A structural state dynamic modeling method based on BIM visualization technology

Abstract

This invention discloses a method for dynamic structural state modeling based on BIM visualization technology, relating to the fields of BIM and 3D visualization technology. The method includes the following steps: S1, constructing a dynamic structural state modeling base; S2, generating a spatially distributed strain dataset; S3, based on an improved DeepONet model, introducing a parasitic bee chemotaxis-homing mechanism to perform directional walking and aggregation to compensate for local abrupt changes, and combining this with physical equilibrium residual constraints to generate a sharpened continuous strain tensor field for the outer skin region without sensor placement; S4, outputting state driving factors; S5, outputting yield state labels; S6, outputting deformation textures; S7, synchronously mapping the deformation textures and tearing effects to the BIM visualization model. This invention overcomes the limitations of traditional methods, such as ignoring physical evolution laws, smoothing distortion in unseen regions, and lack of stress concentration constraints in crack rendering, providing a precise solution for dynamic structural state modeling using BIM.

Description

Technical Field

[0001] This invention relates to the field of BIM and 3D visualization technology, and in particular to a method for dynamic modeling of structural states based on BIM visualization technology. Background Technology

[0002] With the widespread application of BIM visualization technology in the engineering field, the deep integration of structural health monitoring data and 3D models faces severe challenges in real-time representation and accurate mapping. Existing BIM structural condition visualization methods typically employ discrete color band mapping or preset deformation scripts to drive model rendering. This method, based solely on discrete threshold rendering, ignores the continuous mechanical response distribution of the structure under external loads (such as the anisotropic characteristics of the strain tensor field) and the physical evolution of material critical failure, leading to geometric distortion and physical discontinuity in the visualization of the mechanical state, thus limiting the accurate representation of the progressive yielding process of the structure. Furthermore, existing methods often struggle to fully utilize the sensor spatial topology and material mechanical parameters to constrain the spatial extrapolation of the strain field, resulting in artifacts of smooth transitions in areas where no sensors are deployed. This masks the local abrupt changes caused by stress concentration and fails to dynamically evolve tear cracks based on the actual stress concentration coefficient, increasing the risk of misjudging structural failure.

[0003] Therefore, how to provide a method for dynamic modeling of structural states based on BIM visualization technology is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes a dynamic structural state modeling method based on BIM visualization technology. Through a closed-loop feedback optimization process based on physical equilibrium residual constraints and parasitic wasp chemotaxis-homing mechanism, it performs automatic differential derivation of approximate stress tensors on the sharpened continuous strain tensor field, evaluates mechanical equilibrium residuals, and generates force field vector equations by fusing the gradient of the pseudo-pheromone concentration field with the spatial distance to the probe tensor kernel. Multi-step directional walks are then performed using differentiable Euler recursion to extract local mutation compensation features. The physical equilibrium residuals are used as a penalty term for backpropagation constraints to improve the DeepONet model parameter iteration and the learnable parameter updates of the parasitic wasp probe tensor kernel. This mechanism, by establishing a closed-loop feedback path from "mechanical residual measurement" to "operator network parameters," effectively eliminates physical deviations in the spatial extrapolation of the strain field, ensuring that the improved DeepONet model can dynamically maintain the mechanical equilibrium characteristics of the outer skin region without sensor placement. This achieves the technical effect of generating a physically self-consistent sharpened continuous strain tensor field while compensating for local mutation features. This invention overcomes the limitations of traditional methods, such as ignoring physical evolution laws, smoothing distortion in undefined areas, and lack of stress concentration constraints in crack rendering, and provides a precise solution for dynamic modeling of BIM structural states.

[0005] A method for dynamic structural state modeling based on BIM visualization technology according to an embodiment of the present invention specifically includes:

[0006] S1. Extract the geometric data and structural attribute information of the outer skin of the BIM visualization model and overlay it with regular textures. Calculate the initial standard UV coordinate set, bind the local coordinate system of the mesh patch and the material mechanical parameters, and construct the dynamic modeling base of the structural state.

[0007] S2. Based on the dynamic modeling of structural state, collect time-series data of strain on the surface of the structure, and perform spatial topology mapping between the sensor spatial coordinates and the outer skin mesh to generate a spatially distributed strain dataset.

[0008] S3. Based on the improved DeepONet model, the spatially distributed strain dataset and material mechanical parameters are used as dual-domain constraints for joint encoding and operator mapping to generate the initial strain field and apply affine modulation of mechanical parameters. The parasitic bee chemotactic home-seeking mechanism is introduced to perform directional walking and aggregation to compensate for local abrupt changes. Combined with physical equilibrium residual constraints, a sharpened continuous strain tensor field is generated for the area of ​​the outer skin where no sensors are deployed.

[0009] S4. Extract strain features from the sharpened continuous strain tensor field, map and calculate the UV coordinate torsion offset, and convert the critical parameters of the material mechanical parameters based on the BIM structural attribute information into state driving factors.

[0010] S5. Modulate the UV coordinate distortion offset based on the state driving factor to correct the initial standard UV coordinate set to generate the state-marked UV coordinate set, and combine the state driving factor to map and output the yield state label.

[0011] S6. Based on the structural state dynamic modeling base and yield state label, drive deformation and fragment coloring, and perform twist resampling on the adaptive texture pattern with the state label UV coordinate set to output the deformation texture.

[0012] S7. Based on the yield state label and the critical parameters of the material mechanical parameters, perform over-limit judgment, perform separation offset on the mesh line features along the deformation texture of the over-limit area to generate cracks, render tearing effects, and synchronously map the deformation texture and tearing effects to the BIM visualization model.

[0013] Optionally, S1 specifically includes:

[0014] S11. Extract the vertex coordinates and topological connection relationships of the triangular mesh of the outer skin of the BIM visualization model, calculate the area and normal vector of the mesh patch, and construct the three-dimensional geometric feature set of the outer skin.

[0015] S12. Based on the three-dimensional geometric feature set of the outer skin, perform parametric unfolding on the three-dimensional mesh surface, calculate the affine mapping matrix from the two-dimensional parameter domain to the three-dimensional spatial surface and the local deformation gradient during the unfolding process, and establish an undistorted initial standard UV coordinate set.

[0016] S13. The area change distribution characteristics and local deformation gradient distribution characteristics generated by statistical parameterization are used to calculate the spatial scaling factor and anisotropic rotation angle of the texture primitives based on the area change distribution characteristics and local deformation gradient distribution characteristics. An adaptive texture pattern is generated based on the spatial scaling factor and anisotropic rotation angle and mapped to the three-dimensional mesh surface of the outer skin along the two-dimensional parameter direction of the initial standard UV coordinate set to generate a visual representation of the outer skin covered with adaptive regular texture.

[0017] S14. Calculate the local tangent space basis vectors within each mesh patch, construct a local coordinate system composed of orthogonal tangent vectors and normal vectors, and store it in association with the corresponding mesh patch;

[0018] S15. Extract the elastic modulus and Poisson's ratio, as well as the thickness attribute of the mesh surface in the BIM model, and bind the elastic modulus, Poisson's ratio and thickness attribute together with the local coordinate system to the corresponding mesh surface to generate mesh surface features carrying spatial mechanical properties. Summarize and construct the dynamic modeling base of the structural state.

[0019] Optionally, S2 specifically includes:

[0020] S21. Based on the structural state dynamic modeling basis, collect the strain time series data and spatial coordinates of the outer skin sensor nodes;

[0021] S22. Based on spatial coordinates and the three-dimensional geometric feature set of the outer skin, calculate the spatial closest distance and normal projection deviation from the sensor node to the mesh patch, and perform spatial topology mapping to generate node-pattern mapping pairs.

[0022] S23. Based on the node-pattern mapping pair, the strain time series data of the sensor node is transferred to the corresponding mesh patch to generate the mesh patch strain time series sequence;

[0023] S24. Statistically analyze the amplitude distribution characteristics and frequency band energy distribution characteristics of the strain time series of the grid patch, and calculate the adaptive denoising threshold and filtering bandwidth of the strain time series data;

[0024] S25. Based on the adaptive denoising threshold and filtering bandwidth, perform adaptive filtering on the strain time series of grid patches to generate a spatially distributed strain dataset.

[0025] Optionally, the improved DeepONet model includes a sparse boundary coding branch network, a continuous spatial mapping backbone network, a mechanical parameter modulation layer, a chemotactic aggregation enhancement layer, and a physical constraint residual verification layer.

[0026] The sparse boundary coding branch network is used to concatenate the spatially distributed strain dataset with the sensor spatial coordinates to form a sensor input tensor, and perform joint coding with material mechanical parameters as conditions to output a global boundary constraint feature vector.

[0027] The continuous spatial mapping backbone network is used to encode the target spatial coordinates of the area on the outer skin where no sensors are deployed into a spatial query tensor, perform nonlinear feature extraction, and output a spatial basis function tensor.

[0028] The mechanical parameter modulation layer is used to perform an inner product operation between the global boundary constraint feature vector and the spatial basis function tensor to generate an initial continuous strain tensor field; an anisotropic scaling factor is constructed based on the material mechanical parameters, and an affine transformation is applied to the initial continuous strain tensor field to output a modulated continuous strain tensor field.

[0029] The chemotactic aggregation enhancement layer is used to introduce the parasitic wasp chemotactic host-seeking mechanism in the field of insect behavior, and the specific execution process includes:

[0030] The second-order spatial gradient norm is calculated for the modulated continuous strain tensor field, local feature change rates are extracted and mapped to a non-negative quasi-pheromone concentration field; samples are taken along the spatial distribution of the modulated continuous strain tensor field, and a set of learnable parasitic wasp detection tensor kernels are initialized.

[0031] A wandering attraction potential field is constructed based on the gradient direction of the pseudo-pheromone concentration field, and a wandering repulsion force field is constructed based on the spatial distance metric between the parasitic wasp detection tensor kernels. The wandering attraction potential field and the wandering repulsion force field are fused to generate the force field vector equation.

[0032] Based on the force field vector equation, a multi-step directional walk is performed according to the differentiable Euler recursion, causing the parasitic wasp detector tensor kernel to gather towards the pheromone extremum point, i.e. the singular host region. During the multi-step directional walk, local mutation features are absorbed as compensation features. The density distribution of the gathered parasitic wasp detector tensor kernel is transformed into a spatial mask. The modulated continuous strain tensor field is subjected to a dot product masking. The compensation features are aligned and superimposed on the singular host region according to the mask, and the sharpened continuous strain tensor field is output.

[0033] The physical constraint residual verification layer is used to perform automatic differentiation on the sharpened continuous strain tensor field and derive the approximate stress tensor in combination with material mechanics parameters. It evaluates the equilibrium residual of stress and mechanical equilibrium state, and uses it as a penalty term to backpropagate the constraint network parameter iteration and the learnable parameter update of the parasitic bee detection tensor kernel, outputting the sharpened continuous strain tensor field.

[0034] Optionally, S4 specifically includes:

[0035] S41. Based on the sharpened continuous strain tensor field, the tensor is diagonalized and decomposed. The principal direction is determined with the direction of zero shear strain as the reference, and the principal value is determined with the order of extreme values ​​of diagonal elements as the reference. The strain characteristic parameters are then combined to generate the strain characteristic parameters.

[0036] S42. Based on strain characteristic parameters and local deformation gradient, update the calculation parameter domain torsion vector and calculate the UV coordinate torsion offset.

[0037] S43. Based on the dynamic modeling basis of structural state, extract the elastic modulus, Poisson's ratio and thickness attributes, calculate the current equivalent stress of the mesh surface by combining the strain characteristic parameters, and use the thickness attribute to correct the equivalent stress distribution to generate a spatial stress field.

[0038] S44. Based on the elastic modulus and Poisson's ratio, the critical yield stress of the material is calculated by combining the strain characteristic parameters; the spatial stress ratio is calculated by statistically analyzing the spatial stress field and the critical yield stress, and the adaptive deformation threshold is determined based on the distribution characteristics of the spatial stress ratio.

[0039] S45. Perform normalization mapping on the UV coordinate distortion offset based on the adaptive deformation threshold to generate the state driving factor.

[0040] Optionally, S5 specifically includes:

[0041] S51. Perform dynamic modulation correction on the UV coordinate distortion offset based on the state driving factor to generate the state driving offset.

[0042] S52. Retrieve the initial standard UV coordinate set, perform algebraic superposition on the initial standard UV coordinate set based on the state-driven offset, and generate the state-marked UV coordinate set.

[0043] S53. Extract the driving offset features from the UV coordinate set of the state markers, combine them with the state driving factor to define the yield boundary threshold, perform interval mapping, and generate the yield state label of the mesh patch.

[0044] Optionally, S6 specifically includes:

[0045] S61. Based on the features of the mesh facets carrying spatial mechanical properties in the dynamic modeling base of structural state, the mesh deformation is driven by the mesh facet features. The texture sampling coordinates after deformation are updated by replacing the initial standard UV coordinate set with the state marker UV coordinate set. The fragment coloring mapping relationship is constructed by combining the yield state label and the fragment coloring is performed to generate yield visualization fragments.

[0046] S62. Calculate the resampling interpolation weights based on the statistical texture sampling density distribution characteristics of the state marker UV coordinate set;

[0047] S63. Perform distortion resampling on the adaptive texture pattern based on the resampling interpolation weight to generate a deformed texture.

[0048] Optionally, S7 specifically includes:

[0049] S71. Based on the yield state label, perform barycenter coordinate interpolation in the rasterization stage to output a fragment-level continuous yield state field.

[0050] S72. Extract the fragment yield parameter and fragment texture coordinates based on the fragment-level continuous yield state field, obtain the material critical yield judgment parameter, calculate the characteristic distance of the grid line from the fragment texture coordinate to the deformation texture as the stress concentration coefficient, combine the stress concentration coefficient and the fragment yield parameter to correct the failure comparison value, compare the failure comparison value and the material critical yield judgment parameter to output the over-limit judgment result.

[0051] S73. Based on the result of the over-limit judgment, filter the over-limit fragments and output the crack reference positioning coordinates according to the feature distance of the grid line corresponding to the over-limit fragments.

[0052] S74. Calculate the separation offset vector along the direction perpendicular to the grid line based on the crack reference positioning coordinates. Perform coordinate repositioning on the over-limit fragments according to the separation offset vector to generate crack gaps. Render crack pixels to output tearing effects and synchronously map the deformation texture and tearing effects to the BIM visualization model.

[0053] The beneficial effects of this invention are:

[0054] (1) This invention constructs an improved DeepONet model based on physical equilibrium residual constraints and parasitic wasp chemotaxis host-seeking mechanism, realizing the physical self-consistent derivation and singularity feature enhancement from sparse sensor data to a global continuous mechanical field. The mechanical parameter modulation layer constructs an anisotropic scaling factor based on material mechanical parameters to apply affine modulation to the initial continuous strain tensor field. The chemotaxis aggregation enhancement layer calculates the second-order spatial gradient norm and maps it to a pseudo-pheromone concentration field. It also initializes the parasitic wasp detection tensor kernel to perform a differentiable Euler recursive directional walk along the force field vector equation that integrates attraction and repulsion, accurately superimposing the absorbed local mutation features onto the singular host region. The physical constraint residual verification layer performs automatic differential derivation of the approximate stress tensor on the sharpened continuous strain tensor field and evaluates the mechanical equilibrium residual, which serves as a penalty term to backpropagate and constrain the network parameter iteration. This mechanism, through the deep coupling of mechanical boundary constraints and biomimetic optimization, accurately compensates for the abrupt stress concentration characteristics in areas where no sensors are deployed, eliminates the smoothing distortion and physical violations of spatial extrapolation, and generates a continuous strain tensor field that combines local sharpening characteristics with global mechanical equilibrium.

[0055] (2) This invention achieves precise mapping and dynamic evolution of structural mechanical response to visual deformation and tearing effects by constructing a rendering pipeline based on state-driven factor modulation of UV coordinates and mesh line distance constraint separation offset. The principal direction and principal value are extracted from the sharpened continuous strain tensor field to generate strain feature parameters. The spatial stress ratio is calculated and the adaptive deformation threshold is determined by combining material mechanical parameters. This is converted into a state-driven factor to dynamically correct the initial standard UV coordinate set and generate a state-marked UV coordinate set. This drives the adaptive texture pattern to perform distortion resampling and fragment shading. Based on the yield state label, rasterization interpolation is performed to obtain the fragment-level continuous state. The mesh line feature distance from the fragment texture coordinates to the deformation texture is calculated as the stress concentration factor to correct the failure comparison value. A separation offset vector is applied to the over-limit fragment along the direction perpendicular to the mesh line to generate crack gaps. This mechanism drives texture rheology and constrains geometric tearing by stress ratio adaptive threshold and stress concentration coefficient, ensuring that the deformation texture evolves continuously with the mechanical state and that the crack initiation location and direction strictly follow the physical stress concentration law. The deformation texture and tearing effect are synchronously mapped to the BIM visualization model, realizing high-fidelity visual reproduction of the structure's dynamic response. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is an overall flowchart of a structural state dynamic modeling method based on BIM visualization technology proposed in this invention;

[0058] Figure 2 This is a flowchart illustrating the working principle of the improved DeepONet model, a structural state dynamic modeling method based on BIM visualization technology proposed in this invention. Detailed Implementation

[0059] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0060] refer to Figure 1 and Figure 2 A method for dynamic structural state modeling based on BIM visualization technology, specifically including:

[0061] S1. Extract the geometric data and structural attribute information of the outer skin of the BIM visualization model and overlay it with regular textures. Calculate the initial standard UV coordinate set, bind the local coordinate system of the mesh patch and the material mechanical parameters, and construct the dynamic modeling base of the structural state.

[0062] S2. Based on the dynamic modeling of structural state, collect time-series data of strain on the surface of the structure, and perform spatial topology mapping between the sensor spatial coordinates and the outer skin mesh to generate a spatially distributed strain dataset.

[0063] S3. Based on the improved DeepONet model, the spatially distributed strain dataset and material mechanical parameters are used as dual-domain constraints for joint encoding and operator mapping to generate the initial strain field and apply affine modulation of mechanical parameters. The parasitic bee chemotactic home-seeking mechanism is introduced to perform directional walking and aggregation to compensate for local abrupt changes. Combined with physical equilibrium residual constraints, a sharpened continuous strain tensor field is generated for the area of ​​the outer skin where no sensors are deployed.

[0064] S4. Extract strain features from the sharpened continuous strain tensor field, map and calculate the UV coordinate torsion offset, and convert the critical parameters of the material mechanical parameters based on the BIM structural attribute information into state driving factors.

[0065] S5. Modulate the UV coordinate distortion offset based on the state driving factor to correct the initial standard UV coordinate set to generate the state-marked UV coordinate set, and combine the state driving factor to map and output the yield state label.

[0066] S6. Based on the structural state dynamic modeling base and yield state label, drive deformation and fragment coloring, and perform twist resampling on the adaptive texture pattern with the state label UV coordinate set to output the deformation texture.

[0067] S7. Based on the yield state label and the critical parameters of the material mechanical parameters, perform over-limit judgment, perform separation offset on the mesh line features along the deformation texture of the over-limit area to generate cracks, render tearing effects, and synchronously map the deformation texture and tearing effects to the BIM visualization model.

[0068] In this embodiment, S1 specifically includes:

[0069] S11. Extract the vertex coordinates and topological connections of the triangular mesh of the outer skin of the BIM visualization model. Calculate the cross product of the edge vector from vertex A to vertex B and the edge vector from vertex A to vertex C. Multiply the norm of the cross product result by 0.5 to obtain the area of ​​the mesh patch. Divide the cross product result by its own norm to obtain the normal vector of the mesh patch. Combine the vertex coordinates, area and normal vector of all mesh patches to construct the three-dimensional geometric feature set of the outer skin.

[0070] S12. Based on the three-dimensional geometric feature set of the outer skin, perform parametric unfolding on the three-dimensional mesh surface, calculate the affine mapping matrix from the two-dimensional parameter domain to the three-dimensional spatial surface during the unfolding process, use the column vectors of the affine mapping matrix to perform dot product and cross product operations to extract the local deformation gradient, set the unfolding area distortion tolerance to 0.05, and adjust the vertex parameter coordinates through iterative relaxation until the absolute difference between the area distortion ratio and 1 is less than 0.05, and establish an initial standard UV coordinate set without distortion.

[0071] S13. Perform singular value decomposition based on local deformation gradient, take the maximum singular value after decomposition as the maximum scaling factor, take the minimum singular value as the minimum scaling factor, calculate the mean of the maximum scaling factor and the minimum scaling factor as the area scaling ratio; calculate the angle between the principal direction vector corresponding to the maximum singular value and the parameter domain coordinate axis as the anisotropic rotation angle; set the area scaling ratio as the spatial scaling factor of the texture primitive, perform scaling and rotation transformation on the basic checkerboard pattern according to the spatial scaling factor and the anisotropic rotation angle to generate an adaptive texture pattern, map it to the three-dimensional mesh surface of the outer skin along the two-dimensional parameter direction of the initial standard UV coordinate set, and generate a visual representation of the outer skin covered with adaptive regular texture.

[0072] S14. Based on the initial standard UV coordinate set, calculate the biased directional vector along the U-parameter direction and the biased directional vector along the V-parameter direction in each mesh patch. Divide the biased directional vector along the U-parameter direction by the sum of its own norm and the minimum value 0.001 to obtain the first unit tangent vector. Subtract the dot product of the biased directional vector along the V-parameter direction and the first unit tangent vector from the biased directional vector along the V-parameter direction, and then multiply by the first unit tangent vector. Divide the result by the sum of its own norm and the minimum value 0.001 to obtain the second orthogonal unit tangent vector. Construct a local coordinate system composed of the first unit tangent vector, the second orthogonal unit tangent vector, and the normal vector, and store the local coordinate system in association with the corresponding mesh patch.

[0073] S15. Extract the preset values ​​of elastic modulus (30,000 MPa) and Poisson's ratio (0.2) and thickness (10 mm) of the mesh patches from the BIM model. Bind the elastic modulus (30,000 MPa), Poisson's ratio (0.2), and thickness (10 mm) to the corresponding mesh patches along with the first unit tangent vector, the second orthogonal unit tangent vector, and the normal vector of the local coordinate system. Generate mesh patch features carrying spatial mechanical properties. Summarize all mesh patch features to construct the dynamic modeling basis for structural state.

[0074] In this embodiment, S2 specifically includes:

[0075] S21. Based on the structural state dynamic modeling basis, the strain time series data and spatial coordinates of the outer skin sensor nodes are collected. The X-axis coordinate value, Y-axis coordinate value and Z-axis coordinate value of each node in the spatial coordinates are concatenated into a node position vector. The micro-strain numerical sequence collected at a sampling frequency of 50 Hz in the strain time series data is used as the node strain vector. The node position vector and the node strain vector are combined to construct the sensor node input matrix.

[0076] S22. Based on the node position vector in the sensor node input matrix and the vertex coordinates of the mesh patch in the three-dimensional geometric feature set of the outer skin, calculate the vertical distance from the node position vector to the plane where the three vertices of the mesh patch are located as the spatial nearest distance. Calculate the difference between the projection vector of the node position vector on the normal vector of the mesh patch and the position vector of the center point of the mesh patch as the normal projection deviation. Set the spatial nearest distance threshold to 50 mm and the normal projection deviation threshold to 20 mm. Select mesh patches with a spatial nearest distance of less than 50 mm and a normal projection deviation of less than 20 mm as target patches. Pair the sensor node number with the target patch number to generate a node-pattern mapping pair.

[0077] S23. Based on the node patch mapping pair, extract the node strain vector corresponding to the sensor node number in the sensor node input matrix, and directly assign the node strain vector to the mesh patch corresponding to the target patch number in the node patch mapping pair, retaining the original physical properties of the micro-strain values, and obtain the mesh patch strain time sequence.

[0078] S24. Analyze the amplitude distribution characteristics of the strain time series of the grid patch. Calculate the difference between the maximum and minimum values ​​of the strain time series of the grid patch in the amplitude distribution characteristics as the amplitude range. Calculate the amplitude range multiplied by a scaling factor of 0.2 as the adaptive denoising threshold. Perform a fast Fourier transform on the strain time series of the grid patch to obtain the frequency domain amplitude sequence. Calculate all peak frequencies in the frequency domain amplitude sequence whose amplitude is greater than the frequency domain mean. Select the peak frequency with the smallest frequency value minus 5 Hz as the candidate lower limit frequency. Extract the maximum value between the candidate lower limit frequency and 0 Hz as the lower limit cutoff frequency. Select the peak frequency with the largest frequency value plus 5 Hz as the upper limit cutoff frequency.

[0079] S25. Based on the adaptive denoising threshold, lower cutoff frequency, and upper cutoff frequency, adaptive filtering is performed on the strain time series of the mesh patch. First, a fast Fourier transform is performed on the strain time series of the mesh patch to obtain the frequency domain sequence. The amplitudes of frequency components with frequencies below the lower cutoff frequency and above the upper cutoff frequency in the frequency domain sequence are set to zero, while the amplitudes of frequency components within the filtering bandwidth are retained. An inverse fast Fourier transform is performed on the processed frequency domain sequence to obtain the frequency domain filtered time domain strain sequence. Subsequently, data points with absolute values ​​less than the adaptive denoising threshold in the frequency domain filtered time domain strain sequence are set to zero, while data points with absolute values ​​greater than or equal to the adaptive denoising threshold are retained. The purified time domain strain sequence is output, and all purified time domain strain sequences of the mesh patches are summarized to generate a spatially distributed strain dataset.

[0080] In this embodiment, the improved DeepONet model includes a sparse boundary coding branch network, a continuous spatial mapping backbone network, a mechanical parameter modulation layer, a chemotactic aggregation enhancement layer, and a physical constraint residual verification layer.

[0081] A sparse boundary coding branch network is used to concatenate the X-axis, Y-axis, and Z-axis coordinates of each node in the sensor spatial coordinates into a spatial coordinate vector based on a spatially distributed strain dataset and sensor spatial coordinates. The micro-strain values ​​from the spatially distributed strain dataset are then concatenated with the spatial coordinate vector along the channel dimension to generate a sensor input tensor with a dimension equal to the number of nodes multiplied by 4. The elastic modulus (30000 MPa), Poisson's ratio (0.2), and thickness (10 mm) of the mesh patches in the dynamic structural modeling substrate are extracted. Dividing the elastic modulus (30000 MPa) by the normalization constant (10000 MPa) yields a value of 3.0, and dividing the Poisson's ratio (0.2) by the normalization constant (0.5) yields a value of 0.4. The thickness... The degree attribute of 10 mm is divided by the normalization constant of 10 mm to obtain the value 1.0. The values ​​3.0, 0.4 and 1.0 are concatenated to form the material condition vector. The sensor input tensor is used as the backbone input sparse boundary coding branch network to perform three fully connected layer encoding. The number of neurons in the first layer is set to 64 and the activation function is LeakyReLU, the number of neurons in the second layer is set to 32 and the activation function is LeakyReLU, and the number of neurons in the third layer is set to 16 and the activation function is a linear function. The output of the third layer and the material condition vector are used as skip inputs, broadcast and aligned in the feature channel dimension, and then element-wise addition and fusion are performed to output a global boundary constraint feature vector with a dimension of 16.

[0082] A continuous spatial mapping backbone network is used to map the target spatial coordinates based on the region of the outer skin where no sensors are deployed. The X-axis, Y-axis, and Z-axis coordinates of the target spatial coordinates are multiplied by a frequency coefficient of 10 and mapped to sine and cosine feature values, respectively. The sine and cosine feature values ​​are concatenated to generate a spatial encoding vector of dimension 6, which is defined as a spatial query tensor. The spatial query tensor is input into the continuous spatial mapping backbone network, and two layers of multilayer perceptron feature extraction are performed. The first layer has 128 neurons and the activation function is GELU, and the second layer has 64 neurons and the activation function is GELU. A linear mapping layer with an input dimension of 6 and an output dimension of 64 is constructed to increase the spatial encoding vector to dimension 64. The output of the second layer is aligned with the spatial encoding vector of dimension 64 in the feature channel dimension and element-wise addition skip connections are performed to output a spatial basis function tensor of dimension 64.

[0083] The mechanical parameter modulation layer is used to construct a fully connected upscaling layer with an input dimension of 16 and an output dimension of 64, based on the global boundary constraint feature vector and the spatial basis function tensor. The activation function is set to a linear function. The 16-dimensional global boundary constraint feature vector is input into the fully connected upscaling layer and mapped to a 64-dimensional boundary constraint scaling vector. The element-wise product of the boundary constraint scaling vector and the spatial basis function tensor is calculated to generate an initial continuous strain tensor field of dimension 64. The difference between the values ​​3.0 and 0.4 in the material condition vector is extracted to obtain 2.6 as the principal scaling factor for anisotropy scaling. The value 0.4 is extracted as the secondary scaling factor. The first 32 feature channels of the initial continuous strain tensor field are multiplied by the principal scaling factor 2.6, and the last 32 feature channels are multiplied by the secondary scaling factor 0.4. An affine transformation is then performed to output a modulated continuous strain tensor field of dimension 64.

[0084] The chemotactic aggregation enhancement layer is used to introduce the chemotactic host-seeking mechanism of parasitic wasps in the field of insect behavior. The specific execution process includes:

[0085] The second-order spatial gradient norm is calculated for the modulated continuous strain tensor field, and the local feature rate of change is extracted. The local feature rate of change is input into the Sigmoid activation function and mapped to a non-negative quasi-pheromone concentration field with a numerical range of 0 to 1.

[0086] 64 spatial coordinate points are uniformly sampled along the spatial distribution of the modulated continuous strain tensor field. Each spatial coordinate point is initialized as a learnable parameter vector with a dimension of 16. The results are then combined to construct a parasitic wasp detection tensor kernel with a dimension of 64 by 16.

[0087] Calculate the gradient direction of the pseudo-pheromone concentration field, multiply the gradient direction by the attraction coefficient 0.8 to construct the wandering attraction potential field, calculate the Euclidean distance between the corresponding spatial coordinate points of the parasitic wasp detection tensor kernel, multiply the reciprocal of the Euclidean distance by the repulsion coefficient 0.2 to construct the wandering repulsion force field, and perform vector addition and fusion of the wandering attraction potential field and the wandering repulsion force field to generate the force field vector equation;

[0088] The walking step size is set to 0.5 mm and the number of walking steps is 10. Multi-step directional walking is performed according to the force field vector equation and differentiable Euler recursion. The current coordinate is added to the force field vector multiplied by the walking step size of 0.5 mm to obtain the coordinate of the next step. This causes the parasitic wasp detection tensor kernel to gather towards the extreme point of the pseudo-pheromone concentration field. In each walking step, the values ​​of the modulated continuous strain tensor field and the pseudo-pheromone concentration field in the neighborhood of the current coordinate are extracted and spliced. Compensation features are extracted through a fully connected layer with 16 neurons and the activation function of LeakyReLU.

[0089] The spatial coordinate point density distribution of the parasitic bee detection tensor kernel after aggregation is calculated using a Gaussian kernel function. The Gaussian kernel bandwidth is set to 2.0 mm, and the density distribution is normalized to the numerical range of 0 to 1 to generate a spatial mask. The spatial mask is input into a Sigmoid function to generate continuously differentiable soft-gated weights. The modulated continuous strain tensor field is multiplied element-wise with the soft-gated weights to construct a linear mapping layer with an input dimension of 16 and an output dimension of 64, which increases the dimension of the compensation feature to 64. Based on the spatial coordinate points after the walk terminates, the increased dimension compensation feature is scattered and superimposed on the spatial position corresponding to the modulated continuous strain tensor field. Element-wise superposition is performed on the singular host region corresponding to the extreme value of the pseudo-pheromone concentration, and the sharpened continuous strain tensor field is output.

[0090] The physical constraint residual verification layer is used to automatically perform spatial partial derivative calculation on the sharpened continuous strain tensor field. Combining the elastic modulus value of 3.0 and the Poisson's ratio value of 0.4 in the material condition vector, the partial derivative is multiplied by the elastic modulus value of 3.0 and subtracted from the product of the Poisson's ratio value of 0.4 and the orthogonal direction partial derivatives to derive the dimensionless approximate stress tensor. The divergence of the second-order partial derivative of the dimensionless approximate stress tensor with respect to spatial coordinates is calculated, and the norm of the divergence is extracted as the mechanical equilibrium residual. The mechanical equilibrium residual is multiplied by the penalty weight coefficient of 0.01 as a penalty term. The penalty term is added to the network prediction loss to form the total loss function. The gradient is calculated by backpropagation using the total loss function. The optimizer is set to Adam optimizer with a learning rate of 0.001. The network parameters of the sparse boundary encoding branch network, the continuous spatial mapping backbone network, and the learnable parameters of the parasitic bee probe tensor kernel are updated. The iteration terminates when the mechanical equilibrium residual is less than the convergence threshold of 0.001, and the final sharpened continuous strain tensor field is output.

[0091] The improved DeepONet model proposed in this step is similar to the traditional DeepONet model in that it is based on the deep operator regression theory. That is, it projects sparse spatially distributed data to a high-dimensional latent space for nonlinear mapping, uses the weight update mechanism of the branch network and the backbone network to capture the functional relationship between the input coordinates and the physical field response, and uses feature fusion and affine transformation to generate a global continuous strain response field.

[0092] The difference lies in that this invention breaks through the limitations of the traditional DeepONet model's pure data black-box regression, which ignores the mechanisms of mechanical equilibrium and local mutation. It adds a chemotactic aggregation enhancement step to search for singular host regions and replaces the smooth basis function interpolation and global convolution of the traditional DeepONet with a parasitic wasp chemotactic host-seeking mechanism. It maps the second-order spatial gradient norm to a pseudo-pheromone concentration field to perform a directional walk that combines attraction and repulsion, absorbs local mutation features and compensates them by mask superposition. At the same time, it combines a physical constraint residual verification layer to perform mechanical equilibrium residual penalty on the derived approximate stress tensor, rather than the traditional DeepONet's single data-driven loss function regression.

[0093] The beneficial effects of the improvements are that, through biomimetic walking aggregation and residual penalty closed loop, the law of conservation of mechanics and the prior knowledge of singular mutations are hard-embedded into the forward propagation and parameter update of the network. This breaks the limitations of the traditional DeepONet model, which is prone to smoothing distortion in areas without sensor deployment and violates physical equilibrium, leading to extrapolation overshoot. It achieves a precise conversion from black-box fitting to strong constraints of physical mechanisms. This design significantly enhances the ability of the improved DeepONet model to capture sudden fluctuations in stress concentration. It can accurately sharpen local mutation features in continuous strain fields. Combined with residual-driven parameter iteration, it effectively improves the physical self-consistency of dynamic strain field modeling and the absolute reliability of structural state assessment.

[0094] In this embodiment, S4 specifically includes:

[0095] S41. Based on the sharpened continuous strain tensor field, extract the first 6 eigenvectors from the 64-dimensional eigenvectors of each spatial coordinate point in the sharpened continuous strain tensor field and map them into 3 normal strain components and 3 shear strain components. Fill the 3 normal strain components and 3 shear strain components into a 3x3 matrix to generate a symmetric second-order strain tensor. Calculate the eigenvalues ​​and eigenvectors of the second-order strain tensor. Set the direction where the shear strain is zero as the principal direction. Arrange the eigenvalues ​​in descending order and perform synchronous rearrangement of the eigenvectors according to the sorting index of the eigenvalues. Extract the eigenvalue with the first position as the first principal strain value, the eigenvalue with the second position as the second principal strain value, and the eigenvalue with the third position as the third principal strain value. Concatenate the first, second, and third principal strain values ​​into a diagonal principal value vector. Combine the synchronously rearranged principal direction with the diagonal principal value vector to generate strain feature parameters.

[0096] S42. Based on strain characteristic parameters and local deformation gradient, extract the diagonal principal value vector from the strain characteristic parameters, calculate the difference between the first principal strain value and the second principal strain value in the diagonal principal value vector as the local deformation gradient, perform element-wise multiplication of the diagonal principal value vector and the local deformation gradient to obtain the parameter domain torsion vector, divide the parameter domain torsion vector by the normalization constant 100 microstrain to obtain the parameter domain torsion ratio coefficient, and multiply the parameter domain torsion ratio coefficient by the reference offset of 0.5 mm to generate the spatial geometric torsion offset.

[0097] S43. Based on the structural state dynamic modeling base, extract the elastic modulus of 30000 MPa, Poisson's ratio of 0.2, and thickness attribute of 10 mm from the mesh patches in the structural state dynamic modeling base. Extract the first principal strain value, second principal strain value, and third principal strain value from the strain characteristic parameters. Calculate the sum of the squares of the first and second principal strain values ​​plus the sum of the squares of the third principal strain value minus the difference between the first principal strain value multiplied by the second principal strain value and the first principal strain value multiplied by the third principal strain value. Extract the square root of the product of the difference and the value 3.0. Multiply the square root result by the elastic modulus of 30000 MPa to calculate the current equivalent stress of the mesh patch. Multiply the current equivalent stress by the thickness attribute of 10 mm and divide by the reference thickness constant of 10 mm to perform stress gradient correction based on the thickness change rate. Summarize the corrected equivalent stress distribution of all mesh patches to generate a spatial stress field.

[0098] S44. Based on the structural state dynamic modeling base, extract the critical yield stress of 250 MPa for the mesh surface material, extract the equivalent stress value of each mesh surface in the spatial stress field, divide the equivalent stress value by the critical yield stress of 250 MPa to calculate the spatial stress ratio, count the spatial stress ratio values ​​of all mesh surfaces in the spatial stress field, calculate the mean and standard deviation of all spatial stress ratio values, add the mean and standard deviation to the standard deviation and multiply by the coefficient 2.0 to calculate the adaptive deformation threshold.

[0099] S45. Perform normalization mapping on the spatial geometric distortion offset based on the adaptive deformation threshold. Calculate the square root of the sum of squares of the offsets of all coordinate points in the spatial geometric distortion offset as the global offset norm. Divide the spatial geometric distortion offset by the global offset norm and add a zero-prevention constant of 0.001 to calculate the normalized offset ratio. Calculate the square of the normalized offset ratio, add a smoothing constant of 0.001, and extract the square root as the smoothed offset absolute value. Divide the smoothed offset absolute value by the adaptive deformation threshold to perform normalization mapping. Input the mapping result into the Sigmoid activation function to compress it to a value between 0 and 1, and output the state driving factor.

[0100] The mechanical parameter cross-domain mapping and state-driving factor generation process proposed in this step is similar to the traditional BIM visualization deformation-driving process in that both are based on the theory of physical field feature extraction and graphical parameter mapping. That is, by converting structural response data into geometric deformation descriptors, the mapping law from mechanical state to visual deformation is captured by the coordinate space transformation mechanism, and threshold calibration is used to convert physical field response into visualized driving control quantity.

[0101] The difference lies in that this invention breaks through the limitations of traditional pure geometric linear mapping that ignores the material mechanical constitutive and yield boundary. It adds diagonalization decomposition to extract the principal direction and principal value to construct strain characteristic parameters, and replaces the traditional fixed color band interpolation with a spatial stress ratio and critical yield inference mechanism. It maps the elastic modulus, Poisson's ratio and thickness properties to the current equivalent stress and critical yield stress and performs adaptive deformation threshold truncation. Finally, it combines the adaptive deformation threshold to perform normalization mapping on the UV coordinate distortion offset to generate state driving factors, rather than a single geometric displacement scalar prediction.

[0102] The beneficial effects of the improvements are that, through constitutive deduction and stress ratio truncation, the material yield limit constraint is forcibly embedded into the graphics parameter domain, breaking the limitation of traditional methods that are prone to visual distortions that violate physical laws under over-limit conditions, leading to distortion exceeding the limit. This achieves a precise conversion from geometric representation mapping to strong mechanical constraint. The design significantly enhances the physical defense expression capability against sudden increases in local deformation, can accurately truncate non-physical deformation in continuous parameter space, and, combined with thickness correction and critical adaptive threshold fusion, effectively improves the dynamic adaptability of texture distortion and the physical self-consistency of visualization of the structural state under all working conditions.

[0103] In this embodiment, S5 specifically includes:

[0104] S51. Perform dynamic modulation correction on the spatial geometric distortion offset based on the state driving factor, extract the state driving factor and the spatial geometric distortion offset generated in the previous step, input the state driving factor into the exponential mapping function with the natural constant e as the base and truncate the upper limit value to 3.0 to calculate the dynamic modulation coefficient, multiply the spatial geometric distortion offset by the dynamic modulation coefficient and perform nonlinear scaling to generate the state driving offset.

[0105] S52. Retrieve the initial standard UV coordinate set, the diagonal length of the mesh feature (100 mm), and the deformation amplification factor (50.0). Perform algebraic superposition on the initial standard UV coordinate set based on the state-driven offset. Extract the original U-coordinate values ​​and the original V-coordinate values ​​from the initial standard UV coordinate set. Extract the U-axis offset component and the V-axis offset component from the state-driven offset. Divide the U-axis offset component by the diagonal length of the mesh feature (100 mm) and multiply by the deformation amplification factor (50.0) to calculate the dimensionless U-axis offset ratio. Divide the V-axis offset component by the diagonal length of the mesh feature (100 mm) and multiply by the deformation amplification factor (50.0) to calculate the dimensionless V-axis offset ratio. Add the dimensionless U-axis offset ratio to the original U-coordinate values ​​to calculate the corrected U-coordinates. Add the dimensionless V-axis offset ratio to the original V-coordinate values ​​to calculate the corrected V-coordinates. Combine the corrected U-coordinates and corrected V-coordinates to replace the original coordinates and generate the state marker UV coordinate set.

[0106] S53. Extract the driving offset features from the state marker UV coordinate set. Calculate the absolute value of the difference between the corrected U coordinate in the state marker UV coordinate set and the original U coordinate in the initial standard UV coordinate set as the U-direction offset feature value. Calculate the absolute value of the difference between the corrected V coordinate and the original V coordinate as the V-direction offset feature value. Add the U-direction offset feature value and the V-direction offset feature value to obtain the driving offset feature. Retrieve the adaptive deformation threshold from the previous step and multiply it by the deformation transformation coefficient 1.2 to define the yield boundary threshold. Divide the driving offset feature by the offset normalization constant 0.05 to obtain the offset ratio value. Input the offset ratio value into the ReLU activation function to remove negative values. Compare the output value with the yield boundary threshold. When the output value is greater than or equal to the yield boundary threshold, assign the value 1. When the output value is less than the yield boundary threshold, assign the value 0. Generate the yield state label of the mesh patch.

[0107] In this embodiment, S6 specifically includes:

[0108] S61. Based on the features of the mesh facets carrying spatial mechanical properties in the dynamic modeling base of the structural state, the mesh deformation is driven by the state marker UV coordinate set generated in the previous step. The initial standard UV coordinate set is replaced to update the texture sampling coordinates after deformation. The yield state label generated in the previous step is extracted. When the yield state label is 1, the current fragment is defined as the yield region and the mapping color vector is set to the values ​​1.0, 0.2, and 0.0. When the yield state label is 0, the current fragment is defined as the elastic region and the mapping color vector is set to the values ​​0.2, 0.6, and 1.0. The texture sampling coordinates and the mapping color vector are combined to construct the fragment coloring mapping relationship. Fragment coloring is performed to generate the yield visualization fragment.

[0109] S62. Based on the statistical texture sampling density distribution characteristics of the state marker UV coordinate set, retrieve the state marker UV coordinate set generated in the previous step, traverse all directly adjacent meshes of the current mesh, calculate the absolute value of the difference between the current mesh and each adjacent mesh in the state marker UV coordinate set as the U-axis coordinate difference, calculate the absolute value of the difference between the corrected V coordinate as the V-axis coordinate difference, extract the square root of the square of the U-axis coordinate difference and add the square of the V-axis coordinate difference, and add the zero constant 0.001 to calculate the Euclidean coordinate spacing. Divide the value 1.0 by the Euclidean coordinate spacing to calculate the local sampling density between the current mesh and the adjacent mesh, accumulate the local sampling densities between the current mesh and all directly adjacent meshes to calculate the total local sampling density of the current mesh, accumulate the total local sampling density of all directly adjacent meshes to calculate the cumulative total local sampling density of adjacent meshes, and divide the total local sampling density of the current mesh by the sum of the total local sampling density of the current mesh and the cumulative total local sampling density of adjacent meshes to calculate the normalized resampling interpolation weight.

[0110] S63. Perform distorted resampling on the adaptive texture pattern based on the resampling interpolation weights. Extract the resampling interpolation weights generated in the previous step. Retrieve the initial color value of the current original texture pixel and the adjacent color values ​​of adjacent original texture pixels in the adaptive texture pattern. Multiply the resampling interpolation weights by the initial color value of the current original texture pixel to calculate the current color contribution value. Subtract the resampling interpolation weights from the value 1.0 to calculate the adjacent interpolation weights. Multiply the adjacent interpolation weights by the adjacent color values ​​of adjacent original texture pixels to calculate the adjacent color contribution value. Add the current color contribution value to the adjacent color value. The contribution values ​​are weighted and summed to obtain the distorted color values. The red, green, and blue channel values ​​are extracted from the distorted color values. The red channel values ​​are input into the Tanh activation function to compress them to a value range of -1.0 to 1.0, and then an offset constant of 1.0 is added and divided by the value of 2.0 to map them to a value range of 0 to 1.0. The same Tanh activation and offset mapping operation is performed on the green and blue channel values. The mapped red, green, and blue channel values ​​are concatenated to generate distorted texture pixels. All distorted texture pixels are then combined to generate the distorted texture.

[0111] In this embodiment, S7 specifically includes:

[0112] S71. Based on the yield state labels generated in the previous steps, perform centroid coordinate interpolation in the rasterization stage, extract the yield state label values ​​at the mesh vertices and the weight ratios of the rasterized fragments relative to the three vertices of the enclosing triangle, multiply the weight ratios of the three vertices by the yield state label values ​​of the corresponding vertices and sum them up, input the summation result into the Sigmoid activation function to map to a value range between 0.0 and 1.0, and output a fragment-level continuous yield state field.

[0113] S72. Based on the fragment-level continuous yield state field generated in the previous step, extract the fragment yield parameter and fragment texture coordinates. Retrieve the deformation texture generated in the previous step, calculate the Euclidean distance from the fragment texture coordinates to the mesh line features in the deformation texture, and add a zero constant of 0.01 mm as the stress concentration distance. Divide the reference scaling factor of 2.0 mm by the stress concentration distance to calculate the stress concentration factor. Multiply the stress concentration factor by the fragment yield parameter to calculate the stress correction parameter. Extract the adaptive deformation threshold retrieved when defining the yield boundary threshold in the previous step, multiply it by the safety factor of 1.5 to define the material critical yield judgment parameter. Add the stress correction parameter to the fragment yield parameter to calculate the failure comparison value. Compare the failure comparison value with the material critical yield judgment parameter. When the failure comparison value is greater than or equal to the material critical yield judgment parameter, output the value 1 as the over-limit judgment result. When the failure comparison value is less than the material critical yield judgment parameter, output the value 0 as the over-limit judgment result.

[0114] S73. Based on the over-limit judgment results generated in the previous step, filter over-limit fragments, traverse and extract over-limit fragments with an over-limit judgment result of value 1, retrieve the stress concentration distance calculated in the previous step, extract the difference between the value 1.0 mm and the stress concentration distance, input it into the Max activation function and compare it with the value 0.0, extract the larger value and discard the negative offset as the grid line offset, and calculate the crack reference positioning coordinates by adding the grid line offset to the fragment texture coordinates of the over-limit fragments along the direction of the grid line feature normal vector.

[0115] S74. Based on the crack reference positioning coordinates generated in the previous step, calculate the separation offset vector along the direction perpendicular to the grid line. Extract the vertical component of the grid line normal vector in the crack reference positioning coordinates. Multiply the vertical component by the crack opening coefficient of 0.5 mm to calculate the physical separation offset vector. Retrieve the diagonal length of the grid feature (100 mm). Divide the physical separation offset vector by the diagonal length of the grid feature (100 mm) to calculate the dimensionless UV separation offset vector. Perform coordinate repositioning on the over-limit fragment based on the dimensionless UV separation offset vector. Add the dimensionless UV separation offset vector to the original texture coordinates of the over-limit fragment to calculate the crack gap coordinates. Set the pixel color corresponding to the crack gap coordinates to the crack color values ​​of 0.1, 0.1, and 0.1 to render the crack pixels and output the tearing effect. Retrieve the deformation texture and tearing effect generated in the previous step. Map the deformation texture to the grid patch surface of the BIM visualization model. Overlay and map the tearing effect to the over-limit fragment position of the BIM visualization model to complete the visualization synchronous mapping.

[0116] The physical-driven crack rendering process proposed in this step is similar to the traditional BIM visualization damage characterization process in that both are based on the theory of rasterization rendering and texture space mapping in computer graphics. That is, by projecting the grid-level physical state onto the fragment space for interpolation calculation, pixel-level shading and coordinate repositioning mechanisms are used to capture the visual evolution of structural damage, and both use geometric offset and texture fusion to output visual damage effects.

[0117] The difference lies in that this invention breaks through the limitation of traditional pure visual representation simulation ignoring the mechanical stress concentration mechanism. It adds a centroid coordinate interpolation step to construct a fragment-level continuous yield state field, and replaces the traditional preset crack map and Boolean cutting with a mesh line feature distance constraint mechanism. The distance from the fragment texture coordinate to the deformation texture mesh line is mapped to the stress concentration coefficient to correct the failure comparison value and perform critical yield truncation. Finally, the separation offset vector is calculated along the direction perpendicular to the mesh line to perform coordinate repositioning on the over-limit fragment to generate crack gaps, rather than a single artificial model replacement or static texture superposition.

[0118] The beneficial effects of the improvements are that, by using interpolation continuity and distance constraint truncation, the stress concentration evolution law is forcibly embedded into the graphics rendering pipeline, breaking the limitation of traditional methods that are prone to visual distortions that violate mechanical causality under extreme conditions, leading to representational distortion. This achieves a precise conversion from geometric representation simulation to strong physical mechanism constraints. The design significantly enhances the physical defense expression capability for crack initiation direction, can accurately truncate non-stressed cracks in fragment-level parameter space, and, combined with repositioning fusion of vertical separation offset, effectively improves the dynamic adaptability of crack effects and the physical self-consistency of structural damage rendering under all working conditions.

[0119] Example 1: To verify the feasibility of this invention in the health monitoring and digital operation and maintenance of large-scale spatial structures, the method of this invention was applied to the steel structure health monitoring and BIM visualization system of a provincial key transportation hub (hereinafter referred to as "Hall S"). In traditional structural condition monitoring and visualization systems, color band mapping based on discrete threshold triggering or preset deformation script-driven model rendering are commonly used. These methods not only struggle to accurately extrapolate the continuous mechanical state across the entire domain under sparse sensor deployment, but also fail to accurately acquire the local abrupt change characteristics and actual crack evolution laws in stress concentration areas, easily leading to misjudgment or underreporting of structural yielding and failure risks. To solve the above problems, Hall S decided to adopt the structural condition dynamic modeling method based on BIM visualization technology proposed in this invention.

[0120] During implementation, Venue S first extracted the vertex coordinates and topological connections of the triangular mesh on the outer skin of the BIM visualization model, calculated the area and normal vector of the mesh patches to construct a three-dimensional geometric feature set for the outer skin; parametric unfolding was performed on the three-dimensional mesh surface to establish an initial standard UV coordinate set without distortion, and a visual representation of the outer skin with an adaptive regular texture was generated; the local tangent space basis vectors within each mesh patch were calculated to construct a local coordinate system, and the elastic modulus, Poisson's ratio, and thickness attributes were extracted and bound to the corresponding mesh patch to construct a dynamic modeling basis for the structural state. Simultaneously, Venue S's operation and maintenance team used high-frequency strain sensors deployed on the steel structure's outer skin to collect strain time-series data and spatial coordinates, calculated the spatial nearest distance from the sensor nodes to the mesh patches and the normal projection deviation to perform spatial topological mapping, and performed adaptive filtering based on an adaptive denoising threshold and filtering bandwidth to generate a spatially distributed strain dataset.

[0121] Venue S improves the DeepONet model by using spatially distributed strain datasets and material mechanical parameters as dual-domain constraints. The global boundary constraint feature vector and spatial basis function tensor are jointly encoded by a sparse boundary encoding branch network and a continuous spatial mapping backbone network. The mechanical parameter modulation layer constructs an anisotropic scaling factor based on the material mechanical parameters and applies an affine transformation. The chemotaxis aggregation enhancement layer introduces the parasitic wasp chemotaxis host-seeking mechanism, calculates the second-order spatial gradient norm and maps it to a pseudo-pheromone concentration field, initializes the parasitic wasp detection tensor kernel and performs multi-step directional walks along the force field vector equation, superimposing local mutation features onto the singular host region. The physical constraint residual verification layer performs automatic differential derivation of the approximate stress tensor on the sharpened continuous strain tensor field, evaluates the mechanical equilibrium residual and uses it as a penalty term to backpropagate the constraint network parameter iteration, and generates the sharpened continuous strain tensor field of the outer skin region without sensor deployment.

[0122] In the core state-driven and dynamic visualization stage, this invention extracts strain feature parameters from the sharpened continuous strain tensor field, and calculates the UV coordinate distortion offset by combining local deformation gradient; extracts elastic modulus, Poisson's ratio, and thickness attributes to calculate the current equivalent stress and critical yield stress, determines the adaptive deformation threshold based on the spatial stress ratio, and performs normalization mapping on the UV coordinate distortion offset to generate a state-driven factor; dynamically modulates and corrects the initial standard UV coordinate set based on the state-driven factor to generate a state-marked UV coordinate set, and performs interval mapping to generate a yield state label by combining the yield boundary threshold; updates the texture sampling coordinates with the state-marked UV coordinate set to drive fragment shading and adaptive texture pattern distortion resampling, and outputs the deformation texture; performs centroid coordinate interpolation based on the yield state label to obtain the fragment-level continuous yield state field, calculates the mesh line feature distance from the fragment texture coordinates to the deformation texture as the stress concentration coefficient, outputs the over-limit judgment result by combining the material critical yield judgment parameter, calculates the separation offset vector along the direction perpendicular to the mesh line to generate crack gaps, renders tearing effects, and synchronously maps the deformation texture and tearing effects to the BIM visualization model, realizing a closed-loop leap from physical perception to dynamic visual reproduction.

[0123] To further verify the actual performance of the method of the present invention, venue S conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1:

[0124] Table 1. Performance Comparison of Dynamic Modeling Methods for the S-Steel Structure of Venue

[0125] Error in strain field projection for areas without data deployment (%) 18.5 3.2 -82.7% Local stress concentration feature capture rate (%) 65.0 94.5 +29.5% Yield state identification accuracy (%) 78.4 96.1 +17.7% Excessive cracking positioning deviation (mm) 25.0 2.5 -90.0% Physical equilibrium residual convergence iteration count 500 120 -76.0% Dynamic deformation rendering latency (milliseconds) 250 85 -66.0% Physical consistency of crack evolution direction (%) 55.0 92.0 +37.0% False alarms and omissions lead to invalid inspections (times / year) 40 5 -87.5% Time taken (seconds) to synchronize BIM model status update 30 3 -90.0% Operation and maintenance decision response satisfaction (%) 75.0 96.0 +21.0%

[0126] As shown in Table 1, the performance of the structural state dynamic modeling system was comprehensively improved after applying the method of this invention. The strain field extrapolation error in the unspecified area decreased from 18.5% to 3.2% using traditional methods, and the capture rate of local stress concentration features increased from 65.0% to 94.5%, significantly improving the accuracy of state perception and providing a reliable basis for subsequent risk assessment. The yield state identification accuracy increased from 78.4% to 96.1%, effectively avoiding erroneous early warning scheduling. The deviation in over-limit crack positioning was significantly reduced from 25.0 mm to 2.5 mm, and the physical consistency of crack evolution direction increased to 92.0%, significantly enhancing the physical self-consistency of the visual expression. In addition, the dynamic deformation rendering latency was shortened from 250 milliseconds to 85 milliseconds, and the BIM model state synchronization update time was shortened from 30 seconds to 3 seconds, greatly improving the system's real-time performance. The number of invalid inspections caused by false alarms and missed alarms decreased from 40 times / year to 5 times / year, significantly reducing operation and maintenance costs. Satisfaction with operational and maintenance decision-making responses has also improved significantly, increasing from 75.0% to 96.0%.

[0127] Through the method of this invention, venue S has successfully achieved accurate perception and physical self-consistent dynamic visualization of the mechanical state of a large spatial structure, effectively reducing the risk of misjudgment of yield failure, ensuring the safe operation of the hub venue structure, significantly improving the intelligence and digitalization level of steel structure health monitoring, significantly reducing the data analysis burden of operation and maintenance personnel, enhancing the dynamic expression robustness of the BIM visualization model, and providing strong technical support for digital twins of engineering structures.

[0128] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for dynamic structural state modeling based on BIM visualization technology, characterized in that, Includes the following steps: S1. Extract the geometric data and structural attribute information of the outer skin of the BIM visualization model and overlay it with regular textures. Calculate the initial standard UV coordinate set, bind the local coordinate system of the mesh patch and the material mechanical parameters, and construct the dynamic modeling base of the structural state. S2. Based on the dynamic modeling of structural state, collect time-series data of strain on the surface of the structure, and perform spatial topology mapping between the sensor spatial coordinates and the outer skin mesh to generate a spatially distributed strain dataset. S3. Based on the improved DeepONet model, the spatially distributed strain dataset and material mechanical parameters are used as dual-domain constraints for joint encoding and operator mapping to generate the initial strain field and apply affine modulation of mechanical parameters. The parasitic bee chemotactic home-seeking mechanism is introduced to perform directional walking and aggregation to compensate for local abrupt changes. Combined with physical equilibrium residual constraints, a sharpened continuous strain tensor field is generated for the area of ​​the outer skin where no sensors are deployed. S4. Extract strain features from the sharpened continuous strain tensor field, map and calculate the UV coordinate torsion offset, and convert the critical parameters of the material mechanical parameters based on the BIM structural property information into state driving factors. S5. Modulate the UV coordinate distortion offset based on the state driving factor to correct the initial standard UV coordinate set to generate the state-marked UV coordinate set, and combine the state driving factor to map and output the yield state label. S6. Based on the structural state dynamic modeling base and yield state label, drive deformation and fragment coloring, and perform twist resampling on the adaptive texture pattern with the state label UV coordinate set to output the deformation texture. S7. Based on the yield state label and the critical parameters of the material mechanical parameters, perform over-limit judgment, perform separation offset on the mesh line features along the deformation texture of the over-limit area to generate cracks, render tearing effects, and synchronously map the deformation texture and tearing effects to the BIM visualization model.

2. The structural state dynamic modeling method based on BIM visualization technology according to claim 1, characterized in that, S1 specifically includes: S11. Extract the vertex coordinates and topological connections of the triangular mesh of the outer skin of the BIM visualization model, calculate the area and normal vector of the mesh patch, and construct the three-dimensional geometric feature set of the outer skin. S12. Based on the three-dimensional geometric feature set of the outer skin, perform parametric unfolding on the three-dimensional mesh surface, calculate the affine mapping matrix from the two-dimensional parameter domain to the three-dimensional spatial surface and the local deformation gradient during the unfolding process, and establish an undistorted initial standard UV coordinate set. S13. The area change distribution characteristics and local deformation gradient distribution characteristics generated by statistical parameterization are used to calculate the spatial scaling factor and anisotropic rotation angle of the texture primitives based on the area change distribution characteristics and local deformation gradient distribution characteristics. An adaptive texture pattern is generated based on the spatial scaling factor and anisotropic rotation angle and mapped to the three-dimensional mesh surface of the outer skin along the two-dimensional parameter direction of the initial standard UV coordinate set to generate a visual representation of the outer skin covered with adaptive regular texture. S14. Calculate the local tangent space basis vectors within each mesh patch, construct a local coordinate system composed of orthogonal tangent vectors and normal vectors, and store it in association with the corresponding mesh patch; S15. Extract the elastic modulus and Poisson's ratio, as well as the thickness attribute of the mesh surface in the BIM model, and bind the elastic modulus, Poisson's ratio and thickness attribute together with the local coordinate system to the corresponding mesh surface to generate mesh surface features carrying spatial mechanical properties. Summarize and construct the dynamic modeling base of the structural state.

3. The method for dynamic structural state modeling based on BIM visualization technology according to claim 1, characterized in that, S2 specifically includes: S21. Based on the structural state dynamic modeling basis, collect the strain time series data and spatial coordinates of the outer skin sensor nodes; S22. Based on spatial coordinates and the three-dimensional geometric feature set of the outer skin, calculate the spatial closest distance and normal projection deviation from the sensor node to the mesh patch, and perform spatial topology mapping to generate node-pattern mapping pairs. S23. Based on the node-pattern mapping pair, the strain time series data of the sensor node is transferred to the corresponding mesh patch to generate the mesh patch strain time series sequence; S24. Statistically analyze the amplitude distribution characteristics and frequency band energy distribution characteristics of the strain time series of the grid patch, and calculate the adaptive denoising threshold and filtering bandwidth of the strain time series data; S25. Based on the adaptive denoising threshold and filtering bandwidth, perform adaptive filtering on the strain time series of grid patches to generate a spatially distributed strain dataset.

4. The method for dynamic structural state modeling based on BIM visualization technology according to claim 1, characterized in that, The improved DeepONet model includes a sparse boundary coding branch network, a continuous spatial mapping backbone network, a mechanical parameter modulation layer, a chemotactic aggregation enhancement layer, and a physical constraint residual verification layer. The sparse boundary coding branch network is used to concatenate the spatially distributed strain dataset with the sensor spatial coordinates to form a sensor input tensor, and perform joint coding with material mechanical parameters as conditions to output a global boundary constraint feature vector. The continuous spatial mapping backbone network is used to encode the target spatial coordinates of the area on the outer skin where no sensors are deployed into a spatial query tensor, perform nonlinear feature extraction, and output a spatial basis function tensor. The mechanical parameter modulation layer is used to perform an inner product operation between the global boundary constraint feature vector and the spatial basis function tensor to generate an initial continuous strain tensor field; an anisotropic scaling factor is constructed based on the material mechanical parameters, and an affine transformation is applied to the initial continuous strain tensor field to output a modulated continuous strain tensor field. The chemotactic aggregation enhancement layer is used to introduce the parasitic wasp chemotactic host-seeking mechanism in the field of insect behavior, and the specific execution process includes: The second-order spatial gradient norm is calculated for the modulated continuous strain tensor field, local feature change rates are extracted and mapped to a non-negative quasi-pheromone concentration field; samples are taken along the spatial distribution of the modulated continuous strain tensor field, and a set of learnable parasitic wasp detection tensor kernels are initialized. A wandering attraction potential field is constructed based on the gradient direction of the pseudo-pheromone concentration field, and a wandering repulsion force field is constructed based on the spatial distance metric between the parasitic wasp detection tensor kernels. The wandering attraction potential field and the wandering repulsion force field are fused to generate the force field vector equation. Based on the force field vector equation, a multi-step directional walk is performed according to the differentiable Euler recursion, causing the parasitic wasp detector tensor kernel to gather towards the pheromone extremum point, i.e. the singular host region. During the multi-step directional walk, local mutation features are absorbed as compensation features. The density distribution of the gathered parasitic wasp detector tensor kernel is transformed into a spatial mask. The modulated continuous strain tensor field is subjected to a dot product masking. The compensation features are aligned and superimposed on the singular host region according to the mask, and the sharpened continuous strain tensor field is output. The physical constraint residual verification layer is used to perform automatic differentiation on the sharpened continuous strain tensor field and derive the approximate stress tensor in combination with material mechanics parameters. It evaluates the equilibrium residual of stress and mechanical equilibrium state, and uses it as a penalty term to backpropagate the constraint network parameter iteration and the learnable parameter update of the parasitic bee detection tensor kernel, outputting the sharpened continuous strain tensor field.

5. The method for dynamic structural state modeling based on BIM visualization technology according to claim 1, characterized in that, S4 specifically includes: S41. Based on the sharpened continuous strain tensor field, the tensor is diagonalized and decomposed. The principal direction is determined with the direction of zero shear strain as the reference, and the principal value is determined with the order of extreme values ​​of diagonal elements as the reference. The strain characteristic parameters are then combined to generate the strain characteristic parameters. S42. Based on strain characteristic parameters and local deformation gradient, update the calculation parameter domain torsion vector and calculate the UV coordinate torsion offset. S43. Based on the dynamic modeling basis of structural state, extract the elastic modulus, Poisson's ratio and thickness attributes, calculate the current equivalent stress of the mesh surface by combining the strain characteristic parameters, and use the thickness attribute to correct the equivalent stress distribution to generate a spatial stress field. S44. Based on the elastic modulus and Poisson's ratio, the critical yield stress of the material is calculated by combining the strain characteristic parameters; the spatial stress ratio is calculated by statistically analyzing the spatial stress field and the critical yield stress, and the adaptive deformation threshold is determined based on the distribution characteristics of the spatial stress ratio. S45. Perform normalization mapping on the UV coordinate distortion offset based on the adaptive deformation threshold to generate the state driving factor.

6. The method for dynamic structural state modeling based on BIM visualization technology according to claim 1, characterized in that, S5 specifically includes: S51. Perform dynamic modulation correction on the UV coordinate distortion offset based on the state driving factor to generate the state driving offset. S52. Retrieve the initial standard UV coordinate set, perform algebraic superposition on the initial standard UV coordinate set based on the state-driven offset, and generate the state-marked UV coordinate set. S53. Extract the driving offset features from the UV coordinate set of the state markers, combine them with the state driving factor to define the yield boundary threshold, perform interval mapping, and generate the yield state label of the mesh patch.

7. The method for dynamic structural state modeling based on BIM visualization technology according to claim 1, characterized in that, S6 specifically includes: S61. Based on the features of the mesh facets carrying spatial mechanical properties in the dynamic modeling base of structural state, the mesh deformation is driven by the mesh facet features. The texture sampling coordinates after deformation are updated by replacing the initial standard UV coordinate set with the state marker UV coordinate set. The fragment coloring mapping relationship is constructed by combining the yield state label and the fragment coloring is performed to generate yield visualization fragments. S62. Calculate the resampling interpolation weights based on the statistical texture sampling density distribution characteristics of the state marker UV coordinate set; S63. Perform distortion resampling on the adaptive texture pattern based on the resampling interpolation weight to generate a deformed texture.

8. The method for dynamic structural state modeling based on BIM visualization technology according to claim 1, characterized in that, Specifically, S7 includes: S71. Based on the yield state label, perform barycenter coordinate interpolation in the rasterization stage to output a fragment-level continuous yield state field. S72. Extract the fragment yield parameter and fragment texture coordinates based on the fragment-level continuous yield state field, obtain the material critical yield judgment parameter, calculate the characteristic distance of the grid line from the fragment texture coordinate to the deformation texture as the stress concentration coefficient, combine the stress concentration coefficient and the fragment yield parameter to correct the failure comparison value, compare the failure comparison value and the material critical yield judgment parameter to output the over-limit judgment result. S73. Based on the result of the over-limit judgment, filter the over-limit fragments and output the crack reference positioning coordinates according to the feature distance of the grid line corresponding to the over-limit fragments. S74. Calculate the separation offset vector along the direction perpendicular to the grid line based on the crack reference positioning coordinates. Perform coordinate repositioning on the over-limit fragments according to the separation offset vector to generate crack gaps. Render crack pixels to output tearing effects and synchronously map the deformation texture and tearing effects to the BIM visualization model.

Images