a stress sensor array, which are used to collect magnetic field signals, temperature signals and stress signals; The second unit is used to construct a hierarchical coupling calculation scheme, and establish the coupling relationship between the magnetic field, stress field and temperature field based on the magnetic field signal, temperature signal and stress signal; the magnetic field and stress field are coupled through the magnetostrictive constitutive equation containing stress-induced magnetization terms and magnetostrictive strain terms, and the temperature field is coupled with the stress field by establishing a piecewise linear mapping relationship of the influence of temperature on the magnetic properties and mechanical properties of the material; the magnetic field divergence constraint, mechanical equilibrium constraint and heat conduction boundary constraint are introduced as physical constraint conditions, and the adaptive meshing method based on the physical field gradient is used for calculation to obtain multi-physical field coupling data; The third unit is used to reduce noise on the multi-physics field coupling data by using multi-level wavelet packet decomposition based on energy entropy criterion; the reduced noise data is input into a deep neural network, the loss function of the deep neural network introduces field quantity conservation constraints and constitutive relationship constraints, and the spatiotemporal feature fusion method of the attention mechanism is used to fuse the multi-scale features extracted by the deep neural network to generate a defect feature dictionary; Defect localization is performed based on the defect feature dictionary, and the defect size is evaluated by a deep regression network that introduces physical interpretability constraints.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images