A smart prediction method and system for differential settlement of highway widening subgrade

Abstract

The application provides an intelligent prediction method and system for highway widening roadbed differential settlement, and relates to the technical field of highway widening. Real-time collection of new and old pavement joint data, construction of a three-dimensional stress field model, identification of weak areas and generation of a topological structure, combined with pore water pressure gradient, mapping of coordinates to filler particle size parameters, dynamic allocation of creep parameter weights, and generation of a creep cloud map. Extract the strain rate eigenvalue, establish the correlation matrix of traffic load and shear strength degradation, correct the shear strength factor, construct the settlement differential model, and predict the settlement distribution. Use triple residual comparison to optimize parameters, when the similarity coefficient meets the standard, output the settlement early warning atlas, and divide the early warning level. The application integrates multi-source data, improves the early warning coincidence degree, and improves the settlement prediction accuracy and prevention and control timeliness.

Description

Technical Field

[0001] This invention relates to the field of highway widening technology, specifically to an intelligent prediction method and system for differential settlement of highway widening subgrade. Background Technology

[0002] In highway widening projects, the prediction and control of differential settlement of the roadbed has long relied on traditional monitoring methods and empirical models. Existing technologies mostly employ single-point settlement monitoring or finite element simulation based on homogeneous assumptions, making it difficult to achieve a refined characterization of the dynamic response of the entire roadbed cross-section. Especially at the junction of new and old pavements, due to the coupling effect of uneven distribution of fill compaction, groundwater level fluctuations, and the time history effect of traffic loads, conventional methods often ignore the dynamic evolution mechanism of weak areas, resulting in insufficient settlement prediction accuracy. This leads to a significant spatiotemporal misalignment between early warning results and actual areas of abnormal settlement, restricting the timeliness and reliability of engineering risk prevention and control. Summary of the Invention

[0003] The main objective of this invention is to provide an intelligent prediction method and system for differential settlement of highway widening subgrade. To achieve this objective, this invention proposes an intelligent prediction method for differential settlement of highway widening subgrade, applied to an intelligent prediction system for differential settlement of highway widening subgrade. The system includes: a multi-source data acquisition unit, a creep analysis unit, a coupling correction unit, and a settlement prediction unit; the method includes:

[0004] Based on real-time acquisition of compaction distribution data, groundwater level change data and traffic load time history curves at the junction of new and old road surfaces by sensors, a three-dimensional stress field model with a weak zone marker layer is constructed. The three-dimensional stress field model synchronously outputs the stress state parameters of each grid node.

[0005] Creep analysis unit: Maps the weak zone marker layer to the filler particle size distribution parameters, dynamically allocates creep parameter weights based on the coordinates of the weak zone marker layer, and generates a creep cloud map in combination with stress state parameters. The creep cloud map contains a time-varying strain rate distribution band that overlaps with the weak zone marker layer.

[0006] By matching the mesh nodes of the creep cloud map with the three-dimensional stress field model, the strain rate characteristic value corresponding to the weak zone marker layer is extracted, and a correlation matrix of traffic load-shear strength degradation with groundwater level fluctuation amplitude as constraint condition is established. The correlation matrix is ​​dynamically corrected by a two-parameter reduction algorithm based on the water content sensitivity of the weak zone.

[0007] The modified correlation matrix is ​​coupled with the stress state parameters to construct a settlement differential model containing the viscoelastic-plastic constitutive equations of the weak region. A settlement distribution map that integrates the spatial topology of the weak region marker layer is generated by spatiotemporal convolution.

[0008] The system simultaneously optimizes the spatial resolution and creep parameter weights of the weak zone marker layer by comparing the triple residuals of the three-dimensional stress field model, creep cloud map, and settlement distribution map. When the spatial overlap between the weak zone marker layer and the measured settlement anomaly area reaches a set threshold, the system outputs a pavement settlement early warning map.

[0009] Preferably, the construction of the three-dimensional stress field model is based on the fusion of multi-source heterogeneous data. Compaction degree distribution data, groundwater level change data, and traffic load time history curves are aligned using a unified timestamp. Weighted overlay is employed to achieve data format conversion and spatiotemporal synchronization. Specifically, compaction degree data is normalized based on grid nodes, groundwater level data is mapped to the same grid system through spatial interpolation, and traffic load time history curves are converted into equivalent nodal loads based on vehicle axle load distribution. Weight allocation is adjusted according to the measurement accuracy of each data source, with a weight coefficient of 0.6 for compaction degree data, 0.3 for groundwater level data, and 0.1 for traffic load data, for a total weight coefficient of 1. The coefficient of variation method is used to dynamically allocate weights based on the dispersion of the measured data to ensure the physical balance of the data fusion. Spatial clustering analysis uses a density clustering algorithm, with the geometric center of the region where the compaction degree is below a set threshold as the initial seed point, combined with the extreme points of the second derivative of the stress state parameter as the clustering feature vector. The extreme points of the second derivative were calculated using the central difference method. The neighborhood radius was set to 1.2 times the average mesh size, and the minimum sample size was 5, generating a three-dimensional weak zone topology with thickness gradient characteristics. In the finite element mesh generation, the mesh size at the weak zone boundary was reduced to one-fifth to one-third of that in the regular region. This ratio was dynamically adjusted according to the stress gradient amplitude, triggering mesh refinement when the local stress gradient exceeded 10 kPa / m. Cubic spline interpolation was used for the dynamic loading of the pore water pressure gradient. The interpolation nodes were distributed along the groundwater level monitoring points, and the boundary condition was set to ensure the impermeable layer pressure was always zero. The interpolation function was a piecewise cubic polynomial to ensure the continuity and smoothness of the pore water pressure within the weak zone marker layer, ultimately forming a three-dimensional tensor field model containing stress state parameters. The dynamic loading of the pore water pressure gradient in the three-dimensional stress field model used cubic spline interpolation, with the node distribution discretized along the spatial coordinates (x, y, z) of the groundwater level monitoring points. The interpolation function was: Where N i (x), N j (y), N k (z) is a one-dimensional cubic spline basis function. The node sequence is divided at equal intervals according to the distance between monitoring points, and the basis function satisfies continuity; the interpolation coefficient p ijk The measured data were fitted using the least squares method, with a residual tolerance set to 1×10⁻⁶. -4 kPa; at the boundary of the weak zone, the mesh size is reduced to 1 / 5 to 1 / 3 of that in the normal region, the specific proportion depending on the local stress gradient. Decide: when At that time, adaptive mesh refinement is triggered to ensure precise coupling between pore water pressure gradient and stress field.

[0010] Preferably, the creep contour map is constructed by mapping the spatial coordinates of the weak zone marker layer to the filler particle size distribution parameters through coordinate transformation. A bilinear interpolation method is used to transfer parameters between grid nodes, ensuring accurate matching between the coarse particle content data and the geometric topology of the weak zone. The nonlinear mapping relationship is defined as an exponential function relationship between the creep parameter and the coarse particle content, expressed as: μ=μ0·exp(λ·C g )w i Where μ is the creep parameter, μ0 is the reference creep rate, and λ is the gradation sensitivity coefficient, ranging from 0.1 to 0.5, calibrated through indoor triaxial creep tests, C g The weighting coefficient wi represents the proportion of coarse particles; its dynamic allocation is based on the Euclidean distance function from node i to the centroid xc of the coarse particle enrichment region, calculated using the following formula: Where x c Let d be the centroid coordinates of the coarse-grained enrichment region. max The maximum radial dimension of the weak zone is defined; the implicit time integral time step Δt is set to 1 hour, and the convergence condition is that the strain increment residual norm is less than 1 × 10⁻⁶. -6 The tangent modulus is constructed using the viscoelastic constitutive equation; the second deviatoric stress invariant J2 is selected as the coupling reference for the deviatoric stress tensor components, multiplied by the weighting coefficients and input into the viscoelastic constitutive equation to calculate the anisotropic creep strain increment, which is: Δε c =Δt(J2w i ) / (2G v ), where G v The viscoelastic shear modulus is used to generate a time-varying strain rate distribution band, which is then superimposed onto the weak area marker layer to form a creep contour map.

[0011] Preferably, the construction of the correlation matrix involves extracting the principal strain rate gradient from the strain rate eigenvalues ​​using eigenvalue decomposition. This is achieved by performing eigenvalue decomposition on the strain rate tensor of each grid node and taking the direction of the eigenvector corresponding to the largest eigenvalue as the principal strain rate gradient direction. The shear strain rate direction angle is obtained by calculating the ratio of shear strain components using the arctangent function. The row vectors of the two-dimensional matrix represent the groundwater level fluctuation amplitude Δh discretized at 0.1m intervals, the column vectors represent the traffic load spectral density S(f) divided into 0.5Hz frequency bands, and the matrix elements are the shear strength degradation factor η. The specific formula for the two-parameter reduction algorithm is as follows:

[0012] η = η0·[1-β1·tanh(α·Λh)][1-β2·ln(1+γ·S(f))], where η0 is the initial shear strength determined by direct shear test, β1 is the water content sensitivity coefficient calibrated by regression analysis with a value range of 0.05-0.15, β2 is the load number sensitivity coefficient with a value range of 0.001-0.005, α and γ are fitting parameters, optimized by least squares method; during dynamic correction, the negative exponential relationship between the water content m in the weak zone and the matrix suction n is n = n0exp(-km), where n0 is the initial matrix suction, k is the attenuation coefficient, determined by fitting the soil-water characteristic curve. When the water content exceeds the plastic limit, the value range of β1 is expanded from [0.05, 0.1] to [0.1, 0.15], and the two-parameter reduction algorithm is activated to ensure the real-time adaptability of the shear strength degradation factor.

[0013] Preferably, the settlement differential model is constructed based on the viscoelastic-plastic constitutive equations:

[0014]

[0015] Where x is the back stress, Y is the yield stress, the back stress and yield stress are obtained from triaxial creep test and triaxial compression test, and the relaxation time τ is given by τ = η v / G v Determined, η v G is the viscosity coefficient. v The elastic shear modulus, material constants C and n were obtained by fitting data from the steady-state creep stage of triaxial creep tests. The deep neural network adopted a spatiotemporal convolutional architecture with a hollow convolutional kernel. The input layer was a coupled tensor of the correlation matrix and stress state parameters. The convolutional kernel size was 3×3×3 with a void ratio of 2. The number of network layers was set to 5. The loss function adopted a weighted combination of mean square error and Huber loss with a weight ratio of 7:3. The training dataset included compaction degree, groundwater level, and load time history data from historical engineering cases. The adjacency matrix encoding was based on the spatial distance threshold between nodes. When the distance between two nodes was less than 3 times the average particle size of the weak area, the adjacency matrix element was 1, otherwise it was 0. The graph convolutional layer adopted a message passing mechanism to aggregate the strain rate feature values ​​of neighboring nodes and extract long-range spatial correlation features through the ReLU activation function. The traffic load time history phase information was converted into frequency domain components through short-time Fourier transform and superimposed with the spatial topological features of the settlement distribution map to generate a settlement distribution map with fused phase delay effect.

[0016] The preferred method is to optimize the stress residual R using triple residual comparison. This is calculated as the Euclidean distance between the predicted stress and the measured stress, and the displacement residual R. u It is obtained through the L2 norm difference of the nodal displacement vectors; specifically: R total =ω1·||R σ|| 2 +ω2·||R ∈ || 2 +ω3·||R u || 2 The weight coefficient ω is adjusted based on the eigenvalue decomposition of the covariance matrix, and the first three eigenvectors are extracted by principal component analysis. The weight allocation ratio is the normalized result of the eigenvalue proportions. The objective function is defined as the weighted sum of squares of the comprehensive residuals, and the iteration termination condition is that the rate of decrease of the comprehensive residuals is less than 1 × 10⁻⁶. -4 The number of steps or the maximum number of iterations reaches 1000; the similarity coefficient is calculated by converting the predicted weak area and the measured abnormal area into a binary spatial mask. When the ratio of the intersection area to the union area exceeds 0.75, the spatial overlap is judged to meet the standard. The spatial alignment between the measured data and the model output adopts the affine transformation registration algorithm to ensure strict matching of the grid coordinate system.

[0017] Preferably, the settlement early warning map is constructed to intuitively reflect the settlement risk level of different areas through a color coding system. Based on the predicted settlement distribution map, the settlement rate of each area is compared with the design allowable value, and the early warning level is divided according to the rate ratio. When the maximum settlement rate of a certain area does not exceed the allowable value, it is marked as a safe level and represented by color. If the rate ratio reaches the preset first threshold, it is upgraded to a low-risk level and switched to yellow. As the ratio further increases, orange, red and dark red are used to mark the medium-risk, high-risk and extremely high-risk levels respectively.

[0018] This invention provides an intelligent prediction method and system for differential settlement of roadbed during highway widening, which has the following beneficial effects:

[0019] 1. This invention achieves spatiotemporal continuous characterization of the mechanical response of weak roadbed areas by constructing a dynamic analysis system with multi-dimensional physical field coupling. It breaks through the limitations of traditional homogenization models, significantly improves the spatiotemporal resolution and dynamic adaptability of differential settlement prediction, and improves the accuracy of abnormal area identification, especially in complex geological scenarios.

[0020] 2. This invention enables the model to autonomously correct the spatial distribution weights of parameters in weak areas through collaborative optimization, effectively overcoming the prediction bias caused by data heterogeneity in traditional methods, and significantly improving the spatial overlap between its early warning results and the actual subsidence anomaly areas compared to conventional methods.

[0021] 3. This invention, through a settlement risk classification and early warning system based on spatial topological correlation, can generate highly reliable settlement evolution information in real time during the construction and operation phases, thereby shortening the response time for engineering risk prevention and control and providing scientific decision-making support for the full life-cycle safety management of highway widening projects. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] In highway widening projects, to achieve intelligent prediction of differential settlement of the roadbed, the system first collects real-time data on compaction distribution, groundwater level changes, and traffic load time history curves through a sensor network deployed at the junction of the old and new pavements. The compaction data is normalized based on grid nodes, groundwater level data is spatially interpolated to the same grid system, and traffic load is converted into equivalent nodal loads based on vehicle axle load distribution. For example, if the sensor measures a compaction threshold of 92% for a certain road section, areas below this value are marked as potentially weak areas. During data fusion, the weighting coefficients for compaction, groundwater level, and traffic load are set to 0.6, 0.3, and 0.1, respectively, to ensure the physical balance of multi-source heterogeneous data fusion. Subsequently, the system uses a density clustering algorithm to generate a weak area marker layer, based on the geometric characteristics of areas with compaction below the set threshold. The center is used as the initial seed point, and spatial clustering analysis is performed based on the extreme points of the second derivative of the stress state parameters. For example, the geometric center coordinates of a weak zone are (x = 50m, y = 10m, z = 2m), the neighborhood radius is set to 1.2 times the average grid size, and the minimum number of samples is 5. This ultimately generates a three-dimensional weak zone topology with a thickness gradient. In the finite element mesh generation, the grid size at the boundary of the weak zone is reduced to one-fifth to one-third of that in the regular area, with the specific ratio dynamically adjusted by the local stress gradient amplitude. When the local stress gradient exceeds 10 kPa / m, adaptive mesh refinement is triggered to ensure accurate coupling between the pore water pressure gradient and the stress field. The dynamic loading of the pore water pressure gradient uses cubic spline interpolation, with interpolation nodes distributed along the groundwater level monitoring points. The interpolation function satisfies continuity, and the residual tolerance is set to 1 × 10⁻⁶. -4 kPa, ultimately forming a three-dimensional tensor field model containing stress state parameters.

[0024] Based on this, the system maps the spatial coordinates of the weak zone marker layer to the filler particle size distribution parameters; it uses bilinear interpolation to transfer parameters between grid nodes and establishes a nonlinear mapping relationship between coarse particle content and creep parameters in the filler gradation curve; for example, the mass percentage C of coarse particles in a certain area... g The gradation sensitivity coefficient λ was calibrated to 0.3 by indoor triaxial creep tests, and the creep parameter μ was calculated as μ=μ0·exp(λ·C g )=1.2μ0; The dynamic allocation of creep parameter weights is based on the Euclidean distance function from the node to the centroid of the coarse-grained enrichment region, and the calculation formula is: Where dmax is the maximum radial dimension of the weak zone, for example, 5m; the implicit time integration step Δt is set to 1 hour, and the convergence condition is that the strain increment residual norm is less than 1×10. -6 The second deviatoric stress invariant J2 is selected as the coupling reference for the deviatoric stress tensor components. After multiplying with the weighting coefficients, it is input into the viscoelastic constitutive equation to calculate the anisotropic creep strain increment. For example, when J2 = 50 kPa and the viscoelastic shear modulus G... v When Δε = 1000 kPa, c =Δt(J2w i ) / (2Gv)=0.00125, generating a time-varying strain rate distribution band and superimposing it onto the weak area marker layer to form a creep cloud map.

[0025] Subsequently, the system extracts the strain rate eigenvalues ​​corresponding to the weak zone marker layer by matching the creep contour map with the mesh nodes of the three-dimensional stress field model; it extracts the principal strain rate gradient direction using the eigenvalue decomposition method, and the shear strain rate direction angle is obtained by calculating the ratio of shear strain components using the arctangent function; based on this, a two-dimensional correlation matrix is ​​established with the groundwater level fluctuation amplitude as the row vector and the traffic load spectral density as the column vector; the shear strength degradation factor η is calculated using a two-parameter reduction algorithm, and the formula is as follows:

[0026] η=η0·[1-β1·tanh(α·Λh)][1-β2·ln(1+γ·S(f))],

[0027] Where η0 is the initial shear strength, β1 is the moisture content sensitivity coefficient, β2 is the load application number sensitivity coefficient, and α and γ are fitting parameters, obtained through least squares optimization driven by experimental data; for example, the initial shear strength η0 = 100 kPa in a certain weak zone, the moisture content sensitivity coefficient β1 = 0.1, the load application number sensitivity coefficient β2 = 0.003, and the fitting parameters α = 1.2 and γ = 0.8; when the moisture content exceeds the plastic limit of 25%, the value range of β1 is expanded from [0.05, 0.1] to [0.1, 0.15], and the two-parameter reduction algorithm is activated at the same time; if the moisture content of a certain weak zone is m = 30% and the load application number N = 1000 times, then η = 82 kPa is calculated, realizing the dynamic correction of the shear strength factor.

[0028] The modified correlation matrix is ​​coupled with stress state parameters to construct a settlement differential model containing the viscoelastic-plastic constitutive equations of the weak zone; the relaxation time τ is determined by the viscosity coefficient η. v With elastic shear modulus G v The ratio is determined, for example, η. v =500kPa / h, G v=1000kPa, τ=0.5h; material constants C and n are obtained by fitting steady-state creep stage data from triaxial creep tests, for example, C=0.05, n=1.5; spatiotemporal convolution uses a deep neural network with dilated convolution kernels, the input layer is a coupled tensor of the correlation matrix and stress state parameters, the convolution kernel size is 3×3×3, the dilation rate is 2, the number of network layers is set to 5, and the loss function is a weighted combination of mean squared error and Huber loss with a weight ratio of 7:3; adjacency matrix encoding is based on the spatial distance threshold between nodes. When the distance between two nodes is less than 3 times the average particle size of the weak area, such as 0.3m, the adjacency matrix element is set to 1; the graph convolutional layer aggregates the strain rate feature values ​​of neighboring nodes through a message passing mechanism and uses the ReLU activation function to extract long-range spatial correlation features; the traffic load time history phase information is converted into frequency domain components through short-time Fourier transform and superimposed with the spatial topological features of the settlement distribution map to generate a settlement distribution map with fused phase delay effect; for example, the predicted settlement rate of a certain area is 5mm / year, the design allowable value is 8mm / year, and it is marked as a safety level.

[0029] The system further optimizes parameters through triple residual comparison using a three-dimensional stress field model, creep contour plot, and settlement distribution map; stress residual R σ The Euclidean distance between the predicted stress and the measured stress is calculated, and the displacement residual Ru is obtained through the difference in the L2 norm of the nodal displacement vectors; the combined residual R total Defined as the weighted Euclidean distance formula: R total =ω1·||R σ || 2 +ω2·||R ∈ || 2 +ω3·||R u || 2 The weighting coefficients ω are dynamically adjusted based on the eigenvalues ​​of the covariance matrix of each residual. For example, after the eigenvalue decomposition of the covariance matrix, the proportions of the first three eigenvectors are 50%, 30%, and 20%, respectively, then the weighting coefficients ω1 = 0.5, ω2 = 0.3, and ω3 = 0.2. During the optimization process, the conjugate gradient method is used to synchronously update the spatial resolution parameters and creep parameters of the weak region marker layer. The iteration termination condition is that the overall residual decrease rate is less than 1 × 10⁻⁶. -4 The maximum number of iterations can reach 1000. An early warning condition is triggered when the spatial similarity coefficient between the predicted weak area and the measured settlement anomaly area exceeds 0.75. The similarity coefficient is calculated by converting the predicted and measured areas into binary spatial masks, and then calculating the ratio of the intersection area to the union area. For example, if the intersection area is 120m²... 2 If the area of ​​the union is 160m2, then the similarity coefficient is 0.75, which satisfies the threshold condition.

[0030] Ultimately, the system generates a settlement early warning map, which intuitively reflects the settlement risk level of different areas through HSV color space encoding. The warning level is determined based on the ratio of the predicted settlement rate to the design allowable value: when the maximum settlement rate of an area does not exceed the allowable value, it is marked with a safe color; if the ratio reaches 1.0 to 1.2, it is marked yellow, indicating low risk; 1.2 to 1.5 is orange, indicating medium risk; 1.5 to 2.0 is red, indicating high risk; and exceeding 2.0 is dark red, indicating extremely high risk. For example, if the predicted settlement rate of a road section is 12 mm / year, the design allowable value is 10 mm / year, and the ratio is 1.2, the system marks this area as orange (medium risk), indicating the need for reinforcement measures. Through this process, the system realizes a complete chain logic from data acquisition, model building, parameter optimization to early warning output, providing scientific support for the full life-cycle safety management of highway widening projects.

[0031] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart prediction method for differential settlement of roadbed during highway widening, characterized in that, Includes the following steps: Based on real-time acquisition of compaction distribution data, groundwater level change data, and traffic load time history curves at the junction of new and old road surfaces using sensors, a three-dimensional stress field model with a weak zone marker layer is constructed. The three-dimensional stress field model synchronously outputs the stress state parameters of each grid node; where the grid node is the basic unit of the three-dimensional stress field model. The weak zone marker layer is mapped to the filler particle size distribution parameters. The creep parameter weights are dynamically allocated based on the coordinates of the weak zone marker layer. A creep cloud map is generated by combining the stress state parameters. The creep cloud map contains a time-varying strain rate distribution band that overlaps with the weak zone marker layer. By matching the mesh nodes of the creep cloud map with the three-dimensional stress field model, the strain rate characteristic value corresponding to the weak zone marker layer is extracted, and a correlation matrix of traffic load-shear strength degradation with groundwater level fluctuation amplitude as constraint condition is established. The correlation matrix is ​​dynamically corrected by a two-parameter reduction algorithm based on the water content sensitivity of the weak zone. The modified correlation matrix is ​​coupled with the stress state parameters to construct a settlement differential model containing the viscoelastic-plastic constitutive equations of the weak region. A settlement distribution map that integrates the spatial topology of the weak region marker layer is generated by spatiotemporal convolution. By comparing the triple residuals of the three-dimensional stress field model, creep cloud map, and settlement distribution map, the spatial resolution and creep parameter weights of the weak zone marker layer are simultaneously optimized. The triple residuals include: the stress residual of the three-dimensional stress field model, the strain residual of the creep cloud map, and the displacement residual of the settlement distribution map. The residuals are obtained by comparing the measured data collected by the sensor in real time with the model output value. When the spatial overlap between the weak area marker layer and the measured settlement anomaly area reaches a set threshold, a road settlement early warning map is output.

2. The intelligent prediction method for differential settlement of highway widening subgrade according to claim 1, characterized in that: The construction of the three-dimensional stress field model involves fusing multi-source heterogeneous sensor data to establish a three-dimensional stress field model with a finite element mesh centered on the weak zone marker layer. The weak zone marker layer is generated by using the geometric center of the region below a set threshold in the compaction distribution data as the seed point, and performing spatial clustering analysis based on the extreme points of the second derivative of the stress state parameters to generate a three-dimensional topological structure of the weak zone with thickness gradient characteristics. The three-dimensional stress field model sets the mesh size at the boundary of the weak zone marker layer to one-fifth to one-third of that of the conventional region, and uses cubic spline interpolation to dynamically load the pore water pressure gradient caused by groundwater level changes, forming a three-dimensional tensor field containing stress state parameters.

3. The intelligent prediction method for differential settlement of highway widening subgrade according to claim 1, characterized in that: The creep cloud map is constructed by mapping the spatial coordinates of the weak zone marker layer to the filler particle size distribution parameters using an interpolation-based coordinate transformation to establish a nonlinear mapping relationship between the coarse particle content and creep parameters in the filler gradation curve. The dynamic allocation of creep parameter weights is achieved by defining an Euclidean distance function from nodes within the weak zone marker layer to the nearest coarse particle enrichment zone, specifically expressed as: , where w i Let x be the creep parameter weight for node i, λ be the gradation sensitivity coefficient, and x be the creep parameter weight for node i. i Let x be the spatial coordinates of node i within the weak zone marker layer. c Let d be the centroid coordinates of the coarse-grained enrichment region. max The maximum radial dimension of the weak zone is defined as follows: the time-varying strain rate distribution band is generated by coupling the deviatoric stress tensor component in the stress state parameters with the weighting coefficient and inputting it into the viscoelastic constitutive equation, and then calculating the anisotropic creep strain increment through an implicit time integration algorithm.

4. The intelligent prediction method for differential settlement of highway widening subgrade according to claim 1, characterized in that: The correlation matrix is ​​constructed by extracting the principal strain rate gradient and shear strain rate direction angle from the strain rate eigenvalues, establishing a two-dimensional matrix with groundwater level fluctuation amplitude as the row vector and traffic load spectral density as the column vector; the shear strength degradation factor η is calculated using a two-parameter reduction algorithm. Where η0 is the initial shear strength, β1 is the moisture content sensitivity coefficient, β2 is the load application number sensitivity coefficient, α and γ are fitting parameters, Δh is the groundwater level fluctuation amplitude, and S(f) is the traffic load spectral density; the dynamic correction process constructs a negative exponential relationship between the moisture content of the weak zone and the matrix suction, adjusts the value range of β1 in real time, and activates the two-parameter reduction algorithm when the moisture content exceeds the plastic limit.

5. The intelligent prediction method for differential settlement of highway widening subgrade according to claim 1, characterized in that: The settlement differential model is constructed by using the degradation factor in the correlation matrix as input parameters to establish a set of viscoelastic-plastic constitutive equations considering strain hardening and softening: ； Where x is the back stress and Y is the yield stress. Let C be the relaxation time, n be the material constants, σ ​​be the stress tensor, t be time, and E be the elastic modulus. The total strain rate, The plastic strain rate is given by sign(x), which is the sign function. The spatiotemporal convolution uses a deep neural network with dilated convolution kernels to encode the spatial topology of the weak area marker layer into an adjacency matrix. Long-range spatial correlation features are extracted through graph convolution layers, and finally, a settlement distribution map integrating traffic load temporal phase information is output.

6. The intelligent prediction method for differential settlement of highway widening subgrade according to claim 1, characterized in that: The triple residual comparison is performed by defining the stress residual R in the three-dimensional stress field model. σ The strain residual R of the creep contour plot ϵ and the displacement residual R of the settlement distribution map u The weighted Euclidean distance formula is used to calculate the overall residuals: The weight coefficient ω is dynamically adjusted based on the eigenvalues ​​of the covariance matrix of each residual. During the optimization process, the conjugate gradient method is used to synchronously update the spatial resolution parameters and creep parameters of the weak region marker layer. When the similarity coefficient between the weak region marker layer and the measured abnormal region exceeds 0.75, the spatial overlap threshold condition is triggered.

7. The intelligent prediction method for differential settlement of highway widening subgrade according to claim 1, characterized in that: The settlement early warning map is constructed by aligning the predicted results of the settlement distribution map with historical monitoring data in time and space, and using HSV color space encoding to generate a multi-layer overlay map with early warning levels; the early warning level is determined based on the ratio of the maximum settlement rate to the allowable value within the weak zone marker layer.

8. A system for implementing the intelligent prediction method for differential settlement of highway widening subgrade as described in any one of claims 1-7, characterized in that, include: Multi-source data acquisition unit: Based on real-time acquisition of compaction distribution data, groundwater level change data and traffic load time history curve at the junction of new and old road surfaces by sensors, a three-dimensional stress field model with a weak zone marker layer is constructed. The three-dimensional stress field model synchronously outputs the stress state parameters of each grid node. Creep analysis unit: Maps the weak zone marker layer to the filler particle size distribution parameters, dynamically allocates creep parameter weights based on the coordinates of the weak zone marker layer, and generates a creep cloud map in combination with stress state parameters. The creep cloud map contains a time-varying strain rate distribution band that overlaps with the weak zone marker layer. Coupled correction unit: By matching the mesh nodes of the creep cloud map and the three-dimensional stress field model, the strain rate characteristic value corresponding to the weak zone marker layer is extracted, and the correlation matrix of traffic load-shear strength degradation with groundwater level fluctuation amplitude as constraint condition is established. The correlation matrix is ​​dynamically corrected by a two-parameter reduction algorithm based on the water content sensitivity of the weak zone. Settlement prediction unit: The modified correlation matrix is ​​coupled with the stress state parameters to construct a settlement differential model containing the viscoelastic-plastic constitutive equations of the weak zone. The settlement distribution map is generated by spatiotemporal convolution, which integrates the spatial topology of the weak zone marker layer. The system simultaneously optimizes the spatial resolution and creep parameter weights of the weak zone marker layer by comparing the triple residuals of the three-dimensional stress field model, creep cloud map, and settlement distribution map. When the spatial overlap between the weak zone marker layer and the measured settlement anomaly area reaches a set threshold, the system outputs a pavement settlement early warning map.

Images