A complex sensitive environment subway construction management method and device

Abstract

The application provides a complex sensitive environment subway construction management method and device, relates to the technical field of subway construction management, and comprises the following steps: acquiring first information, wherein the first information comprises shield propulsion parameter time sequence data of a construction area, building settlement monitoring data, underground pipeline distribution data and microseismic sensor array data; preprocessing the first information to obtain a three-dimensional geological influence graph integrating underground structure-stress response-dynamic disturbance; jointly optimizing the three-dimensional geological influence graph through semi-variogram modeling and spectral graph co-training clustering algorithm to generate a multi-layer heterogeneous graph structure; cooperatively evolving the multi-layer heterogeneous graph structure through a dynamic graph convolution network and a graph integral propagation model to generate a risk cascade prediction model; and solving the prediction model to obtain a global management scheme of subway construction. The application realizes risk controllability, timely response and dynamic adjustment of a subway construction whole life cycle management strategy.

Description

Technical Field

[0001] This invention relates to the field of subway construction management technology, and more specifically, to a method and apparatus for subway construction management in complex and sensitive environments. Background Technology

[0002] Subway construction in complex and sensitive environments, such as near high-rise buildings, areas with dense underground pipelines, or areas with weak strata, often faces challenges such as high uncertainty in geological structure, complex stress disturbance propagation, and difficulty in accurately predicting the impact range of construction disturbances. Current construction management technologies largely rely on empirical models or rule bases driven by single parameters, such as making empirical judgments by monitoring building settlement or shield tunneling parameter fluctuations, and then combining them with two-dimensional geological profiles for risk estimation. These methods lack the ability to fuse multi-source heterogeneous information and cannot comprehensively characterize the coupling relationship between underground structures and dynamic disturbances. In addition, traditional technologies usually lack the ability to dynamically predict the cascading effects of risks, leading to delayed emergency response and inaccurate judgment of risk propagation paths. Especially in extreme situations such as earthquake-induced disturbances or multi-source disturbance coupling, structural damage and safety accidents are more likely to occur, making it difficult to support refined construction control under multi-objective conditions.

[0003] Therefore, there is an urgent need for a method and device for subway construction management in complex and sensitive environments to solve the above problems. Summary of the Invention

[0004] The purpose of this invention is to provide a method and apparatus for subway construction management in complex and sensitive environments, so as to improve the above-mentioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0005] Firstly, this application provides a method for subway construction management in complex and sensitive environments, including:

[0006] Obtain first information, which includes time-series data of shield tunneling parameters in the construction area, building settlement monitoring data, underground pipeline distribution data, and microseismic sensor array data;

[0007] The first information is preprocessed to obtain a three-dimensional geological impact map that integrates underground structure, stress response, and dynamic disturbance.

[0008] The three-dimensional geological impact map is jointly optimized by semi-variogram modeling and spectrum co-training clustering algorithm to generate a multi-layer heterogeneous map structure.

[0009] The multi-layer heterogeneous graph structure is co-evolved with a dynamic graph convolutional network and a graph integral propagation model to generate a risk cascade prediction model.

[0010] The prediction model is solved, and multi-objective feasible region mapping is performed based on the solution results to obtain a global management scheme for subway construction.

[0011] Secondly, this application also provides a subway construction management device for complex and sensitive environments, comprising:

[0012] The acquisition unit is used to acquire first information, which includes time-series data of shield tunneling parameters in the construction area, building settlement monitoring data, underground pipeline distribution data, and microseismic sensor array data.

[0013] The processing unit is used to preprocess the first information to obtain a three-dimensional geological impact map that integrates underground structure, stress response, and dynamic disturbance.

[0014] The optimization unit is used to jointly optimize the three-dimensional geological impact map by performing semi-variogram modeling and spectrum co-training clustering algorithm to generate a multi-layer heterogeneous map structure.

[0015] An evolution unit is used to co-evolve the multi-layer heterogeneous graph structure with a dynamic graph convolutional network and a graph integral propagation model to generate a risk cascade prediction model.

[0016] The computing unit is used to solve the prediction model and perform multi-objective feasible domain mapping based on the solution results to obtain a global management scheme for subway construction.

[0017] The beneficial effects of this invention are as follows:

[0018] This invention constructs an integrated technical framework encompassing multi-source spatiotemporal data acquisition, 3D geological map construction, multi-layer graph structure modeling, risk prediction, and global construction control. This method achieves, for the first time, 3D modeling integrating underground structure, stress response, and dynamic disturbance mechanisms. Advanced algorithms such as multi-core tensor support machines, coherence point drift algorithms, and spectral co-training are used to obtain graph structure representations with higher spatial consistency and physical correlation. Furthermore, through the co-evolution of dynamic graph convolutional networks and graph integral propagation models, a risk cascade prediction model capable of characterizing vulnerable nodes, crack propagation paths, and propagation thresholds is constructed. Finally, based on fractional-order differential-curvature driven stochastic modeling methods, combined with multi-objective Pareto optimization and variable structure network solution mechanisms, this invention outputs construction parameter envelopes and a 3D management cube, achieving risk-controllable, timely response, and dynamically adjustable strategy management throughout the entire lifecycle of subway construction, significantly improving the safety and intelligence level of construction in complex environments.

[0019] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the process for managing subway construction in complex and sensitive environments as described in this embodiment of the invention;

[0022] Figure 2 This is a schematic diagram of the complex and sensitive environment subway construction management device described in an embodiment of the present invention.

[0023] In the diagram: 701, Acquisition Unit; 702, Processing Unit; 703, Optimization Unit; 704, Evolution Unit; 705, Calculation Unit. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0025] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] Example 1:

[0027] This embodiment provides a method for subway construction management in complex and sensitive environments.

[0028] See Figure 1 The figure shows that the method includes steps S1, S2, S3, S4, and S5.

[0029] Step S1: Obtain first information, which includes time-series data of shield tunneling parameters in the construction area, building settlement monitoring data, underground pipeline distribution data, and microseismic sensor array data.

[0030] Understandably, in this step, the time-series data of shield tunneling parameters typically include dynamic variables such as thrust, torque, tunneling speed, and grouting pressure. These parameters reflect the mechanical response characteristics of the interaction between the shield machine and the stratum, and are important bases for judging the degree of construction disturbance and the stratum response. Secondly, building settlement monitoring data needs to be continuously observed by deploying high-precision levels, laser scanners, or GNSS base stations to monitor the settlement behavior of the ground and existing structures. This is crucial for assessing the impact of construction on surrounding sensitive buildings. Thirdly, underground pipeline distribution data comes from diverse sources, including municipal databases, underground radar detection (GPR), and BIM model information fusion, aiming to establish an accurate distribution map of underground target structures and ensure the control of construction safety distances. Finally, microseismic sensor array data uses a high-density distributed triaxial seismograph array to capture minute geological disturbances caused by shield construction, and analyzes the evolution of micro-fractures within the stratum through P-wave and S-wave propagation characteristics. These data exhibit high heterogeneity in terms of time, space, and physical properties. Therefore, a synchronized clock system, standardized interfaces, and data fault tolerance mechanisms must be introduced during the data acquisition process to ensure comparability and consistency in subsequent modeling. Through this integration process, the coordinated acquisition of construction disturbance information and geological response information was achieved, laying a solid foundation for establishing a comprehensive analysis model that integrates dynamic disturbance, underground structure, and stress response.

[0031] Step S2: Preprocess the first information to obtain a three-dimensional geological impact map that integrates underground structure, stress response, and dynamic disturbance;

[0032] Understandably, this step achieves deep feature fusion and precise spatial registration between data from different physical scales and sources, establishing a dynamic energy channel map from construction disturbance to geological response. This provides accurate, continuous, and structured basic data support for subsequent semi-variogram modeling and heterogeneous map construction, significantly improving the understanding and modeling capabilities of complex underground environment disturbance mechanisms. In this step, step S2 includes steps S21 and S22.

[0033] Step S21: Perform empirical mode decomposition, nonnegative tensor decomposition and generalized regression neural network co-processing on the first information to generate a spatiotemporally aligned structured multi-source dataset;

[0034] It is understandable that this step, by introducing three algorithms for synergistic fusion, achieves multimodal unification and spatiotemporal alignment of shield tunneling disturbance, building response, and underground structural characteristic data. This lays a highly structured data foundation for subsequent map construction and heterogeneous graph neural network modeling, effectively improving the model's accurate representation and spatiotemporal analysis capabilities of the dynamic evolution process. In this step, step S21 includes steps S211, S212, S213, S214, and S215.

[0035] Step S211: Perform empirical mode decomposition processing on the time series data of the shield propulsion parameters to obtain a set of intrinsic mode functions including the main frequency component of the cutterhead torque;

[0036] Understandably, this step first identifies the local maxima and minima of the original torque signal, constructing upper and lower envelopes; then, it obtains the local trend of the signal through local averaging, gradually subtracting to obtain the first set of intrinsic mode functions (IMFs); this process is iterated until the residuals are monotonic functions. The resulting IMF components correspond to the dynamic behavior at different time scales. For shield tunneling torque data, the first few IMFs typically contain high-frequency disturbance components of the mechanical system (such as cutterhead vibration), the middle-order components may correspond to the mid-frequency response caused by soil shear failure or abrupt changes in rock strata, while the later orders may reflect the low-frequency energy accumulation process of the system.

[0037] To extract components strongly correlated with geological disturbances, this step further introduces dominant frequency analysis and kurtosis calculation to screen the core frequency components in the intrinsic mode functions (IMFs). The cutterhead torque dominant frequency reflects the distribution of strata strength and changes in cutting resistance ahead of the tunnel boring machine (TBM). The resulting IMF set not only contains dominant frequency information that is physically representative of construction disturbances but also possesses good temporal continuity and signal-to-noise ratio.

[0038] Step S212: Based on the set of intrinsic mode functions, perform kurtosis-sample entropy joint criterion screening to obtain the denoised shield tunneling parameter dynamic component matrix;

[0039] Understandably, this step first uses kurtosis as a statistical measure of the degree of signal abrupt change, which can identify sharp, discrete, and anomalous fluctuations in the signal. These components are often related to events such as local ground instability and shield tunneling head encountering obstacles, and have high engineering indicative value. In the calculation, by sorting the kurtosis values ​​of each IMF component, components with strong non-Gaussian characteristics can be initially screened out. Secondly, sample entropy, as a nonlinear indicator measuring the complexity and irregularity of a time series, can reflect the systematic change trend and predictability of the signal. In the shield tunneling environment, sudden events usually significantly reduce sample entropy, making the signal more regular and abrupt.

[0040] Therefore, combining kurtosis and sample entropy allows for dual screening of intrinsic mode functions (IMFs) from two complementary perspectives (signal sharpness and complexity): on the one hand, it retains components with high kurtosis that characterize instantaneous disturbances; on the other hand, it filters out redundant components with extremely low entropy values ​​that may represent mechanical periodic vibrations or background noise. Finally, by setting a joint threshold or employing a weighted sorting mechanism, the most representative dynamic IMF components are selected, and a denoised dynamic component matrix of tunnel boring machine parameters is reconstructed, where each row represents disturbance components at different time scales, and each column represents the temporal evolution process.

[0041] Step S213: Perform tensor modeling based on the building settlement monitoring data and underground pipeline distribution data. After third-order non-negative tensor modeling and tensor chain decomposition, obtain the multi-field coupling tensor of the repaired underground structure.

[0042] Understandably, this step first constructs a three-dimensional tensor structure from building settlement monitoring data according to "building number - time - settlement value," while simultaneously mapping underground pipeline distribution data into a tensor expression of "pipeline type - spatial location - state parameters." After standardization and alignment in the spatial coordinate system, the two are merged into a unified third-order tensor, with the three dimensions corresponding to structural units, temporal distribution, and physical state parameters, respectively. To improve the physical interpretability of the data, non-negative tensor modeling is introduced, meaning all tensor elements are only allowed to be non-negative, to conform to the realistic constraints of natural phenomena such as surface settlement and pipeline pressure.

[0043] Next, the third-order tensor is structurally reconstructed and missing values ​​are filled in through tensor chain decomposition. The optimization objective during the decomposition process is to minimize the reconstruction error. Alternating least squares or variational inference methods are used to iteratively optimize each chain node tensor, enabling the model to efficiently fill in outlier or missing data points while preserving the original structural information.

[0044] The final output is the multi-field coupling tensor of the repaired underground structure. This tensor fully expresses the three-dimensional coupling relationship between settlement evolution, structural state and pipeline distribution in urban underground space, laying the foundation for subsequent neural network modeling and three-dimensional geological map construction.

[0045] Step S214: Perform spatiotemporal mapping processing of the generalized regression neural network based on the dynamic component matrix and the multi-field coupling tensor to obtain a fusion feature tensor with spatiotemporal labels;

[0046] It is understandable that in this step, the basic structure of the generalized regression neural network includes an input layer, a pattern layer, a summation layer, and an output layer, and does not rely on complex training mechanisms such as gradient descent. The specific process is as follows: First, the dynamic component matrix (reflecting dynamic characteristics such as shield tunneling disturbances and torque fluctuations) is used as the input variable, and after tensorization, it is expanded into a multi-dimensional vector sequence over time steps; then, the third-order multi-field coupling tensor (reflecting settlement-pipeline state) is expanded in the same time coordinate manner as the target mapping output.

[0047] In the pattern layer, GRNN uses a Gaussian kernel function to calculate the similarity between the input sample and existing samples using Euclidean distance, generating response weights. Then, a weighted summation layer is used to weight and sum the responses of all samples to form the output, avoiding the gradient vanishing and overfitting problems common in deep networks. In this way, GRNN effectively captures the nonlinear functional relationship between shield tunneling disturbance and the response of the underground structure, and naturally embeds timestamps and spatial location indices into the network structure, ultimately generating a fused feature tensor with unified spatial encoding (such as a geographic raster index) and time-series labels.

[0048] Step S215: Perform modal correction processing based on the fused feature tensor and conjugate gradient optimization processing with partial coherence function constraints to obtain a spatiotemporally strictly aligned structured multi-source dataset.

[0049] Understandably, the modal correction step in this process utilizes the partial coherence function (PCF) as a physical constraint indicator to quantify the true causal relationships between different data modes. The PCF is used to eliminate indirect influences between multiple variables, retaining only the net coherent component between two signals, thereby identifying frequency bands or time periods where actual dynamic coupling exists between shield disturbance and underground response. This process is equivalent to a statistical filter, removing components with low signal-to-noise ratios or no physical correlation from the tensor.

[0050] Subsequently, a conjugate gradient optimization algorithm is introduced to correct and relocate the feature paths in the fused tensor under PCF constraints. The conjugate gradient algorithm is an efficient iterative optimization method that does not require explicit computation of massive Hessian matrices, making it particularly suitable for sparse correction scenarios in large-scale tensor fields. Its optimization process iteratively approximates the optimal path direction, combining the conjugate relationship between the gradient direction and the previous step direction to quickly converge to the mode registration solution with minimum loss.

[0051] Ultimately, this process not only achieved precise synchronization between spatial coordinates and timestamps but also corrected physical deviations in the tensor caused by modal mismatch, resulting in a spatiotemporally strictly aligned structured multi-source dataset. This dataset possesses the following characteristics: ① Physical consistency among data modalities; ② Each data point accurately corresponds to a specific time period and geographical coordinates; ③ High fidelity and low redundancy, making it suitable for subsequent use in high-dimensional geological model construction or risk identification.

[0052] Step S22: The spatiotemporally aligned structured multi-source dataset is coupled with a multi-core tensor support machine and a coherent point drift algorithm to construct a three-dimensional geological influence map that integrates underground structure, stress response, and dynamic disturbance.

[0053] It is understandable that this step achieves deep feature fusion and precise spatial registration between data from different physical scales and sources, and establishes a dynamic energy channel map from construction disturbance to geological response. This provides accurate, continuous and structured basic data support for subsequent semivariogram modeling and heterogeneous map construction, and greatly improves the ability to understand and model the disturbance mechanism of complex underground environment. In this step, step S22 includes steps S221, S222, S223, S224 and S225.

[0054] Step S221: Perform multi-kernel tensor support machine processing on the structured multi-source dataset, and extract geological static features, stress dynamic features and shield tunneling disturbance time series features through Gaussian kernel, polynomial kernel and linear kernel function group respectively, and generate cross-scale feature weight matrix;

[0055] Understandably, this step first uses a structured multi-source dataset as a high-dimensional tensor input. This dataset integrates various heterogeneous signals, including building settlement information, tunnel boring machine (TBM) propulsion disturbances, and stress temporal responses. To characterize the static geological field (such as formation hardness and distribution), the dynamic evolution of stress, and the temporal features of TBM disturbances, a multi-kernel learning strategy is employed to construct a set of kernel functions:

[0056] The Gaussian kernel function is designed to capture the local continuity and nonlinear variation patterns in geological structures, making it suitable for modeling the distribution of static subsurface features.

[0057] Polynomial kernel functions are used to simulate the multi-order coupling relationship of stress response over time, and are particularly suitable for analyzing complex dynamic evolution paths under engineering disturbances;

[0058] Linear kernel functions are used to preserve first-order trend information in shield tunneling disturbances, such as linear fluctuation patterns of propulsion speed and cutterhead torque, which helps maintain the interpretability and response sensitivity of the model.

[0059] The aforementioned set of kernel functions works together on the tensor data to construct a multi-scale nested feature space. Through the joint optimization mechanism of multi-kernel support vector tensors, the system can automatically learn the importance of each kernel function to the target feature dimension and output a cross-scale feature weight matrix.

[0060] Step S222: Perform kernel target alignment processing based on the weight matrix to obtain a low-rank tensor feature basis;

[0061] Kernel-target alignment is an optimization method that measures the similarity and information consistency among the outputs of multiple kernel functions. Its core idea is to achieve alignment and fusion of different kernel mapping spaces on the target task by maximizing the Frobenius inner product between the kernel matrix and the target kernel. This step achieves semantic fusion and consistency of multi-kernel feature spaces through the kernel-target alignment mechanism, eliminating the kernel bias problem caused by heterogeneous sources in geological disturbance modeling. At the same time, the introduction of tensor low-rank modeling strategy significantly improves the compactness and interpretability of feature representation, providing stable and refined feature support for subsequent 3D map construction.

[0062] Step S223: Based on the feature base, perform anisotropic coherent point drift processing to non-rigidly align the geological anomaly with the building foundation point cloud to obtain a spatially consistent fused point cloud;

[0063] Understandably, this step first treats the source point cloud (e.g., geological anomaly point cloud) as a sample generated by a Gaussian mixture model (GMM), while the target point cloud (e.g., building foundation structure point cloud) is treated as observation data. The algorithm attempts to register the data by maximizing the probabilistic similarity between the two. Secondly, an anisotropic covariance structure is introduced: unlike the standard CPD which uses the same isovariance structure for all points, the Anisotropic Coherent Point Shift Algorithm (A-CPD) assigns an anisotropic covariance matrix to each source point. This structure is derived from the principal components of the perturbation features in each direction extracted from the feature basis tensor, enabling the point cloud deformation process to adaptively differentiate based on the physical properties in the spatial direction (e.g., the principal axis of shield tunneling perturbation, the principal strain direction of settlement). Next, an expectation-maximization (EM) strategy is used to continuously update the rigid and non-rigid transformation parameters, including the affine transformation matrix, the point position offset vector, and the anisotropic covariance matrix, gradually approximating the optimal spatial alignment result. Then, the perturbation intensity extracted from the tensor feature basis and the coupled tensor tension field are introduced as deformation energy terms into the loss function. This approach balances maximum likelihood alignment with physical plausibility, ensuring that the point cloud alignment not only approximates the geometry but also reflects actual geological deformation characteristics. Finally, a spatially consistent fused point cloud set is obtained.

[0064] Step S224: Construct a three-dimensional energy transfer chain of shield cutter trajectory and micro-seismic events based on the fused point cloud, and generate a spatiotemporally coupled hybrid dimension point cloud set;

[0065] Understandably, this step begins by using B-spline fitting to fit the actual propulsion trajectory of the shield cutterhead based on the spatial locations of the building foundation nodes and geological anomalies marked in the fused point cloud, combined with the cutterhead position, propulsion speed, and rotational torque data from the shield propulsion parameters. Next, it calls upon the event waveform information recorded by the microseismic sensor array and, through a spatial solution method based on multi-station delay inversion, infers the distribution location of the microseismic sources in three-dimensional space. This location is then dynamically paired with the shield trajectory to analyze its energy transfer path and possible response delays. Then, the local energy centroid method is used to estimate the energy of each microseismic event, and, combined with timestamps and spatial locations, the microseismic energy is projected onto the shield cutterhead path, establishing a causal relationship between cutter movement and energy release. Finally, this data is fused according to the spatial dimension (geological structure / building location) and the temporal dimension (shield propulsion and vibration response time) to construct a hybrid-dimensional point cloud set containing the interaction between millimeter-level vibration sources and meter-level propulsion paths. This set not only preserves geometric spatial information but also embeds physical attributes such as energy coupling strength and propagation path direction.

[0066] Step S225: Perform variable-scale fusion processing on the mixed point cloud to construct a three-dimensional geological impact map that includes sub-millimeter fractures and meter-level structural responses.

[0067] Understandably, this step begins by extracting the geometric boundaries of cracks from the high-resolution data representing the microstructure of the cracks in the fused point cloud using surface fitting based on Moving Least Squares, and identifying microscopic properties such as crack propagation direction, density distribution, and aperture through multi-scale wavelet transform. Next, for the low-to-medium resolution point cloud portion representing the building foundation response and underground structure, a mid-scale structural representation at a unified spatial scale is constructed using voxel mesh downsampling and curvature-preserving compression methods, preserving key contours and deformation features.

[0068] Subsequently, a variable-scale fusion strategy was employed for point cloud registration and data unification, specifically through the construction of a Scale Nested Structure (SNS). This structure uses meter-level response features as the main axis, embedding sub-millimeter-level fracture features layer by layer. Local density constraints and scale weighting factors are introduced during the fusion process to achieve a joint expression of physical continuity and geometric consistency. A weighted nearest neighbor interpolation algorithm is used between fusion points to map local perturbation information in the energy transfer chain to a multi-scale space, and a three-dimensional tensor representation is constructed to carry different physical properties (such as strain, energy release rate, fracture density, etc.) at each scale. Finally, a unified visualization encoding is used to convert the processed data into a multi-channel color-vector field overlay map, forming an interactive geological impact map.

[0069] Step S3: The three-dimensional geological impact map is jointly optimized by semi-variogram modeling and spectrum co-training clustering algorithm to generate a multi-layer heterogeneous map structure;

[0070] Understandably, this step significantly enhances the ability to model the interactions of multiple factors in complex geological environments. By accurately capturing spatial statistical features through semi-variogram modeling and uncovering hidden geological heterogeneity zones through spectral co-training and clustering, the resulting multi-layered heterogeneous map provides a hierarchical and structured knowledge support framework for subsequent tectonic mechanical analysis, construction risk prediction, and structural anomaly diagnosis, exhibiting good engineering adaptability and reasoning capabilities. In this step, step S3 includes steps S31, S32, S33, S34, and S35.

[0071] Step S31: Based on the three-dimensional geological influence map, perform anisotropic semivariogram modeling and processing. By jointly optimizing the principal direction range calculation of the fracture density field and the gradient variability of the stress field, a multi-scale spatial correlation matrix is ​​obtained.

[0072] Understandably, this step begins by extracting the principal orientation information from the explicitly expressed fracture density field data in the map. This process typically uses structural tensor analysis to identify the principal axis orientation of fracture distribution in each region, and combines this with range estimation to obtain the spatial autocorrelation length of fracture density in different directions. The range here reflects the maximum distance at which variables maintain spatial correlation in a specific direction, and is a key parameter for constructing an anisotropic semivariogram model.

[0073] Next, the system introduces gradient variability analysis of the stress field as an important means to optimize the fitting accuracy of the semivariogram. This analysis calculates the gradient of the geostress tensor field and further evaluates its directional fluctuations (such as the rate of change of the direction of maximum principal stress and the degree of shear stress disturbance) to determine the spatial nonstationarity of local stress disturbances. Based on this, the system jointly optimizes two factors: the fracture principal direction range and the stress gradient variability, dynamically adjusting the structural parameters of the anisotropic semivariogram model, including the covariance structure of the principal axis direction, the shape of the range tensor, and the anisotropic weight allocation, thereby more accurately characterizing the multi-directional spatial correlation in the geological environment.

[0074] Finally, the aforementioned multi-directional, multivariate semi-variogram model was discretized to construct a multi-scale spatial correlation matrix. This matrix not only reflects the synergy of spatial point pairs at different scales and in different directions, but also integrates the multi-level coupling relationship between microscale crack perturbation and macroscale stress response, providing high-resolution spatial constraints for subsequent graph structure clustering and risk identification.

[0075] Step S32: Perform spectral Laplace construction processing based on the spatial correlation matrix to generate a weighted adjacency matrix that incorporates physical propagation characteristics;

[0076] Understandably, in this first step, the system maps the multi-scale coupling information between geological units contained in the spatial correlation matrix to node similarity in the graph structure, thus constructing preliminary adjacency relationships. Here, each spatial location point or point cloud unit is considered a node in the graph, and the edge weights between nodes are assigned by the spatial synergy between the two points, specifically considering the consistency of fracture range, the covariance characteristics of the stress field, and the degree of response coupling to microseismic disturbances.

[0077] Next, the system introduces the Laplace matrix construction mechanism from spectral graph theory. During construction, a weighted adjacency matrix based on physical correlation is first formed, with its elements representing the physical propagation coupling strength between nodes; then, the node degree matrix is ​​calculated; subsequently, the normalized Laplace matrix of the undirected graph is constructed. To further enhance physical interpretability, the system introduces a direction field weight adjustment mechanism, that is, the edge weights in the Laplace matrix are directionally weighted according to the propagation direction of the microseismic wavefront, the shield tunneling disturbance release path, and the local structural axis, thereby achieving a consistent embedding of propagation characteristics and structural morphology.

[0078] In addition, to prevent low-frequency feature leakage caused by sparse graph structure, the system also introduces a spectral domain regularization mechanism in the construction of the spectral graph, which limits the participation of high-order features (such as the concentration of perturbation energy) in the distribution construction of Laplace features. This makes the final weighted adjacency matrix not only maintain the density of node connections, but also have the low-pass filtering characteristics under frequency domain decomposition, which is beneficial to the subsequent spectral clustering and graph neural network training process.

[0079] Step S33: Perform multi-view spectrum clustering based on the weighted adjacency matrix to obtain a primary graph partition dominated by penetration paths;

[0080] Understandably, this step treats the constructed weighted adjacency matrix as the basis for a graphical representation of the geological environment. Combining the concept of multi-view learning, multiple sets of adjacency matrices are extracted from different physical characteristic dimensions (such as fracture density, stress gradient, microseismic disturbance, etc.) to form a multi-view graph dataset. Next, for each view, the corresponding graph Laplacian matrix is ​​calculated and eigenvalue decomposition is performed to extract low-dimensional spectral embeddings representing the node structure.

[0081] Subsequently, this step employs a multi-view spectral clustering algorithm to jointly optimize the spectral embeddings of each view, thereby promoting the coordinated expression of geological features under different physical attributes in a unified space. This algorithm typically includes an iteratively updated objective function, balancing consistency and diversity among views. This ensures that the clustering results not only reflect the structure from a single perspective but also integrate information from various views, improving the robustness and accuracy of the partitioning.

[0082] The clustering process focuses on seepage pathways, i.e., networks of fractures and weak surfaces where underground fluids or gases can easily migrate, ensuring that the resulting map partitions are dominated by seepage pathways, highlighting the spatial connectivity and transmission characteristics of high-risk areas. Through this partitioning, the system can clearly define the boundaries and key nodes of geologically sensitive areas, forming a preliminary risk identification framework.

[0083] Step S34: Perform vulnerable node enhancement processing based on the primary graph partition to generate an enhanced graph partition containing structural vulnerability propagation chains;

[0084] Understandably, this step, based on the initial graph partitioning results, focuses on node characteristic analysis. It uses vulnerability indices (such as degree centrality, betweenness centrality, and propagation potential) to quantitatively assess each node in the graph, identifying key vulnerable nodes that may become the starting point or bottleneck for risk propagation. Next, combining geological and physical properties and structural stress states, these vulnerable nodes undergo attribute enhancement processing, i.e., adding weights, connecting edges, or introducing auxiliary nodes into the graph structure to simulate the spatial diffusion path and impact range of vulnerability.

[0085] Subsequently, this step constructs a propagation chain model, connecting multiple vulnerable nodes in series according to propagation relationships and geological connectivity to form a structural vulnerability propagation chain. This chain structure can reflect the interactions between vulnerable nodes and their contribution to the overall risk diffusion, revealing potential cascading failure paths. The specific establishment of the propagation chain typically employs path mining algorithms from graph theory and propagation dynamics simulations, iteratively updating propagation probabilities and node states to characterize the cumulative risk effect of vulnerable node clusters.

[0086] Step S35: Perform multi-layer graph fusion processing based on the enhanced graph partitioning, and construct a multi-layer heterogeneous graph structure that characterizes millimeter-level crack propagation and meter-level structural response using tensor product graph operators.

[0087] Understandably, this step extracts layer information at different scales based on the enhanced graph partitioning results, including microscopic layers reflecting the distribution of minute cracks and macroscopic layers reflecting the overall structural response. Next, these layers are fused using a tensor product graph operator. This operator, through the mathematical mechanism of tensor products, couples the nodes and edge structures of each layer in a high-dimensional space, forming a heterogeneous graph network with a multidimensional tensor representation, enabling the interaction and integration of information at different scales and attributes.

[0088] Specifically, the tensor product graph operator first calculates the connectivity between node pairs in different layers, and then constructs a joint adjacency matrix through product mapping. This allows the fracture propagation path and the mechanical response of the structure to be expressed in parallel within a unified graph model. Simultaneously, this method effectively preserves the characteristic differences and topological structure of each graph layer, achieving cross-scale information fusion and transmission, and avoiding the limitations of single-scale models in representing complex geological phenomena. The resulting multi-layered heterogeneous graph structure is stored in tensor form, which can meticulously reflect the spatiotemporal coupling characteristics of fracture micro-behavior and macroscopic structural response in the geological environment.

[0089] Step S4: The multi-layer heterogeneous graph structure is co-evolved with a dynamic graph convolutional network and a graph integral propagation model to generate a risk cascade prediction model.

[0090] It is understandable that this step, through the co-evolution of dynamic graph convolutional networks and graph integral propagation models, not only achieves deep fusion and dynamic evolution capture of multi-scale and multi-modal geological information, but also effectively simulates the complex physical mechanisms of risk propagation, improving the accuracy and timeliness of risk prediction, and providing strong technical support for risk management in subway construction under complex and sensitive environments. In this step, step S4 includes steps S41, S42, S43, S44, and S45.

[0091] Step S41: Perform spatiotemporal graph node embedding processing based on the multi-layer heterogeneous graph structure to generate a spatiotemporally unified graph node embedding representation;

[0092] Understandably, this step first extracts multidimensional attributes for each node from the multi-layered heterogeneous graph, including geological fracture characteristics, stress response, and shield tunneling disturbances. These attributes cover different spatial scales and physical magnitudes. Next, spatiotemporal graph embedding techniques are used to encode the node data, typically including sequence modeling of temporal features (such as using temporal convolution or recurrent neural networks to capture the evolution of nodes over time) and encoding of spatial neighborhood relationships (adjacency relationships based on the graph structure). This process achieves spatiotemporal coupled embedding of node features by jointly considering the temporal dynamics of the nodes and the multi-layered spatial structure.

[0093] Step S42: Perform gated dynamic graph convolution processing based on the node embedding representation to obtain a spatiotemporal evolution feature matrix that integrates mechanical vibration and geological response;

[0094] Understandably, this step first embeds the spatiotemporal information of the nodes into an input-gated dynamic graph convolutional layer. This layer is designed with dynamic graph convolution kernels and gating mechanisms, which can adaptively adjust the influence weights of different time steps and neighboring nodes on the current node's state. The gating mechanism is similar to the gated unit (LSTM) in a recurrent neural network, which can effectively filter and memorize key temporal information, suppress noise and redundant data, thereby capturing the complex nonlinear coupling relationship between mechanical vibration signals and geological responses.

[0095] Next, dynamic graph convolution updates node representations alternately in the temporal and spatial dimensions based on dynamically changing adjacency relationships and node features, achieving joint modeling of mechanical disturbances and geological structure responses during construction. The final output is a spatiotemporal evolution feature matrix, which systematically integrates the instantaneous characteristics of mechanical vibrations and the dynamic response of geological structures, reflecting the spatial diffusion and temporal evolution characteristics of construction impacts.

[0096] Step S43: Perform graph integral propagation modeling based on the spatiotemporal feature matrix, calculate the state transition integral kernel using the node vulnerability index, and generate a propagation probability density field that considers the soil hysteresis effect.

[0097] Understandably, this step first extracts vulnerability indicators for each graph node based on the spatiotemporal feature matrix. These indicators comprehensively reflect the geological structural stability, stress state, and mechanical disturbance sensitivity of the node within the subway construction environment. Then, a state transition integral kernel function is constructed using these vulnerability indicators. This kernel function quantifies the probability and intensity of risk state transmission between graph nodes.

[0098] The kernel function is as follows:

[0099]

[0100] Among them, K i→j (t) represents the risk state transition kernel function from node i to node j within time t, x i Let x represent the spatial coordinate vector of node i. j Let σ represent the spatial coordinate vector of node j, and φ represent the spatial scale control parameter. i ψ represents the geological stability vulnerability factor of node i. j γ represents the stress response vulnerability factor of node j. ij (s) represents the perturbation coupling function at time s, t is the upper limit of integration, represents the propagation time window, and d is the differential symbol.

[0101] Next, using an integral propagation model, the state transition integral kernel is applied to the graph structure to model the dynamic propagation process of risk in the network. This process not only considers traditional spatial diffusion but also specifically introduces the soil hysteresis effect, namely the nonlinear hysteresis response and energy dissipation phenomenon of soil after being subjected to force. By introducing the hysteresis effect, the propagation probability density field can more realistically reflect the accumulation and delayed response characteristics of risk in the geological environment.

[0102] The resulting propagation probability density field is a spatiotemporally coupled probability distribution that accurately describes the diffusion pattern and temporal evolution of risk along geological structures and construction disturbance paths. This provides a solid mathematical foundation and physically meaningful model support for subsequent risk cascade analysis and prediction. The technical effect of this step is that it significantly improves the dynamic response accuracy and physical simulation realism of the risk propagation model, especially enhancing its ability to simulate complex nonlinear soil behavior.

[0103] Step S44: Perform multimodal graph signal coupling processing based on the probability density field, and obtain a risk-cascaded spatiotemporal joint propagation graph by connecting graph residuals weighted by node betweenness centrality.

[0104] Understandably, this step first extracts various risk propagation-related signal modes from the probability density field, such as multi-dimensional information like different risk types, propagation paths, and propagation speeds. Then, these modal signals are fused and coupled on a graph structure to fully capture the multi-dimensional, multi-scale propagation characteristics of risk in complex underground structures. Next, the importance of each node in the graph is quantified using the node betweenness centrality index from graph theory. This centrality is then introduced as a weight into the graph's residual connections. This residual connection mechanism effectively enhances the signal transmission capability of important nodes, strengthens the dominant role of key nodes in the risk propagation chain, and prevents information from being diluted or lost during multi-layered propagation.

[0105] Finally, a spatiotemporal joint propagation graph of risk cascades is constructed through weighted graph residual connections. This graph not only reflects the propagation path of risk along the spatial network but also incorporates the temporal evolution characteristics during the propagation process. This spatiotemporal joint propagation graph provides refined dynamic structural support for subsequent risk identification, early warning, and decision-making.

[0106] Step S45: Perform phase transition critical point prediction processing based on the propagation diagram to obtain a three-dimensional risk cascade prediction model that includes risk threshold and propagation path.

[0107] Understandably, this step first analyzes the state change trends of nodes and edges in the propagation graph, and, combined with criticality theory, identifies phase transition critical points in the network structure that may trigger rapid risk propagation. This step typically involves constructing a risk propagation dynamics model, using multiple indicators such as propagation probability density, node betweenness centrality, and the cumulative impact of propagation paths to dynamically assess the threshold for the system to transition from a steady state to an unstable state. Next, critical point detection algorithms are used, such as critical threshold search based on the maximum entropy principle, or continuous / discrete critical interval division based on phase transition theory, to accurately locate key nodes and propagation paths that may erupt during the cascading risk propagation process.

[0108] Finally, by integrating the above analysis results, a three-dimensional risk cascade prediction model was formed, incorporating risk thresholds and specific propagation paths. This model not only depicts the spatial distribution characteristics of risk propagation but also reflects the changing patterns of propagation intensity over time. This model can provide engineering managers with critical indicators for risk early warning and tools for tracking risk paths, facilitating the early implementation of preventative measures.

[0109] Step S5: Solve the prediction model and perform multi-objective feasible region mapping based on the solution results to obtain a global management scheme for subway construction.

[0110] It is understandable that this step, through system solution and multi-objective mapping, constructs a scientific and reasonable risk control and construction management decision support framework, improving the overall safety and efficiency of construction, effectively reducing potential losses caused by risk cascading, and ensuring the smooth progress of the subway project and the safety of the surrounding environment. In this step, step S5 includes steps S51 and S52.

[0111] Step S51: Solve the prediction model by combining fractional path integral and curvature-driven stochastic differentiation to derive the multidimensional gradient response surface of the construction parameters to the geological response.

[0112] It is understood that in this step, step S51 includes steps S511, S512, S513, S514 and S515.

[0113] Step S511: Based on the risk cascade prediction model, construct fractional-order control equations to generate fractional-order stochastic differential equations with time delay terms.

[0114] Understandably, this step first introduces fractional derivative operations, leveraging their ability to characterize the long memory effect and nonlocal behavior of the system to capture the hysteresis effect of soil and the historical dependence characteristics in the structural response. Next, considering the stochasticity and time delay of construction risk propagation, a time delay term is added to the model to reflect the delay in information transmission and response. Simultaneously, stochastic differential equations are introduced to describe the effects of uncertain disturbances and noise in the system. This fractional-order stochastic differential equation form integrates non-integer-order differential operators, time delay factors, and stochastic excitations, achieving high-precision dynamic modeling of the cascading evolution of risks.

[0115] The stochastic differential equation is as follows:

[0116]

[0117] in, Let Σ(X(t)) represent the Caputo fractional derivative of order α∈(0,1] applied to the state function X(t), reflecting the system's historical memory characteristic and often used to describe the "hysteresis behavior" of formation response. A represents the deterministic state matrix of the system, B represents the time-delay feedback matrix, reflecting the influence of historical states on current changes, τ represents the time delay length (a constant), representing the lag time of the system in the propagation response, f(X(t)) represents the nonlinear control function of the state, used to capture complex physical coupling behavior in the system, and Σ(X(t)) is the state-dependent diffusion coefficient matrix, representing the coupling relationship between the disturbance intensity and the system state. The Wiener process (white noise) represents the random disturbances experienced by the system, mainly simulating uncertainties such as underground micro-seismic activity and ground stress fluctuations. X(t) is the state variable, which is usually a vector combination of key parameters of the construction system, such as cutterhead propulsion speed, shield attitude, and underground stress state. t represents the time variable.

[0118] In practical applications, this modeling approach breaks through the limitations of traditional integer-order models and can meticulously simulate the complex nonlinear time-varying characteristics of risk propagation in the subway construction environment. In particular, it demonstrates stronger sensitivity and accuracy in predicting the evolution trend of node risks and key critical behaviors.

[0119] Step S512: Perform path integral solving based on the fractional equation to obtain the path probability weight distribution considering the anisotropy of the strata;

[0120] Understandably, this step first transforms the fractional-order stochastic differential equation into an integral expression over the path space. By integrating the weights of all possible paths, the probability contribution of different propagation paths is systematically calculated. Next, considering the anisotropic parameters of the formation, such as elastic modulus, permeability, and fracture steering, the path weights are weighted and corrected during the path integration process. This reflects the impact of the heterogeneous response of the geological medium in different directions on the selection of risk propagation paths.

[0121] This process can meticulously capture the spatial dependence and directional differences in the risk propagation process, overcoming the limitations of traditional methods that ignore stratigraphic heterogeneity, and making the path probability weight distribution more consistent with actual geological conditions. In practical applications, this path integral solution not only improves the spatial resolution of risk path prediction, but also provides accurate basis for identifying key prevention and control areas during construction.

[0122] Step S513: Perform curvature-driven field construction processing based on the weight distribution to generate a stochastic dynamic system that incorporates geometric constraints;

[0123] Understandably, this step first calculates the curvature distribution of the risk propagation path in space. This curvature information reflects the degree of curvature and geometric complexity of the path, thereby revealing the impact of geological structure changes on path evolution during risk propagation. Next, this curvature field is used as the driving force in the dynamic system and introduced into the stochastic dynamic system framework. By constructing partial differential equations containing stochastic perturbation terms and curvature driving terms, the evolution of risk under complex geological geometric constraints is simulated.

[0124] The partial differential equations containing random perturbation terms and curvature-driven terms are shown below:

[0125]

[0126] Where R(x,t) represents the risk response function of spatial location x at time t, reflecting the local risk intensity or density. The curvature-driven diffusion term describes the geometric modulation effect of the degree of geological structural curvature on the risk diffusion path. Let k(x) represent the rate of evolution of the risk response over time, and k(x) be the path curvature field at spatial location x. For the spatial gradient of risk response, The random disturbance term simulates uncertainties such as micro-seismic events, mechanical disturbances, and parameter uncertainties. σ(x,t) is the disturbance intensity function, controlling the spatial and temporal intensity distribution of the disturbance. The stochastic evolution of the system is introduced by the space-time Brownian motion derivative (white noise).

[0127] In this process, geometric constraints not only control the spatial morphology of the risk propagation path but also reflect the influence of unpredictable factors in the geological environment through random perturbations, enabling the system to dynamically adapt to changing underground conditions. This stochastic dynamic system, by incorporating curvature information, can more accurately simulate the nonlinear, multi-scale dynamic behavior of risk propagation. In practical applications, this step enhances the analytical capability for the morphology of risk diffusion paths, providing theoretical support for spatial early warning and dynamic management of construction risks.

[0128] Step S514: Perform joint solution processing on the stochastic dynamical system to obtain the multi-scale covariant gradient distribution in the parameter space;

[0129] Understandably, this step first constructs a joint set of equations including curvature driving force and random perturbation terms, and then uses the finite element method to iteratively calculate this high-dimensional complex system. The joint solution process not only considers the coupling effects between parameters, but also decomposes the covariant behavior at different spatial scales through multi-scale analysis, achieving accurate estimation of parameter gradients in both local and global ranges.

[0130] This multi-scale covariant gradient distribution reveals the sensitivity and uncertainty of key parameters in the system to risk propagation paths and evolution processes, helping to identify potential risk diffusion mechanisms at different scales. In practical applications, this step provides a sophisticated mathematical foundation for subsequent risk control strategy optimization and parameter tuning, making risk management more targeted and effective. The technical benefits are manifested in obtaining a multi-scale, spatially correlated gradient distribution through joint solution, significantly improving the model's analytical capability and prediction accuracy for dynamic risk changes in complex geological and construction environments.

[0131] Step S515: Perform manifold dimensionality reduction processing based on the multi-scale covariant gradient distribution to generate a multi-dimensional gradient response surface that reflects the synergistic effect of thrust, rotational speed, and grouting volume.

[0132] Understandably, this step first captures the local geometric relationships between data points by constructing an adjacency graph of the high-dimensional gradient data. Then, manifold learning methods, such as Locally Linear Embedding (LLE), Laplacian Eigenmaps, or t-SNE, are applied to gradually map the complex high-dimensional gradient distribution onto a two-dimensional or three-dimensional low-dimensional manifold, forming a continuous and smooth multidimensional gradient response surface. This surface visually demonstrates the nonlinear coupling relationship between thrust, rotational speed, and grouting volume, and their combined impact on formation response.

[0133] By using manifold dimensionality reduction, this step effectively avoids information loss and linearity assumptions in traditional dimensionality reduction methods, enabling complex geomechanical response laws to be expressed in a visual and physically meaningful way, which facilitates subsequent risk assessment and parameter control.

[0134] Step S52: Perform variable structure backpropagation and multi-objective feasible region mapping processing on the response surface to output a global management scheme that includes construction parameter control charts, risk warning distribution and dynamic adjustment strategies.

[0135] It is understandable that this step can not only reflect changes in construction risks in real time and improve the accuracy and timeliness of risk warnings, but also guide on-site operators to dynamically adjust construction parameters according to different risk levels, significantly improving the safety and management level of subway construction. Finally, it outputs a global management solution that includes construction parameter control charts, risk warning distribution and dynamic adjustment strategies, supporting intelligent control of the entire construction process. In this step, step S52 includes steps S521, S522, S523, S524 and S525.

[0136] Step S521: Perform variable structure neural network construction processing based on the multidimensional gradient response surface to generate an adaptive network architecture with parameter sensitivity weighting.

[0137] Understandably, this step employs a variable structure neural network construction method. First, sensitivity analysis is performed on each parameter dimension of the response surface to identify the relative strength of the impact of construction parameters such as thrust, rotational speed, and grouting volume on the overall system performance. Based on this sensitivity information, the network structure automatically adjusts its number of layers, nodes, and connection weights to achieve weighted parameter allocation. This means assigning higher weights and richer expressive power to key parameters, while simplifying the processing of parameters with less influence, thereby effectively reducing model complexity while ensuring prediction accuracy.

[0138] Next, the network employs an adaptive training mechanism, continuously iteratively updating weights and structure to achieve real-time adaptation to changes in the construction environment and multi-scale response characteristics. The variable structure neural network can flexibly adjust its structure according to changes in data distribution and input characteristics, avoiding underfitting or overfitting problems that may occur with fixed-structure networks. Furthermore, this network design incorporates multi-dimensional information from gradient response surfaces, enhancing the model's ability to express the nonlinear coupling relationships of construction parameters. The main technical effects are: improved accuracy in capturing and dynamically modeling the sensitivity of construction parameters, enabling the model to not only accurately reflect the impact of each parameter on risk and efficiency, but also to adaptively optimize its structure in response to environmental changes, significantly enhancing the reliability and robustness of risk prediction and construction control schemes, thus providing a solid intelligent foundation for subsequent multi-objective optimization and decision-making.

[0139] Step S522: Perform feasible region boundary detection processing based on the adaptive network architecture to obtain the parametric feasible region topology mapping under non-convex constraints;

[0140] Understandably, this step first inputs construction parameters (such as thrust, rotation speed, grouting volume, etc.) and their dynamic response data, and then uses a network to predict and evaluate the risk indicators and constraints corresponding to each parameter combination. Next, using high-dimensional non-convex optimization techniques combined with topology data analysis methods, parameter combinations that meet safety, efficiency, and environmental constraints in the prediction results are screened and marked, gradually outlining the feasible domain boundary in the parameter space that satisfies all constraints.

[0141] In this process, the system dynamically identifies complex non-convex shapes and multi-connected subsets in the feasible region through a nonlinear boundary detection algorithm, avoiding the simplification bias of traditional convex optimization methods in the face of real construction environments. Subsequently, using topology mapping technology, the high-dimensional parameter feasible region is projected onto a low-dimensional space while preserving its topological structural features, generating a topology mapping map of the parameter feasible region. This mapping not only clearly shows the overall layout and boundary morphology of feasible parameter combinations but also provides a visual basis for multi-objective optimization and decision-making.

[0142] Step S523: Perform multi-objective Pareto front search processing based on the parameter feasible region mapping to generate a collaborative control solution set of thrust-speed-grouting quantity;

[0143] Understandably, this step treats these three types of parameters as mutually coupled control variables, combining multi-dimensional objective functions such as risk warning, construction efficiency, and environmental impact, and iteratively searches within the high-dimensional feasible region using a multi-objective particle swarm optimization algorithm. First, an initial solution set is randomly generated. Then, in each generation, selection, crossover, and mutation operations are used to maintain the diversity and superiority of the solution set. Simultaneously, non-dominated sorting is used to classify the solution set into levels, ensuring that all solutions are Pareto optimal solutions that cannot be simultaneously dominated by other solutions under the current conditions.

[0144] Next, the algorithm calculates the congestion distance based on the objective function value, further maintaining the uniformity of the solution set distribution and avoiding local optima traps. As iterations proceed, the solution set gradually converges to the Pareto front, representing the optimal coordinated control scheme that balances thrust, rotational speed, and grouting volume. The final output coordinated control solution set not only reflects the multi-objective trade-offs of construction parameters but also takes into account safety, efficiency, and environmental friendliness during subway construction.

[0145] Step S524: Perform reverse tracking of the risk propagation path based on the collaborative control solution set of thrust-rotation speed-grouting volume to obtain the dynamic risk threshold surface;

[0146] Understandably, this step starts with a known risk cascade prediction model and propagation map, and uses a backpropagation algorithm or a reverse Monte Carlo path sampling method to trace the risk diffusion path in reverse along key nodes in the risk propagation network that may trigger safety accidents or geological anomalies. This process achieves precise location of risk diffusion paths under different construction conditions by dynamically adjusting the propagation probability density and node vulnerability indicators, combined with the changing trajectory of multi-dimensional construction parameters.

[0147] Next, this step integrates the risk propagation information obtained from reverse tracing in a three-dimensional parameter space (thrust, rotational speed, grouting volume) to construct a dynamically changing risk threshold surface. This surface dynamically reflects the critical boundary between formation stability and safety risk under various combinations of construction parameters, and can reveal in real time which parameter combinations will trigger critical states of risk during construction.

[0148] Step S525: Perform multi-scale strategy fusion processing based on the dynamic risk threshold surface, and generate a three-dimensional construction management cube containing construction parameter envelope, risk heat map and phase adjustment matrix through affine transformation.

[0149] Understandably, this step begins by considering different spatial and temporal scales, taking into account the dynamic range of construction parameters (such as thrust, rotation speed, and grouting volume), the characteristics of risk thermal distribution, and the diversity of construction stages. Affine transformation is then used to geometrically map and scale this multidimensional information. Affine transformation preserves the linear relationship and spatial topology of the data, ensuring that the correspondence between risk thresholds and construction parameters remains undistorted during the transformation process.

[0150] Next, this step generates a three-dimensional construction management cube containing three parts: First, the construction parameter envelope, which reflects the fluctuation range of construction parameters within the safety boundary and provides a boundary reference for parameter adjustment; Second, the risk heat map, which intuitively displays the risk distribution density and intensity under different parameter combinations, helping construction managers to quickly identify high-risk areas; Third, the phase adjustment matrix, which describes the phase relationship between parameter timing and risk response during the construction process, supporting timing control and strategy optimization.

[0151] Example 2:

[0152] like Figure 2 As shown, this embodiment provides a subway construction management device for complex and sensitive environments. See [link / reference]. Figure 2 The device includes an acquisition unit 701, a processing unit 702, an optimization unit 703, an evolution unit 704, and a calculation unit 705.

[0153] The acquisition unit 701 is used to acquire first information, which includes time-series data of shield tunneling parameters in the construction area, building settlement monitoring data, underground pipeline distribution data, and microseismic sensor array data.

[0154] Processing unit 702 is used to preprocess the first information to obtain a three-dimensional geological impact map that integrates underground structure, stress response and dynamic disturbance;

[0155] Optimization unit 703 is used to jointly optimize the three-dimensional geological influence map by performing semi-variogram modeling and spectrum co-training clustering algorithm to generate a multi-layer heterogeneous map structure.

[0156] Evolution unit 704 is used to perform dynamic graph convolutional network and graph integral propagation model co-evolution on the multi-layer heterogeneous graph structure to generate a risk cascade prediction model.

[0157] The calculation unit 705 is used to solve the prediction model and perform multi-objective feasible domain mapping based on the solution results to obtain a global management scheme for subway construction.

[0158] It should be noted that the specific manner in which each module performs its operation in the apparatus described in the above embodiments has been described in detail in the embodiments of the method, and will not be elaborated here.

[0159] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0160] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for managing subway construction in complex and sensitive environments, characterized in that, include: Obtain first information, which includes time-series data of shield tunneling parameters in the construction area, building settlement monitoring data, underground pipeline distribution data, and microseismic sensor array data; The first information is preprocessed to obtain a three-dimensional geological impact map that integrates underground structure, stress response, and dynamic disturbance. The three-dimensional geological impact map is jointly optimized by semi-variogram modeling and spectrum co-training clustering algorithm to generate a multi-layer heterogeneous map structure. The multi-layer heterogeneous graph structure is co-evolved with a dynamic graph convolutional network and a graph integral propagation model to generate a risk cascade prediction model. The prediction model is solved, and multi-objective feasible region mapping is performed based on the solution results to obtain a global management scheme for subway construction. The first information is preprocessed to obtain a three-dimensional geological impact map integrating underground structure, stress response, and dynamic disturbance, including: The first information is processed by empirical mode decomposition, nonnegative tensor decomposition and generalized regression neural network to generate a spatiotemporally aligned structured multi-source dataset. The spatiotemporally aligned structured multi-source dataset is coupled with a multi-core tensor support machine and a coherent point drift algorithm to construct a three-dimensional geological impact map that integrates underground structure, stress response, and dynamic disturbance. Specifically, the multi-layer heterogeneous graph structure is co-evolved with a dynamic graph convolutional network and a graph integral propagation model to generate a risk cascade prediction model, including: Based on the multi-layer heterogeneous graph structure, spatiotemporal graph node embedding processing is performed to generate a spatiotemporally unified graph node embedding representation; Gated dynamic graph convolution processing is performed based on the node embedding representation to obtain a spatiotemporal evolution feature matrix that integrates mechanical vibration and geological response; Based on the spatiotemporal evolution feature matrix, graph integral propagation modeling is performed, and the state transition integral kernel is calculated using the node vulnerability index to generate a propagation probability density field that considers the soil hysteresis effect. Multimodal graph signal coupling processing is performed based on the probability density field, and a risk-cascaded spatiotemporal joint propagation graph is obtained through graph residual connection weighted by node betweenness centrality. Based on the propagation diagram, a phase transition critical point prediction process is performed to obtain a three-dimensional risk cascade prediction model that includes risk thresholds and propagation paths. The prediction model is solved, and multi-objective feasible region mapping is performed based on the solution results to obtain a global management scheme for subway construction, including: The prediction model is solved by combining fractional path integral and curvature-driven stochastic differentiation to derive the multidimensional gradient response surface of construction parameters to geological response. The response surface is processed by variable structure backpropagation and multi-objective feasible domain mapping to output a global management scheme that includes construction parameter control charts, risk warning distribution, and dynamic adjustment strategies.

2. The method for managing subway construction in complex and sensitive environments according to claim 1, characterized in that... The first information is processed by a combination of empirical mode decomposition, nonnegative tensor decomposition, and generalized regression neural network, including: Empirical mode decomposition is performed on the time series data of the tunnel boring machine propulsion parameters to obtain a set of intrinsic mode functions that include the main frequency component of the cutterhead torque. Based on the eigenmode function set, a kurtosis-sample entropy joint criterion screening process is performed to obtain the denoised shield tunneling parameter dynamic component matrix; Tensor modeling was performed based on the building settlement monitoring data and underground pipeline distribution data. After third-order non-negative tensor modeling and tensor chain decomposition, the multi-field coupling tensor of the repaired underground structure was obtained. Based on the dynamic component matrix and the multi-field coupling tensor, a spatiotemporal mapping process of a generalized regression neural network is performed to obtain a fusion feature tensor with spatiotemporal labels. Modal correction is performed based on the fused feature tensor, followed by conjugate gradient optimization with partial coherence function constraints, resulting in a spatiotemporally strictly aligned structured multi-source dataset.

3. The method for managing subway construction in complex and sensitive environments according to claim 1, characterized in that... The spatiotemporally aligned structured multi-source dataset is coupled with a multi-core tensor support machine and a coherence point drift algorithm, including: Based on the structured multi-source dataset, a multi-kernel tensor support machine is used to process the data. Geological static features, stress dynamic features, and shield tunneling disturbance time series features are extracted by Gaussian kernel, polynomial kernel, and linear kernel function group, respectively, to generate a cross-scale feature weight matrix. Based on the weight matrix, kernel target alignment is performed to obtain a low-rank tensor feature basis. Anisotropic coherent point drift processing is performed based on the aforementioned feature base to non-rigidly align the geological anomaly with the building foundation point cloud, resulting in a spatially consistent fused point cloud. Based on the fused point cloud, a three-dimensional energy transfer chain of shield cutter trajectory and microseismic events is constructed, generating a spatiotemporally coupled hybrid dimension point cloud set; By performing variable-scale fusion processing on a mixed-dimensional point cloud set, a three-dimensional geological impact map containing sub-millimeter fractures and meter-level structural responses is constructed.

4. The method for managing subway construction in complex and sensitive environments according to claim 1, characterized in that... The three-dimensional geological impact map is jointly optimized using a semi-variogram model and a spectrum co-training clustering algorithm to generate a multi-layer heterogeneous map structure, including: Based on the three-dimensional geological influence map, anisotropic semivariogram modeling was performed. By jointly optimizing the principal direction range calculation of the fracture density field and the gradient variability of the stress field, a multi-scale spatial correlation matrix was obtained. Based on the spatial correlation matrix, a spectral Laplace construction process is performed to generate a weighted adjacency matrix that incorporates physical propagation characteristics; Multi-view spectrum clustering is performed based on the weighted adjacency matrix to obtain a primary graph partition dominated by penetration paths; Based on the primary graph partition, vulnerable nodes are enhanced to generate an enhanced graph partition containing the structural vulnerability propagation chain; Multi-layer graph fusion processing is performed based on the enhanced graph partitioning, and a multi-layer heterogeneous graph structure characterizing millimeter-level crack propagation and meter-level structural response is constructed using tensor product graph operators.

5. The method for managing subway construction in complex and sensitive environments according to claim 1, characterized in that... The prediction model is solved jointly by fractional-order path integrals and curvature-driven stochastic differentials, including: The risk cascade prediction model is used to construct fractional-order control equations to generate fractional-order stochastic differential equations with time delay terms. The path integral is solved based on the fractional-order stochastic differential equation to obtain the path probability weight distribution considering the anisotropy of the strata. Based on the weight distribution, curvature-driven field construction is performed to generate a stochastic dynamic system that incorporates geometric constraints; By performing joint solution processing on the stochastic dynamical system, a multi-scale covariant gradient distribution in the parameter space is obtained; Based on the multi-scale covariant gradient distribution, manifold dimensionality reduction is performed to generate a multi-dimensional gradient response surface that reflects the synergistic effect of thrust, rotational speed, and grouting volume.

6. The method for managing subway construction in complex and sensitive environments according to claim 1, characterized in that... The response surface is subjected to variable structure backpropagation and multi-objective feasible region mapping, including: Based on the multidimensional gradient response surface, a variable structure neural network is constructed to generate a parameter sensitivity-weighted adaptive network architecture. Based on the adaptive network architecture, feasible region boundary detection processing is performed to obtain the parametric feasible region topology mapping under non-convex constraints. Based on the feasible region mapping of the parameters, a multi-objective Pareto front search is performed to generate a collaborative control solution set of thrust-speed-grouting quantity. Based on the collaborative control solution set of thrust-rotation speed-grouting volume, reverse tracing of the risk propagation path is performed to obtain the dynamic risk threshold surface; Based on the dynamic risk threshold surface, multi-scale strategy fusion processing is performed, and a three-dimensional construction management cube containing construction parameter envelopes, risk heatmaps, and phase adjustment matrices is generated through affine transformation.

7. A subway construction management device for complex and sensitive environments, characterized in that, include: The acquisition unit is used to acquire first information, which includes time-series data of shield tunneling parameters in the construction area, building settlement monitoring data, underground pipeline distribution data, and microseismic sensor array data. The processing unit is used to preprocess the first information to obtain a three-dimensional geological impact map that integrates underground structure, stress response, and dynamic disturbance. The optimization unit is used to jointly optimize the three-dimensional geological impact map by performing semi-variogram modeling and spectrum co-training clustering algorithm to generate a multi-layer heterogeneous map structure. An evolution unit is used to co-evolve the multi-layer heterogeneous graph structure with a dynamic graph convolutional network and a graph integral propagation model to generate a risk cascade prediction model. The computing unit is used to solve the prediction model and perform multi-objective feasible domain mapping based on the solution results to obtain a global management scheme for subway construction. The first information is preprocessed to obtain a three-dimensional geological impact map integrating underground structure, stress response, and dynamic disturbance, including: The first information is processed by empirical mode decomposition, nonnegative tensor decomposition and generalized regression neural network to generate a spatiotemporally aligned structured multi-source dataset. The spatiotemporally aligned structured multi-source dataset is coupled with a multi-core tensor support machine and a coherent point drift algorithm to construct a three-dimensional geological impact map that integrates underground structure, stress response, and dynamic disturbance. Specifically, the multi-layer heterogeneous graph structure is co-evolved with a dynamic graph convolutional network and a graph integral propagation model to generate a risk cascade prediction model, including: Based on the multi-layer heterogeneous graph structure, spatiotemporal graph node embedding processing is performed to generate a spatiotemporally unified graph node embedding representation; Gated dynamic graph convolution processing is performed based on the node embedding representation to obtain a spatiotemporal evolution feature matrix that integrates mechanical vibration and geological response; Based on the spatiotemporal evolution feature matrix, graph integral propagation modeling is performed, and the state transition integral kernel is calculated using the node vulnerability index to generate a propagation probability density field that considers the soil hysteresis effect. Multimodal graph signal coupling processing is performed based on the probability density field, and a risk-cascaded spatiotemporal joint propagation graph is obtained through graph residual connection weighted by node betweenness centrality. Based on the propagation diagram, a phase transition critical point prediction process is performed to obtain a three-dimensional risk cascade prediction model that includes risk thresholds and propagation paths. The prediction model is solved, and multi-objective feasible region mapping is performed based on the solution results to obtain a global management scheme for subway construction, including: The prediction model is solved by combining fractional path integral and curvature-driven stochastic differentiation to derive the multidimensional gradient response surface of construction parameters to geological response. The response surface is processed by variable structure backpropagation and multi-objective feasible domain mapping to output a global management scheme that includes construction parameter control charts, risk warning distribution, and dynamic adjustment strategies.

Images