Sluggy soil processing effect data prediction method based on deep learning

Abstract

The invention discloses a muddy soil treatment effect data prediction method based on deep learning, and particularly relates to the technical field of treatment effect prediction. The problems that in the solidification and drainage process of mucky soil, the pore structure evolution is complex, and permeability mutation is difficult to predict are solved; the method comprises the following steps: scanning a target mucky soil sample in different solidification stages to generate a spatial structure voxel model, extracting pore topological parameters, and constructing a pore multilevel communication feature tensor; identifying critical mutation pore nodes through a graph traversal algorithm, and extracting corresponding pore geometric variation characteristic quantities; inputting the pore geometric variation characteristic quantity into a graph neural network model, and outputting a microcosmic pore structure mutation risk value; a multi-modal attention mechanism is introduced in combination with soil body solidification environment parameters, and a permeability coefficient transition prediction value is generated; according to the permeability coefficient transition prediction value, drainage scheme optimization solution is executed, and the prediction precision and drainage decision reasonability in the sludge soil foundation treatment process are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of treatment effect prediction technology, and more specifically, to a method for predicting the treatment effect of silty soil based on deep learning. Background Technology

[0002] Existing methods for predicting the treatment effect of silty soil mostly rely on macroscopic empirical parameters or statistical models, which are insufficient to reflect the complex evolution of the microscopic pore structure of the soil during solidification and drainage. Because the pore structure of silty soil exhibits non-uniformity and instability at different solidification stages, and the connectivity of the pore network has a non-linear amplification effect on permeability, existing methods generally neglect the crucial influence of critical changes in connectivity within the microstructure on the formation of permeability paths. This results in insufficient accuracy in predicting permeability performance when faced with abrupt changes in local structure (such as pore closure or connectivity breach), making it difficult to meet the requirements of accurate prediction of drainage behavior and solidification effects in engineering projects.

[0003] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a deep learning-based method for predicting the treatment effect data of silty soil to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based method for predicting the treatment effect of silty soil includes the following steps: S1: Scan the target silty soil sample to obtain microscopic images of the soil at different solidification stages, and generate a spatial structure voxel model containing pore distribution through a three-dimensional image reconstruction algorithm. S2: Extract pore topology parameters based on spatial structure voxel model, apply maximum sphere algorithm to identify pore channel connectivity structure, and construct pore multi-level connectivity feature tensor; S3: Perform seepage path search analysis on the multi-level connectivity feature tensor of pores, locate critically abrupt ... S4: Input the pore geometric variation feature quantity and the seepage path dataset into the pre-trained graph neural network model, and output the mutation risk value of the micropore structure; S5: Based on the mutation risk value of micropore structure and soil solidification environment parameters, the prediction weight is adjusted by using a multimodal attention mechanism to generate permeability coefficient transition prediction value; S6: Optimize the drainage scheme based on the predicted value of the permeability coefficient transition.

[0006] In a preferred embodiment, S1 specifically refers to: Multi-angle tomographic scanning was performed on the target silty soil sample, and microscopic slice images of the soil were acquired at the initial solidification stage, the intermediate solidification stage and the final solidification stage. Spatial registration is performed on the microscopic slice images acquired at each curing stage to generate a three-dimensional grayscale matrix; Noise suppression is applied to the three-dimensional grayscale matrix to remove scanning artifacts and preserve the true pore boundaries; The denoised 3D grayscale matrix is ​​divided into a solid phase voxel set and a pore voxel set; A 3D image reconstruction algorithm is executed based on the solid voxel set and the pore voxel set to generate a spatial structure voxel model containing the pore distribution.

[0007] In a preferred embodiment, S2 specifically refers to: Within the spatial structure voxel model, the pore voxel set is labeled with connected components to identify each independent connected sub-region; The maximum inscribed sphere algorithm is applied to each independent connected subregion to locate the key connected sphere and calculate the corresponding radius. Extract the shape factor, connected path length, and aperture distribution index for each key connected sphere; The shape factor, connected path length, critical connected sphere radius, and aperture distribution index are combined into a multi-level connectivity feature tensor based on the porosity connectivity hierarchy.

[0008] In a preferred embodiment, S3 specifically refers to: A seepage network graph is constructed based on the multi-level connectivity feature tensor of pores. The voxels in each independent connected sub-region are identified as nodes of the network graph, and the connectivity between nodes is identified as edges of the network graph. A graph traversal algorithm is performed on the seepage network diagram to identify critical abrupt change pore nodes in the seepage network diagram using a depth-first traversal approach. Data is recorded on critically abruptly changed pore nodes and adjacent connection information to generate a seepage path dataset. Extract the pore geometric variation feature corresponding to each critical mutation pore node from the seepage path dataset.

[0009] In a preferred embodiment, S4 specifically refers to: Construct a node feature matrix and an edge feature matrix based on each critical mutation pore node and its corresponding pore geometric variation feature in the seepage path dataset. The node feature matrix and edge feature matrix are input into the graph convolutional layer of the pre-trained graph neural network model, and node information fusion is achieved through layer-by-layer neighborhood aggregation operation. The deep feature representations output by the graph neural network model are sequentially mapped through fully connected layers and regression layers to obtain the microscopic pore structure mutation risk value corresponding to each critical mutation pore node.

[0010] In a preferred embodiment, S5 specifically refers to: The microscopic pore structure mutation risk value is matched with the soil solidification environmental parameters of the corresponding solidification stage in time to construct a joint feature matrix of microscopic risk-environmental parameters. The joint feature matrix of micro-risk-environment parameters is input into a deep prediction network containing a multimodal attention mechanism, and feature transformation and attention weighting are performed in each modal branch respectively; By fusing the attention weighting results of each modal branch, a fused prediction feature vector covering different solidification stages and different connectivity structure types is generated; The fused prediction feature vector is input into the regression subnetwork, and the predicted value of the penetration coefficient transition at the corresponding time step is output.

[0011] In a preferred embodiment, S6 specifically refers to: An objective function for optimizing drainage schemes is constructed using the predicted values ​​of permeability coefficient transitions at different time steps as input conditions. The objective function for optimizing the drainage scheme is set, and optimization constraints are set. The optimization constraints include the stage characteristics of the soil solidification process, the risk value of abrupt changes in the micropore structure, the relationship between the predicted value of the permeability coefficient transition and the corresponding time step. The optimization algorithm is used to optimize the objective function of the drainage scheme to obtain an optimized drainage scheme that meets the optimization constraints. The optimized drainage scheme includes the drainage layout location, drainage rate, and drainage duration.

[0012] The technical effects and advantages of the deep learning-based data prediction method for silty soil treatment effects of this invention are as follows: By acquiring microscopic images of silty soil at different solidification stages and reconstructing spatial structure voxel models, the dynamic evolution characteristics of pore structures can be accurately characterized, facilitating the identification of spatial heterogeneity of microstructures during solidification. Extracting pore topological parameters and applying the maximum inscribed sphere algorithm to construct a multi-level connectivity feature tensor of pores allows for the quantitative description of the connectivity hierarchy and geometric characteristics of pore channels. Utilizing graph traversal algorithms to identify critical abrupt change pore nodes and extract their variational features helps capture local structural changes that trigger permeability transitions. Combining abrupt change risk value generated by a graph neural network model enables in-depth modeling of microstructural stability. Introducing a multimodal attention mechanism to fuse abrupt change risk value with soil solidification environment parameters significantly enhances the ability to perceive permeability transition trends at different solidification stages. Optimizing drainage schemes based on permeability coefficient transition predictions allows for precise control of drainage strategies, improving solidification efficiency and reducing engineering risks. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the deep learning-based data prediction method for the treatment effect of silty soil according to the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0015] Example Figure 1 This invention presents a data prediction method for the treatment effect of silty soil based on deep learning, which includes the following steps: S1: Scan the target silty soil sample to obtain microscopic images of the soil at different solidification stages, and generate a spatial structure voxel model containing pore distribution through a three-dimensional image reconstruction algorithm. S2: Extract pore topology parameters based on spatial structure voxel model, apply maximum sphere algorithm to identify pore channel connectivity structure, and construct pore multi-level connectivity feature tensor; S3: Perform seepage path search analysis on the multi-level connectivity feature tensor of pores, locate critically abrupt ... S4: Input the pore geometric variation feature quantity and the seepage path dataset into the pre-trained graph neural network model, and output the mutation risk value of the micropore structure; S5: Based on the mutation risk value of micropore structure and soil solidification environment parameters, the prediction weight is adjusted by using a multimodal attention mechanism to generate permeability coefficient transition prediction value; S6: Optimize the drainage scheme based on the predicted value of the permeability coefficient transition.

[0016] The target silty soil sample was scanned to obtain microscopic images of the soil at different solidification stages. A spatial structure voxel model containing the pore distribution was generated using a 3D image reconstruction algorithm, including: Multi-angle tomographic scanning was performed on the target silty soil sample, and microscopic slice images of the soil were acquired at the initial solidification stage, the intermediate solidification stage and the final solidification stage. Industrial computed tomography (CT) equipment was used to perform tomographic scanning on the target silty soil samples. The equipment was equipped with a high-precision detector and a moving tray, achieving a scanning accuracy of no less than 1 / 20th of a millimeter per pixel. The soil samples were placed in a fixed support, and a robotic arm control device positioned the samples at the scanning center. The scanning process was conducted at three time points: the first time point was the initial curing stage, with the first scan performed 12 hours after the start of treatment; the second time point was the intermediate curing stage, with the second scan performed 18 hours after the start of treatment and the curing agent was applied; and the third time point was the final curing stage, with the third scan performed 36 hours after the start of treatment. Each scan was performed at a fixed angular interval, rotating once every one degree of rotation, resulting in a total of 360 consecutive tomographic images. Each image had a resolution of 2048 x 2048 pixels and a grayscale depth of 16 bits. Environmental parameters such as sample temperature, humidity, and curing agent concentration were recorded during the scanning process for image correction.

[0017] Spatial registration is performed on the microscopic slice images acquired at each curing stage to generate a three-dimensional grayscale matrix; The microscopic slice images from the three time points above are input into the image registration processing module. The registration processing module uses a three-dimensional rigid registration algorithm based on the mutual information criterion. The configuration is as follows: The slice image acquired at the first time point is used as the reference image, and the slice image acquired at the second time point is imported into the registration module. First, a two-dimensional affine transformation estimation is performed between adjacent layers, including translation, rotation, and scaling components. The objective function is to maximize the mutual information of image grayscale. The transformation parameters are adjusted iteratively using gradient descent, with the iteration terminating when the rate of change of the transformation parameters is less than one-hundredth or the mutual information gain is less than one ten-thousandth.

[0018] After two-dimensional registration, the registered second-time-node slices are stacked with the reference image along the slice direction, and a three-dimensional interpolation algorithm is used to connect the continuous slices into volumetric data. The above steps are repeated to register the third-time-node slices with the reference image. Through multiple iterations, a spatially consistent set of grayscale images is finally generated from the slice data of the three time nodes in a unified spatial coordinate system.

[0019] Based on the registration results, the slice data of each time node are arranged in the slice order, and the corresponding pixel gray values ​​are organized into a three-dimensional gray matrix.

[0020] Noise suppression is applied to the three-dimensional grayscale matrix to remove scanning artifacts and preserve the true pore boundaries; To address potential artifacts such as metal ring artifacts, motion blur noise, and detector scattering artifacts during the scanning process, a three-dimensional wavelet transform denoising algorithm is applied to the three-dimensional grayscale matrix.

[0021] A three-dimensional discrete wavelet transform is used to decompose the three-dimensional grayscale matrix into several high-frequency and low-frequency sub-bands. Symmetrical orthogonal wavelet basis functions are selected for decomposition to ensure the continuity of the image boundaries after reconstruction. The number of decomposition levels is set to three.

[0022] Within each high-frequency sub-band, a soft thresholding process is applied based on the noise statistics. The threshold value is calculated according to the sub-band's root mean square noise level, and the threshold formula is: threshold equals the noise standard deviation multiplied by the square root of two logarithmic constant. The soft thresholding function subtracts the threshold from coefficients greater than it, and retains coefficients less than the threshold as zero.

[0023] The high-frequency sub-band and low-frequency sub-band, after soft thresholding, are fused using inverse three-dimensional wavelet transform to generate a denoised three-dimensional grayscale matrix. Comparing this to the original grayscale matrix, structural similarity indices show that the pore boundary detail retention rate is no less than 95%.

[0024] The denoised 3D grayscale matrix is ​​divided into a solid phase voxel set and a pore voxel set; The denoised 3D grayscale matrix is ​​segmented into 3D voxels using a 3D region growing and fusion global thresholding method. Based on gray-level histogram analysis, the globally optimal threshold of the denoised 3D gray-level matrix is ​​first calculated. A gray-level distribution probability density function is established using a bimodal hypothesis model, and the threshold point is determined using the maximum inter-class variance criterion. The globally optimal threshold best separates solid voxels from porous voxels.

[0025] Using the global optimal threshold as an initial condition, voxels with gray values ​​higher than the global optimal threshold are labeled as solid phase candidate voxels in three-dimensional space, and voxels with gray values ​​lower than the global optimal threshold are labeled as pore candidate voxels. In the gray-level matrix at each time point, starting from multiple predefined seed points, a neighborhood traversal is performed on voxels in adjacent hexahedral directions. When the gray value difference between the adjacent voxel and the seed point is less than a set difference (the difference is set to 25 gray levels), the adjacent voxel is assigned to the corresponding category, and the traversal of its neighborhood voxels continues.

[0026] The region growth results were subjected to hole filling and small pseudo-connected region removal: First, a three-dimensional connectivity analysis was performed on the pore voxel set, and regions with fewer than two hundred voxels were considered pseudo-connected regions and removed. Then, a morphological closing operation was performed on the secondary noise points appearing at the boundary between the solid phase voxel set and the pore voxel set, with a radius set to three voxels, to eliminate hole artifacts. Finally, a pure solid phase voxel set and a pure pore voxel set were obtained.

[0027] A three-dimensional image reconstruction algorithm is executed based on the solid phase voxel set and the pore voxel set to generate a spatial structure voxel model containing the pore distribution. After voxel segmentation, the solid voxel set and the porous voxel set are fused into a spatial structure voxel model using a 3D reconstruction algorithm. The resulting solid-phase voxel sets and pore voxel sets are then meshed in three dimensions according to a fixed voxel size (voxel side length of 0.01 mm). The position of each voxel element is uniformly represented by the voxel center coordinates.

[0028] An improved version of the 3D traveling cubes algorithm is used to extract the solid surface from a set of solid voxels. The 3D traveling cubes algorithm first classifies the scalar field values ​​at each voxel cube node, then searches for a standard 3D traveling cube surface template based on the classification, generating triangular mesh fragments. All voxel cubes are traversed, the triangular mesh fragments are pieced together, and redundant adjacent triangular faces are removed, ultimately forming a complete solid surface mesh.

[0029] Based on a solid-state surface mesh, the connectivity strength between pore voxels is predicted at their centers using Gaussian fuzzing by expanding the boundary information of the pore voxel set. Specifically, for each pore voxel center, a weighted sum of adjacent voxels is performed using a 3D Gaussian kernel function, with a weighted function parameter having a weight decay radius of 0.05 mm. Based on the weighted results, the connectivity between adjacent pore voxels is determined. By traversing all pore voxels, pore connectivity information is obtained, and the pore connectivity boundaries are represented by thin triangular facets.

[0030] The solid surface mesh is fused with the pore connectivity boundary to form a spatial structure voxel model that includes the pore distribution. The spatial structure voxel model records the soil microstructure using 3D mesh nodes and mesh cell patches. Mesh nodes are labeled with attributes including solid phase or pore type, voxel center coordinates, and pore connectivity status. The spatial structure voxel model is saved in a standard 3D mesh data format for easy pore topology analysis and seepage simulation.

[0031] Based on the spatial structure voxel model, pore topological parameters are extracted, and the maximum sphere algorithm is applied to identify the pore channel connectivity structure. A multi-level connectivity feature tensor for pores is constructed, including: Within the spatial structure voxel model, the pore voxel set is labeled with connected components to identify each independent connected sub-region; The spatial structure voxel model stores the attribute identifiers of each voxel in a 3D mesh format, including the voxel center coordinates and voxel type information (solid phase or porosity). A 3D voxel connected component labeling algorithm is applied to the porosity voxel set.

[0032] Adjacency is determined based on the distance between voxel centers in units of side length (one voxel side length), and connectivity is determined using the hexahedral neighborhood rule. Specifically, if the difference between the center of a pore voxel and the center of an adjacent voxel in each of the three coordinate directions is equal to one voxel side length, then the two are considered adjacent.

[0033] Scan the entire set of pore voxels, starting with the first pore voxel that has not yet been identified as a connected domain, and use it as the seed voxel of the first connected domain, and assign the connected domain the first number.

[0034] Starting with the seed voxel, traverse all unidentified pore voxels in the neighborhood that satisfy the adjacency condition, incorporate these neighborhood voxels into the same connected component number, and continue repeating this step until all voxels in the current connected component have been traversed.

[0035] After completing the labeling of the current connected component, if there are still unlabeled pore voxels, the next unlabeled pore voxel is used as the seed of the new connected component, assigned a second number, and the region recursive expansion process is repeated until all pore voxels have completed the labeling of connected components.

[0036] After the connected components are labeled, all pore voxels are stored in a data structure according to their respective connected component numbers. Each set of connected component voxels is matched one-to-one with its number, resulting in a set of pore voxels for several independent connected sub-regions.

[0037] The maximum inscribed sphere algorithm is applied to each independent connected subregion to locate the key connected sphere and calculate the corresponding radius. Based on the set of pore voxels in each independent connected sub-region, a distance transformation method is performed in each connected sub-region to determine the location and radius of the maximum inscribed sphere.

[0038] For a single connected sub-region, all its porosity voxel and solid phase voxel information is imported into the distance transformation module. The distance transformation module first performs a binarization operation on the solid phase voxel set, marking the solid phase voxels as obstacle voxels and the porosity voxels as the voxels to be measured.

[0039] The Euclidean distance transformation algorithm is employed. By scanning the set of porosity voxels layer by layer, the Euclidean distance from the center of each porosity voxel to the center of its neighboring solid phase voxels is calculated, and the minimum distance value of each porosity voxel is recorded. The minimum distance value represents the radius of the largest sphere that can be inscribed around that porosity voxel.

[0040] After the scan is completed, select the voxel with the largest distance value from all the porosity voxels in the current connected sub-region, take the voxel as the center of the key connected sphere, and take its corresponding distance value as the radius of the key connected sphere.

[0041] The distance values ​​of pore voxels are globally sorted, and the top few voxels with radii close to the maximum value are selected as candidate key connected spheres. A local neighborhood check is applied to each candidate voxel, i.e., the distance is recalculated within the radius neighborhood of that voxel, to verify whether the candidate voxel is the true maximum inscribed sphere. Through iterative verification, one or more key connected spheres are finally determined, and their voxel center coordinates and radius information are recorded.

[0042] The coordinates of the center of the key connected sphere and the corresponding radius of each connected sub-region are stored in a data table. The data table records fields including the connected sub-region number, the voxel index of the center sphere, and the maximum inscribed sphere radius value.

[0043] Extract the shape factor, connected path length, and aperture distribution index for each key connected sphere; Based on the coordinates and radius information of the center of the key connected sphere, the shape factor, the length of the connected path, and the aperture distribution index are calculated and stored in the multi-level connectivity feature table of the pores.

[0044] For each key connected sphere, construct several equidistant annular scanning surfaces around the sphere's center. Using the sphere's center coordinates and radius, extract a series of concentric spherical profiles from the grid data, and statistically analyze the boundary morphology of the pore voxels intersecting the spheres on these profiles. The shape factor is derived by comparing the ratio of the number of overlapping voxels between the inscribed sphere voxels and the sphere to the theoretical number of sphere voxels. A shape factor closer to one indicates a more regular overall shape for the corresponding spherical voxel; a value deviating from one indicates a flatter or more elongated pore morphology near the sphere's center.

[0045] Starting with the center voxel of the key connected sphere, a graph traversal method is used to calculate the shortest path length in the porosity voxel network of the current connected sub-region. The nodes in the porosity voxel network use the coordinates of the voxel centers, and edges are defined as the connectivity between adjacent voxel centers within a hexahedral neighborhood. Dijkstra's algorithm is used to traverse the undirected weighted graph. Path weights are calculated by summing the distance between adjacent voxel centers and the reciprocal of the porosity opening area; that is, the weight of an adjacent edge equals the sum of the distance and the reciprocal of the opening area. The final calculated shortest path length refers to the shortest cost distance from the center node to any porosity voxel at the boundary of the connected sub-region. The path length value reflects the ease with which the region containing the key connected sphere connects to external porosities.

[0046] Statistical analysis was performed on the minimum distance values ​​of pore voxels across all pore voxels in the connected sub-regions, based on the distance transformation results. The statistical dimension was divided into several level intervals, categorized by the range of distance values ​​from fewest to most, such as upper, middle, and lower levels. The percentage of pore voxels within each level interval relative to the total number of pore voxels was calculated as a pore size distribution index. This index reflects the width distribution of pore connectivity channels at different scales.

[0047] The shape factor, connected path length, and aperture distribution index, along with the corresponding connected sub-region number, key connected sphere center index, and radius information, are stored in a structured table. Table fields include connected sub-region number, sphere center voxel index, inscribed sphere radius value, shape factor value, connected path length value, and aperture distribution level ratio.

[0048] The shape factor, connected path length, critical connected sphere radius, and aperture distribution index are combined into a multi-level porosity connectivity feature tensor according to the porosity connectivity hierarchy. The feature results of multiple connected sub-regions are summarized into a multidimensional tensor format to form a porosity multilevel connected feature tensor.

[0049] First, each independent connected sub-region is sorted from largest to smallest according to the radius value of the key connected sphere. The connected sub-region with the largest radius value is defined as the first connected level, the second largest as the second connected level, and so on. Each connected level contains at least one key connected sphere.

[0050] The pore multi-level connectivity feature tensor is designed as a four-dimensional tensor format. The first dimension represents the connectivity level number, the second dimension represents the key connected sphere index under the same connectivity level, the third dimension represents the feature type (shape factor, connected path length, inscribed sphere radius, pore size distribution index) corresponding to the key connected sphere, and the fourth dimension represents the value of the feature in the specific numerical dimension.

[0051] To ensure comparability between different features, a unified normalization strategy is used to process the feature values. The normalization strategy is as follows: for each feature, subtract the minimum value of that feature across all connected levels, and then divide by the difference between the maximum and minimum values. After normalization, the value range is uniformly mapped to between zero and one.

[0052] Iterate through all connected levels and their corresponding key connected spheres, and sequentially fill the third and fourth dimensions of the tensor with the normalized shape factor, connected path length, inscribed sphere radius, and aperture distribution index, based on the index position. If multiple key connected spheres exist within the same connected level, they are stored in the tensor in descending order of their radii. If there is only one key connected sphere within a certain connected level, zero values ​​are filled into the corresponding positions in the tensor to maintain consistency with the tensor shape compared to other levels.

[0053] The porosity multi-level connectivity feature tensor is output as a tensor data file by the tensor generation module. The tensor data file is saved in a specified format, such as storing the connectivity level number, key connected sphere index, and four feature normalization values ​​line by line in text format; or directly storing four-dimensional tensor data blocks in binary format.

[0054] A seepage path search analysis is performed on the multi-level connectivity feature tensor of the pores. A graph traversal algorithm is used to locate critically abruptly changed pore nodes and generate a seepage path dataset. The pore geometric variation features of these critically abruptly changed pore nodes are recorded, including: A seepage network graph is constructed based on the multi-level connectivity feature tensor of pores. The voxels in each independent connected sub-region are identified as nodes of the network graph, and the connectivity between nodes is identified as edges of the network graph. The pore multi-level connectivity feature tensor is read and stored in a four-dimensional format. The first dimension represents the connectivity level number, the second dimension represents the index of the key connected spheres within the same connectivity level, the third dimension represents the feature type index (shape factor, connected path length, inscribed sphere radius, pore size distribution index), and the fourth dimension represents the normalized value of the corresponding feature. Based on the pore multi-level connectivity feature tensor, a seepage network diagram is constructed. By traversing the first and second dimensions of the porosity multi-level connectivity feature tensor, a unique node identifier is generated for the voxel center position corresponding to each key connected sphere. The node identifier includes the connectivity level number, the key connected sphere index, and the coordinates of the voxel center in the 3D mesh. By mapping the voxel center coordinates corresponding to the tensor index to the 3D mesh, the precise position of each node in the spatial structure voxel model is obtained.

[0055] For key connected sphere nodes at the same connectivity level, if two nodes are directly spatially adjacent in the original spatial structure voxel model (i.e., the center coordinates of the corresponding voxels of the two nodes are within a hexahedral neighborhood), an undirected edge is established in the seepage network graph. The edge attributes include the Euclidean distance between the two nodes and the connectivity strength of the corresponding neighborhood pores. The pore connectivity strength is calculated using a weighting function, which is the reciprocal of the distance between the center coordinates of adjacent voxels multiplied by the ratio of pore opening areas. This weighting method can reflect the seepage potential of adjacent pore regions.

[0056] To simulate the interconnected seepage paths between different levels, cross-level connections need to be established between consecutive levels. Nearest-neighbor voxel matching is performed on nodes in the upper and lower levels: the coordinates of each key connected sphere node in the first level are traversed, the Euclidean distance is calculated for all second-level node coordinates, and the paired nodes with the smallest distance that meet a threshold condition are selected to establish an edge. The threshold condition is that the distance between the two nodes does not exceed twice the voxel's side length. The weight of cross-level connections is also weighted by the ratio of the inverse distance to the pore opening area. If a node in the same level cannot find a paired node in the lower level that meets the threshold, that node is skipped.

[0057] The constructed seepage network graph is stored in an adjacency list, and a hash table is used to map node identifiers to the list of adjacent nodes. Each list item records the neighboring node identifier, edge weight, and edge spatial direction vector information (for rendering and visualization). The node attribute table records the shape factor value, connected path length value, inscribed sphere radius value, and pore size distribution index value for critical mutation diagnosis.

[0058] The seepage network diagram was copied and displayed using visualization tools, and the positions of the nodes in the 3D mesh were compared with the pore regions in the original spatial structure voxel model. When the positions of the nodes and pore voxels in the visualization result highly coincide, and the direction of the connecting edges is consistent with the actual connection direction of the pores, it indicates that the network diagram was constructed accurately.

[0059] A graph traversal algorithm is performed on the seepage network diagram to identify critical abrupt change pore nodes in the seepage network diagram using a depth-first traversal approach. Critical abrupt change pore nodes are defined as nodes in the seepage network graph where the connectivity strength or radius changes abruptly. These nodes are identified using a depth-first search algorithm. The shape factor, connected path length, inscribed sphere radius, and aperture distribution index of each node are compared with those of its neighboring nodes. A node is considered a candidate for critical mutation when the difference in inscribed sphere radius between it and any neighboring node exceeds a 5% threshold and the change rate of its shape factor exceeds 2%. The percentage difference is calculated by dividing the difference of the same feature value between the two nodes by the larger feature value.

[0060] The percolation network graph is recursively scanned starting from all unvisited nodes in the graph using a depth-first traversal approach. Specifically, an explicit stack structure is used. First, the first critical connected sphere node of the first connected level is pushed onto the stack, and this node is marked as visited in the visit mark table.

[0061] Starting from the top node of the stack, traverse all unvisited adjacent nodes of that node and perform the following operations in sequence: Push adjacent nodes onto the stack and mark them as visited in the access table; For a newly pushed adjacent node, its shape factor and inscribed sphere radius are compared with the parent node of the adjacent node according to the critical mutation judgment rule; if the mutation judgment is met, the adjacent node is recorded as a critical mutation pore node. If the current adjacent node has more unvisited neighbor nodes, repeat the above operation until all nodes in the branch have been visited. Once all neighboring nodes of the current node have been visited, the current node is popped from the stack, and the process returns to the previous node to continue scanning the remaining neighboring nodes.

[0062] After completing the depth-first traversal of all nodes in the first connected level, the same traversal process is performed on the first critical connected sphere node in the second connected level, and each connected level is scanned in turn until all nodes in the entire percolation network graph have been visited.

[0063] During the traversal, the node identifier, associated parent node identifier, mutation time (related to the current traversal order), and triggering feature type (such as inscribed sphere radius or shape factor) of each porosity node identified as a critical mutation node are stored in the critical mutation node table. The fields of the critical mutation node table include node identifier, parent node identifier, mutation feature type, feature value before mutation, and triggering level number of feature value after mutation.

[0064] Data is recorded on critically abruptly changed pore nodes and adjacent connection information to generate a seepage path dataset. After identifying the critical abrupt change pore node, starting from the critical abrupt change pore node and combining it with its adjacent edges, record the complete seepage path information: For each critical mutation pore node, backtrack along the parent node chain from the critical mutation pore node during the depth-first traversal until the initial node with a connectivity level number of one is reached. The initial node can be considered as the node farthest from the processing surface. During the backtracking process, the parent-child node identifier and the corresponding edge weight for each hop are recorded sequentially.

[0065] Create an independent data record for each path. The data record fields include path number, critical mutation node identifier, node sequence list, edge weight list, and total path length. The node sequence list stores all traversed node identifiers in backtracking order, and the edge weight list stores the edge weights between corresponding adjacent nodes. The total path length is obtained by summing all edge weights in the list.

[0066] If a critical mutation node has multiple parent node branches during traversal (i.e., the same node is visited multiple times under different traversal branches), then path backtracking is performed on each branch, and a separate path data record is generated for each branch. The shortest and longest backtracking paths can be determined by comparing the total length of the paths of each branch. All branch paths are stored in the seepage path dataset.

[0067] The generated seepage path dataset is filtered, and records with a total path length significantly greater than twice the sample space size or more than eighty path nodes are removed to avoid abnormal paths interfering with subsequent feature extraction. The path filtering criteria are adjusted according to the actual sample size and pore distribution, with the default maximum path length threshold being twice the sample size.

[0068] The final retained seepage path data records are exported as a structured table. The table fields include path number, starting node identifier, ending node identifier, number of nodes, total path length, node identifier sequence, and edge weight sequence.

[0069] By employing path backtracking and multi-branch processing strategies, the seepage path information of critical mutation nodes and their connected networks is fully recorded, enabling the seepage path dataset to contain both path topology and path weight information, thus providing data support for the extraction of geometric variation features.

[0070] Extract the pore geometric variation feature quantity corresponding to each critical mutation pore node from the seepage path dataset; For the seepage path dataset, the geometric variation features corresponding to the critical mutation nodes before and after the mutation are extracted sequentially, and the results are stored in the pore geometric variation feature table: In the seepage path dataset, each node identifier sequence field contains the identifiers of all nodes on each path. Based on the original normalized feature values ​​in the pore multi-level connectivity feature tensor corresponding to the node identifiers, the inscribed sphere radius and shape factor values ​​of the current critical mutation node before (previous traversal step) and after (current traversal step) the mutation is obtained. The feature value fields before and after mutation are read from the labeling of the critical mutation node entries in the seepage path dataset.

[0071] Based on the total path length field and the edge weights between critical nodes and the next node in the path, the shortest path length from the critical node to the boundary of the connected sub-region is calculated before and after the mutation. Specifically, the connected path length is first extracted from the original pore multi-level connectivity feature tensor, and then the accurate path length is obtained by summing the adjacent edge weights in the seepage path dataset. The path length mutation amount field is recorded, and its value is equal to the path length after the mutation minus the path length before the mutation.

[0072] In the seepage path dataset, the normalized distance values ​​of all pore voxels in the connected sub-regions where the critical node abruptly occurred are statistically analyzed using the node identifier sequence field. By comparing the changes in the maximum, minimum, and average distance values ​​of the connected sub-regions before and after the abrupt change, the pore size distribution abrupt change index is calculated. The normalized distance values ​​of all pore voxels in the connected sub-regions before and after the abrupt change are grouped into equally divided intervals, and the frequency distribution is statistically analyzed. The change in the frequency proportion of each interval is recorded to differentiate whether the overall pore size distribution of the connected sub-region has undergone significant distortion.

[0073] In the seepage path dataset, the shape factor values ​​of critical nodes before and after the mutation are read, and the mutation amount of the shape factor is directly calculated, that is, the value after the mutation minus the value before the mutation. The value reflects the degree to which the shape of the pore voxel corresponding to the mutation node changes from being close to a sphere to becoming flattened or elongated after the mutation.

[0074] Write the following fields into the pore geometric variation characteristic table: node identifier, connectivity level number, inscribed sphere radius before mutation, inscribed sphere radius after mutation, inscribed sphere radius mutation amount, shape factor before mutation, shape factor after mutation, shape factor mutation amount, path length before mutation, path length after mutation, path length mutation amount, and aperture distribution mutation index.

[0075] The pore geometric variation features and seepage path dataset are input into a pre-trained graph neural network model, which outputs a mutation risk value of the micropore structure, including: Construct a node feature matrix and an edge feature matrix based on each critical mutation pore node and its corresponding pore geometric variation feature in the seepage path dataset. Referring to the pore geometry variation feature table in the seepage path dataset, each record includes a node identifier, the mutation amount of the inscribed sphere radius, the mutation amount of the shape factor, the mutation amount of the path length, and the mutation index of the pore size distribution. Each record is converted into an independent feature vector. The feature vectors are arranged in the following order: first, the mutation amount of the inscribed sphere radius is filled in; then, the mutation amount of the shape factor is filled in; then, the mutation amount of the path length is filled in; and finally, the mutation index of the pore size distribution is filled in. All feature vectors are filled into the node feature matrix row by row according to the sorting order of the node identifiers in the seepage path dataset.

[0076] The number of rows in the matrix equals the total number of critical abrupt change pore nodes identified in the seepage path dataset; the number of columns in the matrix equals the number of the four geometric variation features.

[0077] Based on the node sequence list and edge weight list recorded in the seepage path dataset, the identifier of each pair of adjacent nodes and its corresponding edge weight are extracted. The adjacent node pair information is transformed into an edge feature vector. The edge feature vector consists of two parts: the first part is the row index number of the adjacent node in the node feature matrix, and the second part is the connectivity strength weight value of the edge. All edge feature vectors are filled into the edge feature matrix row by row in the order recorded in the seepage path dataset.

[0078] The number of rows in the matrix equals the total number of adjacent node pairs in the percolation path dataset; the matrix has a fixed three columns, storing the starting node index, the target node index, and the edge weight value, respectively. The matrix data type uses a mixed floating-point and integer mode: the first two columns are integer node indices, and the third column is a floating-point connectivity strength weight value.

[0079] The node feature matrix and edge feature matrix are input into the graph convolutional layer of the pre-trained graph neural network model, and node information fusion is achieved through layer-by-layer neighborhood aggregation operation. Graph Neural Network Model Architecture and Pre-training Methods: The graph neural network model consists of three layers of graph convolutional units. Each graph convolutional unit is followed by an activation function and a dropout rate operation, and then skip connections are used to retain the output features of the previous layers. The input and output dimensions of each graph convolutional unit are configured as follows: the first layer has an input dimension equal to the node feature dimension (i.e., four dimensions) and an output dimension of sixty-four; the second layer has an input dimension of sixty-four and an output dimension of thirty-two; the third layer has an input dimension of thirty-two and an output dimension of sixteen.

[0080] The activation function is a linear rectified unit function. Each activation function is applied to the weighted aggregation result of the graph convolutional unit output to introduce non-linear feature representation capabilities.

[0081] The dropout rate operation uses a random zeroing method to control the risk of overfitting, with the dropout rate value set to 0.1.

[0082] Skip connections refer to the element-wise addition of the input features and the output features of each graph convolutional unit to preserve shallow feature information.

[0083] Within each graph convolutional unit, a neighborhood traversal is performed sequentially for each node identifier in the node feature matrix. The neighborhood traversal retrieves all target node identifiers connected to the current node identifier based on the node pair relationships and edge weights recorded in the edge feature matrix. For each connected pair, the corresponding edge weight value is read from the edge feature matrix.

[0084] The neighborhood aggregation operation first performs a weighted average of the feature vectors of all neighboring nodes in the upper-level representation space according to the edge weights. Specifically, it performs a pairwise operation on the list of neighboring node identifiers and the list of weights, multiplies the feature vector of each neighboring node by the corresponding edge weight value, sums them up, and then divides the sum by the number of neighboring nodes.

[0085] The weighted average result is superimposed and fused with the feature vector of the current node's identifier in the previous layer's representation space. The fusion rule is: the weighted value after superposition equals the sum of the neighborhood weighted average result and the current node's feature vector. The superposition result is then passed to the linear mapping operation of the current layer. The linear mapping operation uses a trainable weight matrix and a bias vector to map the fused vector to the next layer's feature dimension. The mapping formula is: the linear mapping result equals the product of the fused vector and the weight matrix plus the bias vector.

[0086] The linear mapping result is processed by an activation function to obtain the output feature vector of the current layer. The output feature vector of the current layer is arranged in rows to form a new node feature matrix, which is used by the graph convolutional units of the next layer.

[0087] After linear mapping and activation function, a random dropout operation is performed on the output feature vector, randomly setting some dimensions to zero according to a set dropout rate. After the dropout operation, the initial input feature vectors of the nodes are element-wise added to the current layer's dropped output feature vector to form a skip connection result.

[0088] The skip connection result serves as the final output feature vector of the current layer. Combining the output feature vectors of all node identifiers forms a new node feature matrix, which is then used as the input to the graph convolutional unit of the next layer.

[0089] The pre-trained graphical neural network model has been trained on an independent soil experimental dataset, which includes an artificially constructed seepage path network and its corresponding risk label values. During training, the mean squared error function is used as the loss function, and the network weight parameters are iteratively updated using the backpropagation algorithm and stochastic gradient descent optimizer until the mean squared error of the validation set converges or the number of iterations reaches a preset upper limit.

[0090] Through layer-by-layer neighborhood aggregation operations, the features of neighboring nodes are weighted and fused using the edge weights in the edge feature matrix. Combined with activation functions and skip connection strategies, this achieves deep representation and information fusion of node features. The pre-trained model parameters ensure the network's ability to discriminate microscopic abrupt changes in the seepage path.

[0091] The deep feature representations output by the graph neural network model are sequentially mapped through fully connected layers and regression layers to obtain the microscopic pore structure mutation risk value corresponding to each critical mutation pore node; The deep feature representation of each node is derived from the node feature matrix output by the third-layer graph convolutional unit. The deep feature representation has a dimension of sixteen. To extract non-linear features, two fully connected sub-networks are set up.

[0092] The first fully connected subnetwork has a 16-dimensional input and an 8-dimensional output. This subnetwork linearly maps the input depth feature vector before connecting it to a rectified linear unit (CLU) for non-linear transformation. The linear mapping operation involves multiplying the depth feature vector by the weight matrix of the first fully connected layer and adding the bias vector. The output is an 8-dimensional vector, which is then processed by the activation function.

[0093] The second fully connected sub-network has an input dimension of eight and an output dimension of one. This layer performs a linear mapping operation on the eight-dimensional vector output from the first layer, using the result as the input to the regression layer. This layer no longer uses a non-linear activation function, directly outputting a scalar value. This scalar value is then mapped by the regression layer to a risk value for abrupt changes in the microporous structure.

[0094] The regression layer performs linear scaling and offset operations on the one-dimensional values ​​output by the second fully connected sub-network, mapping the original values ​​to a risk value range between zero and one. The linear scaling parameters include a scaling factor and an offset, which are obtained during the pre-training phase through data normalization to ensure that the closer the risk value of the microporous structure mutation is to one after mapping, the higher the risk, and the closer it is to zero, the lower the risk.

[0095] After mapping all node identifiers, the risk values ​​of microscopic pore structure mutations are stored row by row in a list of microscopic pore structure mutation risk values. The row order of this list is consistent with the original node feature matrix. The output of the list of microscopic pore structure mutation risk values ​​is a matrix, and the columns are called the microscopic pore structure mutation risk value vector.

[0096] Based on the abrupt change risk value of micropore structure and soil consolidation environment parameters, a multimodal attention mechanism is used to adjust the prediction weights to generate permeability coefficient transition prediction values, including: The microscopic pore structure mutation risk value is matched with the soil solidification environmental parameters of the corresponding solidification stage in time to construct a joint feature matrix of microscopic risk-environmental parameters. Soil solidification environmental parameters include, but are not limited to, temperature, moisture content, external stress level, solidification age, and drainage boundary type. The time synchronization matching method is as follows: based on the time nodes of the solidification stage, the microscopic pore structure mutation risk values ​​obtained at the same time step are sequentially combined with the solidification environmental parameters to form a risk parameter alignment sample.

[0097] Batch stitching is performed on the risk parameter aligned samples across all time steps to obtain a joint feature matrix composed of the micropore structure mutation risk value and curing environment parameters. Each row of the matrix represents a feature sample at a time step, and each column represents a specific index dimension. The column order of the matrix is ​​fixed, and the corresponding indices are: micropore structure mutation risk value, temperature, moisture content, stress level, curing age, and drainage boundary type.

[0098] The joint feature matrix of micro-risk-environment parameters is input into a deep prediction network containing a multimodal attention mechanism, and feature transformation and attention weighting are performed in each modal branch respectively; The joint feature matrix is ​​input into a deep prediction network incorporating a multimodal attention mechanism. The deep prediction network consists of three modal branches, handling microstructure, environmental physics, and temporal evolution modes, respectively. Each modal branch contains two feature transformation modules, each consisting of a linear mapping layer and a nonlinear activation function. The feature transformation modules are used to normalize the input feature dimensions.

[0099] After feature transformation, an attention weighting module is introduced into each modality branch. This module calculates the importance weight of each dimension in the current prediction task based on the attention score function, and multiplies this importance weight by the corresponding feature value to obtain a weighted feature representation. The weighting function is a weighted inner product of the input features and the learning parameters, and the result is normalized to ensure that the sum of all weight values ​​is one.

[0100] By fusing the attention weighting results of each modal branch, a fused prediction feature vector covering different solidification stages and different connectivity structure types is generated; The weighted feature representations output from the three modal branches are concatenated to generate a fused prediction feature vector. The fused prediction feature vector contains weighted information on microstructural change trends, solidified environment change characteristics, and historical evolution trends, and the vector dimension is the sum of the concatenated feature dimensions.

[0101] The fused prediction feature vector is input into the regression subnetwork, and the predicted value of the penetration coefficient transition at the corresponding time step is output. The fused predicted feature vector is input into the regression subnetwork. The regression subnetwork consists of a fully connected mapping layer and an output regression layer. The fully connected mapping layer maps the fused predicted feature vector to an intermediate prediction space, and the regression layer performs numerical regression, outputting permeability coefficient transition prediction values ​​corresponding one-to-one with each time step. Each permeability coefficient transition prediction value is a real scalar, representing the transition trend of the soil permeability coefficient at the current time. The permeability coefficient transition prediction values ​​are summarized into a numerical vector, which serves as the predicted output of permeability evolution and is used for drainage scheme optimization.

[0102] The drainage scheme is optimized based on the predicted values ​​of permeability coefficient transitions, including: An objective function for optimizing drainage schemes is constructed using the predicted values ​​of permeability coefficient transitions at different time steps as input conditions. The predicted permeability coefficient transition values ​​in the permeability coefficient transition vector are arranged in chronological order, corresponding to the permeability evolution trends in the initial, middle, and late solidification stages, respectively. The objective function for optimizing the drainage scheme must simultaneously consider reducing the permeability coefficient and minimizing drainage costs. The objective function is defined as the weighted sum of the cumulative deviations between the predicted permeability coefficient values ​​and the design threshold across all time steps, and the cumulative values ​​of the drainage energy consumption index. The construction process is as follows: The predicted permeability coefficient transition value at each time step is compared with the pre-set design permeability coefficient benchmark value. If the predicted permeability coefficient transition value is higher than the design permeability coefficient benchmark value, it is considered a deviation value. The deviation value is expressed as the absolute value of the difference between the predicted permeability coefficient transition value and the design permeability coefficient benchmark value; if the predicted permeability coefficient transition value is lower than the design permeability coefficient benchmark value, the deviation value is zero.

[0103] The cumulative permeability deviation value is obtained by summing the deviation values ​​at all time steps. The cumulative permeability deviation value reflects the weighted sum of the degree of permeability exceeding the standard in each time period under the current drainage scheme.

[0104] When calculating drainage energy consumption, the product of drainage hydraulic power and drainage duration is chosen as an approximation of energy consumption. Drainage hydraulic power is calculated based on drainage rate, drainage head, and drainage equipment efficiency. The product of drainage rate and head is then divided by the equipment efficiency to obtain the energy consumption per unit time. Finally, the energy consumption at each time step is multiplied by the duration of that time step to obtain the energy consumption for that time step.

[0105] The cumulative drainage energy consumption value is obtained by summing up the energy consumption at all time steps.

[0106] The cumulative infiltration deviation value and the cumulative drainage energy consumption value are weighted and summed according to a preset ratio to form the objective function for optimizing the drainage scheme. The preset ratio value is determined by the engineering designer based on the on-site economic and safety requirements. For example, the weight of infiltration deviation can be set to be greater than the weight of drainage energy consumption to ensure that infiltration exceeding the standard will be penalized first. The objective function is expressed as: the objective function value for optimizing the drainage scheme equals the cumulative infiltration deviation value multiplied by the infiltration deviation weight plus the cumulative drainage energy consumption value multiplied by the energy consumption weight.

[0107] Optimize the objective function of the drainage scheme and set optimization constraints; The optimization constraints include the stage characteristics of the soil solidification process, the risk value of abrupt changes in the micropore structure, the relationship between the predicted value of the permeability coefficient transition and the corresponding time step.

[0108] After the objective function is constructed, multiple sets of constraints need to be applied to the drainage scheme optimization process to ensure that the optimization results are within the feasible engineering range. The optimization constraints include: During the initial curing, intermediate curing, and final curing stages, the drainage rate and duration are limited. Specifically, the drainage rate during the initial curing stage must not exceed the set maximum drainage rate threshold to prevent excessive initial drainage from disturbing the foundation; the drainage rate during the intermediate curing stage should be dynamically adjusted according to the curing agent's reaction intensity to ensure that the drainage rate is between the minimum and maximum rates during the intermediate stage; and the drainage rate during the final curing stage should be lower than the minimum rate during the intermediate stage to ensure sufficient settling of fine particles.

[0109] At each time step, corresponding to the micropore structure mutation risk value in the micropore structure mutation risk value vector, it is necessary to ensure that a minimum drainage coverage distance is set in the high-risk node area so as to concentrate more drainage resources in the high-risk area.

[0110] During the intermediate and late curing stages, the rate of decrease in the predicted permeability coefficient must be greater than or equal to the minimum rate threshold to ensure effective curing. The minimum rate threshold can be determined based on experience from similar projects or experimental data.

[0111] The optimization algorithm is used to optimize the objective function of the drainage scheme, and the optimized drainage scheme that satisfies the optimization constraints is obtained. Choose an algorithm suitable for multi-objective constrained optimization to solve the objective function. For example, a genetic algorithm combined with a penalty function strategy can be used for global search, specifically: Genetic algorithm population initialization: Several candidate drainage schemes are generated within the workable area according to rules. The encoding of each candidate drainage scheme includes a drainage well coordinate layout scheme, a drainage rate distribution scheme, and a drainage duration scheme. The drainage well coordinates are generated using a grid system, with the workable grid range determined by on-site measurement data. Drainage wells can be placed at grid intersections. The drainage rate and duration encoding use finite discrete values; for example, the rate can be selected from three levels (low, medium, and high), and the duration can be selected from several fixed time intervals.

[0112] Fitness assessment and penalty function design: Each candidate drainage scheme is decoded into specific drainage scheme parameters. The cumulative infiltration deviation and cumulative drainage energy consumption are calculated based on the objective function, and the objective function value is calculated according to preset weights. For individuals that violate constraints, a penalty function is used to penalize their objective function value. The penalty function is the objective function value plus the penalty term for each constraint violation multiplied by a penalty coefficient, where the penalty term value is the excess amount of the candidate drainage scheme in the corresponding constraint dimension.

[0113] Genetic operator settings: Several superior individuals are selected from the population, and a new population is generated using crossover and mutation operators. The crossover operator exchanges drainage coordinates and rate time periods between two parent schemes, and can use uniform crossover or partially matched crossover. The mutation operator randomly fine-tunes or randomly switches to adjacent selectable values ​​for a specific gene locus (such as the location of a drainage well or the rate during a certain time period) of a single individual. The crossover probability and mutation probability can be set based on engineering experience. The crossover probability is set to a medium level to ensure diversity, and the mutation probability is set to a low level to maintain the transmission of superior genes.

[0114] Iteration Termination Criteria: After several generations of genetic iterations, when the fluctuation range of the optimal value of the objective function is less than a preset threshold or the maximum number of iterations is reached over several consecutive generations, the algorithm is considered to have converged and the iteration stops. The preset maximum number of iterations can be determined based on the complexity of the drainage scheme and the expected computation time, for example, set to one hundred generations.

[0115] Output the optimal individual: Select the individual with the smallest objective function value that satisfies all constraints from the final population, and decode it into specific drainage scheme parameters, including the coordinate layout of drainage wells, the drainage rate level of each drainage well, and the configuration of drainage duration. The optimal individual is the optimized drainage scheme.

[0116] After completing the optimization solution, decoding the optimal individual yields an optimized drainage scheme that follows the information below: Drainage layout location: Based on the coordinate information of the drainage wells in the optimal individual system, the geographical coordinates of each drainage well are marked on the site construction drawing. The coordinate marking is based on the engineering survey coordinate system, using vertical and horizontal coordinate values ​​to accurately describe the location of the drainage wells.

[0117] Drainage rate: Based on the rate level information of each drainage well in the optimal individual, a specific flow rate value is assigned to each drainage well.

[0118] Drainage duration: Convert the duration information of each drainage well in each solidification stage in the optimal individual into time periods.

[0119] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0120] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0121] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0122] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0123] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0124] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0125] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0126] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0127] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0128] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the treatment effect of silty soil based on deep learning, characterized in that, Includes the following steps: S1: Scan the target silty soil sample to obtain microscopic images of the soil at different solidification stages, and generate a spatial structure voxel model containing pore distribution through a three-dimensional image reconstruction algorithm. S2: Extract pore topology parameters based on spatial structure voxel model, apply maximum sphere algorithm to identify pore channel connectivity structure, and construct pore multi-level connectivity feature tensor; S3: Perform seepage path search analysis on the multi-level connectivity feature tensor of pores, locate critically abrupt ... S4: Input the pore geometric variation feature quantity and the seepage path dataset into the pre-trained graph neural network model, and output the mutation risk value of the micropore structure; S5: Based on the mutation risk value of micropore structure and soil solidification environment parameters, the prediction weight is adjusted by using a multimodal attention mechanism to generate permeability coefficient transition prediction value; S6: Optimize the drainage scheme based on the predicted value of the permeability coefficient transition.

2. The method for predicting the treatment effect of silty soil based on deep learning according to claim 1, characterized in that, S1, specifically: Multi-angle tomographic scanning was performed on the target silty soil sample, and microscopic slice images of the soil were acquired at the initial solidification stage, the intermediate solidification stage and the final solidification stage. Spatial registration is performed on the microscopic slice images acquired at each curing stage to generate a three-dimensional grayscale matrix; Noise suppression is applied to the three-dimensional grayscale matrix to remove scanning artifacts and preserve the true pore boundaries; The denoised 3D grayscale matrix is ​​divided into a solid phase voxel set and a pore voxel set; A 3D image reconstruction algorithm is executed based on the solid voxel set and the pore voxel set to generate a spatial structure voxel model containing the pore distribution.

3. The method for predicting the treatment effect of silty soil based on deep learning according to claim 2, characterized in that, S2, specifically: Within the spatial structure voxel model, the pore voxel set is labeled with connected components to identify each independent connected sub-region; The maximum inscribed sphere algorithm is applied to each independent connected subregion to locate the key connected sphere and calculate the corresponding radius. Extract the shape factor, connected path length, and aperture distribution index for each key connected sphere; The shape factor, connected path length, critical connected sphere radius, and aperture distribution index are combined into a multi-level connectivity feature tensor based on the porosity connectivity hierarchy.

4. The method for predicting the treatment effect of silty soil based on deep learning according to claim 3, characterized in that, S3, specifically: A seepage network graph is constructed based on the multi-level connectivity feature tensor of pores. The voxels in each independent connected sub-region are identified as nodes of the network graph, and the connectivity between nodes is identified as edges of the network graph. A graph traversal algorithm is performed on the seepage network diagram to identify critical abrupt change pore nodes in the seepage network diagram using a depth-first traversal approach. Data is recorded on critically abruptly changed pore nodes and adjacent connection information to generate a seepage path dataset. Extract the pore geometric variation feature corresponding to each critical mutation pore node from the seepage path dataset.

5. The method for predicting the treatment effect of silty soil based on deep learning according to claim 4, characterized in that, S4, specifically: Construct a node feature matrix and an edge feature matrix based on each critical mutation pore node and its corresponding pore geometric variation feature in the seepage path dataset. The node feature matrix and edge feature matrix are input into the graph convolutional layer of the pre-trained graph neural network model, and node information fusion is achieved through layer-by-layer neighborhood aggregation operation. The deep feature representations output by the graph neural network model are sequentially mapped through fully connected layers and regression layers to obtain the microscopic pore structure mutation risk value corresponding to each critical mutation pore node.

6. The method for predicting the treatment effect of silty soil based on deep learning according to claim 5, characterized in that, S5, specifically: The microscopic pore structure mutation risk value is matched with the soil solidification environmental parameters of the corresponding solidification stage in time to construct a joint feature matrix of microscopic risk-environmental parameters. The joint feature matrix of micro-risk-environment parameters is input into a deep prediction network containing a multimodal attention mechanism, and feature transformation and attention weighting are performed in each modal branch respectively; By fusing the attention weighting results of each modal branch, a fused prediction feature vector covering different solidification stages and different connectivity structure types is generated; The fused prediction feature vector is input into the regression subnetwork, and the predicted value of the penetration coefficient transition at the corresponding time step is output.

7. The method for predicting the treatment effect of silty soil based on deep learning according to claim 6, characterized in that, S6, specifically: An objective function for optimizing drainage schemes is constructed using the predicted values ​​of permeability coefficient transitions at different time steps as input conditions. The objective function for optimizing the drainage scheme is set, and optimization constraints are set. The optimization constraints include the stage characteristics of the soil solidification process, the risk value of abrupt changes in the micropore structure, the relationship between the predicted value of the permeability coefficient transition and the corresponding time step. The optimization algorithm is used to optimize the objective function of the drainage scheme to obtain an optimized drainage scheme that meets the optimization constraints. The optimized drainage scheme includes the drainage layout location, drainage rate, and drainage duration.

Images