Decision-making method for underground construction based on 3D geological modeling and risk hotspot identification

Abstract

The present invention discloses a method for underground construction decision-making based on three-dimensional geological modeling and risk hot zone identification, which relates to the field of integration of artificial intelligence and geological engineering. The method first obtains drilling data, geological radar images and seismic reflection layer information, constructs a three-dimensional geological voxel model with spatial topological constraints, and uses an irregular grid method to accurately describe the geological unit structure. Subsequently, the temporal characteristic indicators under construction disturbance are extracted to form a continuous time series feature vector, which is input into a convolutional recursive neural network model with a spatial attention aggregation mechanism and a deep memory unit to predict the risk heat value of each spatial position. On this basis, dynamic high-risk hot zones are identified through heat gradient clustering and neighborhood consistency analysis, and a risk hot zone map is constructed. Finally, a multi-objective auxiliary decision function is constructed by combining the construction stage, equipment plan and sensor feedback information to generate a construction decision result including operation path reconstruction, rhythm adjustment and power limit control suggestions.

Description

Technical Field

[0001] The present invention relates to the field of neural network technology in artificial intelligence, and in particular to an underground construction decision-making method based on three-dimensional geological modeling and risk hot zone identification. Background Art

[0002] Existing technologies for risk assessment and decision-making in underground construction rely primarily on two-dimensional geological profiles, empirical parameter models, and limited monitoring data. These methods often employ rule-based analysis methods or simple statistical models for risk classification and construction path planning. These methods typically focus on static risk assessment and are limited in their ability to express underground spatial structures and model dynamic changes, making it difficult to accurately capture the evolution of risk under complex geological conditions. In recent years, some studies have attempted to incorporate three-dimensional modeling and machine learning methods to aid in decision-making, but these efforts have largely remained at the experimental verification stage and have yet to establish a closed-loop decision-making system.

[0003] However, existing technologies still have numerous shortcomings in processing multi-source heterogeneous geological data, responding to dynamic construction disturbances, and identifying high-risk areas. For one thing, 3D modeling often uses a regular grid approach, making it difficult to faithfully reproduce the irregular structure of geological bodies. Furthermore, traditional models struggle to effectively model the nonlinear coupling characteristics of geological processes, resulting in delayed and inaccurate hotspot identification. Furthermore, current construction assistance systems primarily rely on offline analysis, lacking linkage with real-time monitoring information and ineffectively supporting construction decision-making optimization under multi-objective conditions.

[0004] In order to solve the above problems, it is urgent to provide an underground construction decision-making assistance method that has artificial intelligence modeling capabilities and supports multi-source data fusion and dynamic response analysis. Summary of the Invention

[0005] This application provides an underground construction decision-making method based on three-dimensional geological modeling and risk hot zone identification. It realizes the dynamic identification of underground risk hot zones under construction disturbance through deep neural networks, thereby providing a quantifiable, updateable, and spatially continuous intelligent auxiliary decision-making basis for the underground construction process.

[0006] This application provides an underground construction decision-making method based on three-dimensional geological modeling and risk hot zone identification, including:

[0007] Acquire multi-source geological data including borehole data, geological radar images, and seismic reflection layer information to construct a three-dimensional geological voxel model with spatial topological constraints. The three-dimensional geological voxel model uses an irregular grid reconstruction method to describe the distribution characteristics of each geological unit;

[0008] Based on the three-dimensional geological voxel model, extracting time series characteristic indicators under the influence of construction disturbance, including local pressure gradient changes, seepage velocity estimation and structural response heterogeneity factors, to form a continuous time series characteristic vector;

[0009] Inputting the continuous time series feature vector into a trained convolutional recurrent neural network model, which has a spatial attention aggregation mechanism and deep memory units, is used to model the nonlinear risk propagation behavior during the construction evolution process and output the risk heat value of each spatial location;

[0010] Based on the distribution characteristics of risk heat values ​​in three-dimensional space, the heat gradient clustering and neighborhood structure consistency screening algorithm are used to identify spatially continuous and dynamically changing high-risk hot spots, forming a dynamic risk hot spot map.

[0011] Combining the current construction stage, equipment deployment plan and sensor feedback information, a multi-objective auxiliary decision-making function is constructed based on the dynamic risk hot zone map. Taking construction efficiency, personnel safety and energy consumption constraints into comprehensive consideration, an auxiliary decision-making result including operation path reconstruction, construction rhythm adjustment and equipment power limit control suggestions is generated through iterative parameter optimization.

[0012] The beneficial effects of the technical solution provided by this application include:

[0013] (1) By using irregular grids to construct a three-dimensional geological voxel model with spatial topological constraints, the distribution of irregular underground geological units can be more realistically restored, providing a high-fidelity data foundation for subsequent risk analysis. (2) The introduction of a convolutional recurrent neural network with spatial attention aggregation and deep memory mechanism effectively captures the nonlinear coupling characteristics of geological parameters evolving over time, significantly improving the timeliness and accuracy of risk prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a flowchart of an underground construction decision-making method based on three-dimensional geological modeling and risk hot zone identification provided in the first embodiment of the present application. DETAILED DESCRIPTION

[0015] The following description sets forth many specific details to facilitate a thorough understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of the present application. Therefore, the present application is not limited to the specific implementations disclosed below.

[0016] The first embodiment of the present application provides an underground construction decision-making method based on three-dimensional geological modeling and risk hot zone identification. Figure 1 , which is a schematic diagram of the first embodiment of this application. Figure 1The first embodiment of the present application provides a detailed description of an underground construction decision-making method based on three-dimensional geological modeling and risk hot zone identification.

[0017] Step S101: Acquire multi-source geological data including borehole data, geological radar images, and seismic reflection layer information, and construct a three-dimensional geological voxel model with spatial topological constraints. The three-dimensional geological voxel model uses an irregular grid reconstruction method to describe the distribution characteristics of each geological unit.

[0018] In step S101, the system first needs to obtain multi-source heterogeneous original geological data from the underground engineering area. The data should include but not be limited to drilling data, geological radar images and seismic reflection layer information. Drilling data usually comes from on-site geophysical and geological exploration activities, recording the lithology distribution, water content, porosity, compaction and other properties of strata at different depths. The accuracy of its spatial distribution depends on the drilling density and layout method. Geological radar images obtain images of shallow underground structures through high-frequency electromagnetic wave scanning, which are used to identify abnormal areas such as stratum interfaces, fault fracture zones and underground cavities. Seismic reflection layer information is analyzed by laying out excitation sources and detector arrays to form reflection profiles formed by seismic waves reflected at different medium interfaces, which can be used to identify deep underground structural morphology and rock continuity.

[0019] To ensure modeling accuracy, all acquired data undergoes standardized preprocessing, including coordinate unification, sampling interval resampling, outlier removal, interpolation completion, and unified attribute encoding. For borehole data, Kriging interpolation or inverse distance weighted methods can be used to transition point data to volume units. For geological radar images and seismic reflection profiles, image registration and spatial projection transformation are required to map them and the borehole data to a common three-dimensional coordinate system.

[0020] After data fusion is complete, irregular grid modeling is used to model the 3D geological volume. Compared to traditional regular grids (such as cubic voxels or equidistant tetrahedral grids), irregular grid modeling is more suitable for irregular geological interfaces and tectonic zones. The grid density and shape can be adaptively adjusted according to the geological complexity of different regions. For example, higher-density polyhedral cells can be used in fault zones, highly deformed areas, or areas with heterogeneous rock formations, while larger simplified cells can be used in geologically stable areas to improve overall modeling efficiency and accuracy.

[0021] The modeling process includes the following key sub-steps. First, voxel segmentation is initialized. A control point network is generated based on the intersection of drillholes and radar lines, and an initial irregular grid is constructed using constrained Delaunay 3D triangulation. A geological unit labeling strategy based on geological stratification logic is then introduced. The model space is labeled by stratigraphic type, geological age, or lithologic characteristics, and each voxel node is assigned a corresponding geological attribute value, such as lithology, porosity, elastic modulus, and conductivity.

[0022] To ensure that 3D geological voxel models possess spatial topological constraints, the connectivity relationships between voxel units must be explicitly recorded during model construction. This relationship includes not only geometric adjacency (e.g., shared planarity, shared edges, shared vertices) but also geological logical topology, such as structural boundaries like interlayer interfaces, fracture surfaces, and contact interfaces. Using an adjacency matrix or graph data structure allows for a complete representation of spatial topological information, facilitating subsequent tracing of information propagation paths in 3D space, identifying risk coupling areas, and performing simulations.

[0023] The resulting 3D geological voxel model is not only a spatial structural entity but also capable of data loading that evolves over time. This means that in addition to preserving structural properties for each voxel unit, it also reserves dynamic data interfaces for binding time-varying parameter values, such as construction disturbance response, water pressure fluctuations, or temperature changes. This capability provides a high-dimensional, high-resolution input foundation for subsequent dynamic modeling and neural network learning.

[0024] In summary, step S101 not only completes the multi-source fusion and standardized expression of geological data, but also realizes the complete modeling foundation of the geological environment from static geometric structure to dynamic construction interaction process through high-fidelity irregular three-dimensional voxel modeling, providing irreplaceable input data support and spatial logic framework for subsequent construction disturbance feature extraction, risk heat prediction and auxiliary decision analysis.

[0025] Furthermore, the acquisition of multi-source geological data including borehole data, geological radar images, and seismic reflection layer information to construct a three-dimensional geological voxel model with spatial topological constraints includes:

[0026] Based on the borehole data, geological radar images and seismic reflection layer information, a geological profile set is generated and a preliminary model of a three-dimensional geological structure is constructed to drive the division boundary of the three-dimensional geological voxel grid;

[0027] Based on the fault zones, lithologic transition zones, and structural boundary areas identified in the preliminary 3D geological structure model, a voxel refinement strategy is implemented to locally densify the grid in highly heterogeneous areas, thereby forming an irregular grid partitioning structure with controlled partition density.

[0028] The geological information entropy distribution map is calculated based on multi-source geological data to evaluate the geological complexity of each region in three-dimensional space. Based on this, the density of the voxel grid is dynamically adjusted to refine the information-complex areas and coarsen the information-uniform areas.

[0029] For each voxel node after irregular grid division, geological unit number information and spatial topology label are added. The former is used to identify the stratigraphic category and lithologic type to which it belongs, and the latter is used to identify the spatial connection relationship and structural boundary attributes between it and adjacent voxels.

[0030] A three-dimensional geological voxel model with spatial topological constraints was constructed. The model uses irregular grid reconstruction to describe the distribution characteristics of various geological units. It has the ability to express fracture structures, stratigraphic boundaries and spatial discontinuities, providing a structural basis for subsequent underground construction risk analysis.

[0031] In the present invention, the three-dimensional geological voxel model serves as the structural basis of the entire underground construction risk identification and auxiliary decision-making process. Its accuracy, spatial expression ability and topological structure integrity will directly determine the effectiveness of subsequent key steps such as disturbance modeling, hot zone identification and path control.

[0032] The present invention first relies on multi-source geological data to construct a geological profile and a three-dimensional structural preliminary model as the basic constraint framework for voxel grid division. The multi-source geological data includes at least borehole data, geological radar images, and seismic reflection layer information. Borehole data provides vertical profile information such as stratum distribution, lithologic boundaries, and groundwater levels, and is the data source with the highest depth accuracy in geological modeling; geological radar images are mainly used to detect shallow geological structures and local anomaly areas. They have high resolution but limited penetration depth and are suitable for filling in blank areas in profiles; seismic reflection layer information can reveal faults, structural interfaces, and interlayer discontinuities, and is suitable for macroscopic stratigraphic trend analysis and fault boundary extraction.

[0033] During the data fusion process, the borehole sampling points are first spatially located, and their profile data is standardized to extract parameters such as stratigraphic depth, lithology, and porosity. Radar images are then compared laterally with seismic data, and stratigraphic extension surfaces are spatially established using image recognition and horizon tracing techniques. These surfaces are then calibrated against existing geological maps to form a preliminary set of geological profiles. Subsequently, spatial reconstruction algorithms such as finite-difference interpolation, spline fitting, or bidirectionally constrained grid interpolation are used to construct a preliminary 3D geological structural model that reflects the actual geological boundary trends and structural morphology. This preliminary model is not used directly for analysis, but rather serves as a basis for determining boundary locations and grid density during subsequent voxel decomposition.

[0034] Based on the initial 3D geological structure model, the present invention further identifies key areas of high heterogeneity within the model, including fault zones, lithologic transition zones, and tectonic boundaries. Local mesh refinement strategies are implemented accordingly. Fault zones are identified based on the intersection of seismic reflection anomalies, lithologic transition boundaries, and radar high-reflectivity zones, using a boundary line extraction algorithm to construct fracture boundary surfaces. Lithologic transition zones are located based on lithologic number variation trends. Tectonic boundaries are extracted using interlayer angle variations and a mechanical interface recognition module. For these areas, the present invention does not use uniform voxel size. Instead, based on the identification results, a higher spatial partitioning density is applied within the affected area (typically defined as 3–5 times the voxel distance on either side of the boundary). This ensures that each complex boundary segment is encapsulated by finer-grained voxel units, improving representation. Non-heterogeneous regions maintain the same base partitioning density, resulting in an overall irregular grid structure with controllable density zoning.

[0035] After completing the initial density partitioning based on structural recognition, to further optimize the grid layout and avoid unnecessary high-density computational overhead, this paper introduces an information entropy-driven dynamic density control mechanism, combining multi-source data to quantitatively assess geological complexity. First, the 3D modeling space is divided into several basic evaluation sub-blocks (e.g., 20m×20m×10m), and the information entropy of the geological data within each sub-block is calculated. The information entropy index is in the following form:

[0036]

[0037] in, Indicates that in this sub-block The proportion of lithologic or structural types, is the total number of lithologic categories. A higher information entropy indicates a more complex geological structure; a lower information entropy indicates a single dominant lithology and a simpler geological structure. Based on this, the grid density is maintained or even increased in areas with information entropy above a set threshold, while the grid is appropriately coarsened in areas with lower entropy. This achieves a dynamic match between the complexity of the geological information structure and the voxel resolution, significantly improving overall modeling efficiency and consistency of geological representation.

[0038] After completing the above-mentioned density control, the entire three-dimensional geological space will be divided into a set of voxel units with spatial differentiation. To ensure that each voxel not only has geometric properties but also possesses geological and structural identification capabilities, the present invention adds a set of structural identification information to each voxel node. This information includes at least two components: one is the geological unit number, which is used to identify geological information such as the stratigraphic type, lithologic code, and whether it crosses a fault zone; the other is the spatial topological label, which is used to describe the connection relationship between the voxel and its adjacent voxels in three-dimensional space. The topological label not only indicates whether the voxel is coplanar with the surrounding voxels in the six main directions (front and back, left and right, and up and down) and whether there are any boundary mutations, but also records its boundary orientation, contact angle, and the presence of structural discontinuities, facilitating the accurate assessment of disturbance propagation paths and coupling areas in subsequent risk hotspot analysis.

[0039] After all the above processing steps are completed, a 3D geological voxel model with spatial topological constraints can be constructed. This model has the following core capabilities: (1) it supports high-resolution representation of heterogeneous structures in underground space, including difficult-to-construct objects such as fractures, interlayers, and lithologic transitions; and (2) it has the ability to identify spatial structures, providing a basic topological connection graph structure for performing irregular spatial convolution operations in subsequent neural network models.

[0040] Step S102: Based on the three-dimensional geological voxel model, extract the time series characteristic indicators under the influence of construction disturbance, including local pressure gradient change, seepage velocity estimation and structural response heterogeneity factor, to form a continuous time series characteristic vector.

[0041] In step S102, the system uses the 3D geological voxel model constructed in step S101 as its basic input to further extract the dynamic response characteristics of geological units under underground construction disturbance conditions. These response characteristics are converted into structured vector data with temporal continuity for subsequent risk prediction modeling. Specifically, the core of this step is to establish a mapping relationship between geological structural changes and construction impact responses, thereby forming a time series feature vector that can be processed by the neural network.

[0042] First, the impact range and temporal resolution of the construction disturbance must be determined. Based on the current construction process (such as shield tunneling, tunnel excavation, and foundation pit support) and its positional relationship within the 3D model, a target voxel set within the construction impact area is determined. Within this area, the system collects geological parameters related to the disturbance response and continuously monitors or numerically simulates them, generating a time-varying parameter sequence.

[0043] In order to achieve multi-dimensional extraction of response characteristics, the following main characteristic quantities need to be calculated for each target voxel: First, the local pressure gradient change value, which is usually calculated based on physical simulation models or sensor data. By solving the pore water pressure change rate between adjacent voxels, it reflects the transient stress disturbance of the construction on the stratum structure; second, the estimated seepage velocity value. Combined with the groundwater seepage model and the permeability coefficient distribution, the fluid movement trend in the voxel at different times is simulated and calculated to reflect the spatial migration of hydraulic conditions during construction; third, the structural response heterogeneity factor, which is used to measure the degree of response difference between adjacent voxels under disturbance. It can be obtained by calculating the elastic modulus and shear modulus change trend of the geological body and the deformation gradient of adjacent units, and is often described by a normalized tensor.

[0044] To ensure that these features can be effectively recognized and modeled by subsequent neural networks, they need to be uniformly encoded as a vector structure. Specifically, for each target voxel node, at each sampling time step, the three indicators are combined to form a multidimensional feature subvector. Over the entire perturbation analysis cycle, the subvectors from multiple time steps are concatenated to form a complete time series feature vector. This vector not only preserves the temporal characteristics but also captures the nonlinear variation of the geological response as the perturbation evolves.

[0045] During processing, the system can further introduce data enhancement and compression techniques such as differential encoding, sliding window normalization, and Z-score standardization to improve feature stability and expressiveness. Furthermore, for regions with clear structural coupling between voxels, the system can construct spatial position embedding information, encoding the topological relationships of voxels in three-dimensional space as additional spatial index features, thereby enhancing the spatial resolution of the feature vector.

[0046] Ultimately, step S102 outputs a set of continuous time series feature vectors that can be fed into a deep neural network for inference modeling. This set structurally corresponds to multiple voxel nodes within the construction-affected area of ​​the 3D geological model. Semantically, it captures the dynamic response and evolution of geological parameters under the influence of construction disturbances. Its temporal, spatial, and multimodal nature provides sufficient and reliable input data for subsequent nonlinear risk heat prediction.

[0047] Therefore, this step not only completes the transition from static geological modeling to dynamic behavioral modeling, but also establishes a quantitative coupling mechanism between construction disturbances and geological responses, ensuring that the system can truly restore and learn the complex risk propagation paths during underground construction.

[0048] To accurately identify risk hotspots during underground construction, the 3D geological voxel model constructed in step S101 must be used to further extract temporal characteristic indicators for each voxel within the construction disturbance impact zone. The following uses shield construction as an example to illustrate how to use the model to calculate local pressure gradient changes, seepage velocity estimates, and structural response heterogeneity factors.

[0049] First, set: Each voxel unit in the model space is recorded as , whose center has position coordinates and dimensions are 1 m. The sampling time series is .

[0050] 1. Local pressure gradient changes

[0051] At each point in time , voxel The pore water pressure is recorded as . Select its six directly adjacent voxels (upper and lower, front and back, left and right), and the corresponding pressure values ​​are , the center distance is The local pressure gradient can be calculated using the following formula:

[0052]

[0053] in, : Voxel In time Pressure gradient (unit: )

[0054] is the number of adjacent voxels, up to 6

[0055] Voxel In time Pore ​​water pressure (unit: Pa)

[0056] Neighbor voxel In time Pore ​​water pressure (unit: Pa)

[0057] is the center distance, fixed at 1 m.

[0058] 2. Estimation of infiltration flow rate

[0059] Based on Darcy's law, voxels In time The groundwater flow velocity is estimated as:

[0060]

[0061] in, For groundwater in voxels Flow rate in (unit: );

[0062] Voxel Permeability coefficient (unit: ), determined by its lithology;

[0063] is the groundwater dynamic viscosity (the standard value is )

[0064] : The voxel at time Pressure gradient

[0065] For example, if the sand layer voxel ,and ,but:

[0066]

[0067] Indicates that the direction of water flow is opposite to the direction of pressure gradient.

[0068] 3. Structural Response Heterogeneity Factor

[0069] Voxel response includes three physical quantities: displacement (Unit: m), shear strain (dimensionless), volumetric strain (dimensionless). Select adjacent voxels The corresponding physical quantity is , then the construction response heterogeneity factor is defined as:

[0070]

[0071] in, Voxel The structural response heterogeneity factor of

[0072] Separate voxels In time Response;

[0073] is the adjacent voxel The larger the factor, the more concentrated the stress and severe the deformation in the area, and the higher the potential risk.

[0074] Through the above three indicators, each voxel can be Construct a complete time series feature vector:

[0075]

[0076] This vector can be directly input into the neural network for risk heat modeling, where 、 as well as Represents a time series.

[0077] Step S103: Input the continuous time series feature vector into the trained convolutional recurrent neural network model. The convolutional recurrent neural network model has a spatial attention aggregation mechanism and a deep memory unit, which is used to model the nonlinear risk propagation behavior in the construction evolution process and output the risk heat value of each spatial location.

[0078] In step S103, the system uses the continuous time series feature vectors extracted in step S102 as input and utilizes a trained deep neural network model to perform nonlinear modeling and inference on the risk status of each spatial location. The system then outputs a risk heat value for subsequent hotspot identification. The goal of this step is to establish a mapping relationship between the evolution of geological disturbances and potential risk levels, enabling high-resolution spatial risk prediction during dynamic construction processes.

[0079] In the specific implementation, we first need to construct the input feature matrix corresponding to the 3D geological voxel model. , which has been obtained in step S102 in a continuous time period Feature sequence within , the sequence is determined by the local pressure gradient , permeation flow rate and structural response heterogeneity The model consists of multidimensional physical quantities such as voxels, arranged in chronological order, forming a two-dimensional tensor structure with a length of T and a dimension of d (usually 3 or its extended form). The set of feature tensors of all voxels constitutes the batch input data of the model, with time series characteristics, spatial location indexing, and physical consistency.

[0080] To accurately capture the propagation paths and risk aggregation effects of construction disturbances in geological space, a deep neural network structure combining convolutional and recursive mechanisms was employed. The overall network architecture consists of several functional layers: First, a set of temporal convolutional layers extracts the local temporal evolution patterns of the feature vector within each voxel, such as periodic disturbance responses, stress wave propagation trends, or sudden seepage changes. The convolution kernel size can be set based on the sampling interval and construction cadence, typically taking 3 to 5 time steps.

[0081] This is followed by a set of spatial attention aggregation units. This module constructs a dynamic correlation matrix between each voxel and its spatial neighbors by incorporating the adjacency topology between geological voxels. Based on the activation value of each voxel's current input sequence and the states of its neighboring voxels, the system calculates the attention weight distribution for each voxel, automatically focusing on regions that are highly coupled to its geological response at the current moment, thereby enhancing the ability to identify geological structural heterogeneity. This attention mechanism is typically implemented as dot-product attention or graph convolution, and softmax normalization can be used to ensure weight interpretability.

[0082] The network then introduces multiple stacked bidirectional recurrent neural network layers, using either LSTM or GRU architectures to handle long-term dependencies in the temporal dimension. This bidirectional architecture allows the system to simultaneously consider past construction disturbance history and future trend forecasts, enabling a more comprehensive modeling of stress diffusion and risk accumulation effects. The recurrent layers output a feature representation vector for each voxel node within the entire time series window.

[0083] Finally, the network projects this high-dimensional spatiotemporal feature vector into a scalar through a set of fully connected mapping layers: the risk heat value corresponding to the current voxel at the current construction moment. This heat value is normalized to the range [0, 1], representing its relative risk level. Values ​​closer to 1 indicate a higher likelihood of being in a high-risk area, while values ​​closer to 0 indicate a lower risk area under construction disturbance conditions.

[0084] During the model training phase, historical construction data or simulated disturbance data are used as training samples. The objective function can use a mean squared error loss function or a cross-entropy loss function based on actual disaster labels. Backpropagation and a gradient descent optimizer are combined for iterative parameter updates. If true labeled data is unavailable, a semi-supervised training strategy can be employed, or unsupervised heat clustering can be used as pseudo-labels.

[0085] To improve inference efficiency, the system uses a regional partitioning strategy to perform forward propagation operations only on voxels within the current construction impact radius during each inference process, thereby reducing the computing resources required for full-map modeling. Inference results are output as a spatial point cloud, with each voxel node assigned a risk heat value and its corresponding spatial coordinates retained for subsequent construction of hot zone maps and decision support.

[0086] In summary, step S103 not only transforms the time series characteristics of geological disturbance response into spatially explicit risk assessment results, but also realizes intelligent perception and structured modeling of the underground construction risk diffusion process through the attention mechanism and recursive structure. It has the technical advantages of strong feasibility, good scalability, and high prediction accuracy, and constitutes the core technical link of this method for realizing dynamic risk identification.

[0087] Furthermore, the convolutional recursive neural network model includes a temporal geological disturbance feature extraction layer, a spatial topology perception convolutional encoding layer, an attention regulation recursive modeling layer, and a heat regression decoding layer;

[0088] The time-series geological disturbance feature extraction layer is used to receive continuous time series feature vectors extracted based on the three-dimensional geological voxel model, and the input is a multi-channel time series tensor containing pressure gradient, seepage velocity and structural heterogeneity factor. It uses a one-dimensional time series convolution method to realize dynamic pattern recognition across time windows and output the local disturbance response spectrum corresponding to each voxel.

[0089] The spatial topology-aware convolutional coding layer is used to integrate the geological structural adjacency relationship between voxel nodes. The input is the disturbance response spectrum and the topological graph structure tensor. The topological embedding coding method is combined with the three-dimensional irregular convolution operator to achieve modeling of spatially irregular geological connection relationships. The output is a topological convolution feature map. The topological graph structure tensor is constructed by the spatial adjacency relationship and the geological structural boundary relationship between voxels in the three-dimensional geological voxel model. It is used to clarify the topological connection mode of each voxel node and serves as the adjacency input in the spatial convolution operation.

[0090] The attention control recursive modeling layer is used to receive the topological convolution feature map, combine the historical disturbance evolution state, and construct a multi-head spatial attention gating mechanism and a bidirectional recurrent unit to dynamically adjust the risk dependency weight of each region voxel in the multi-step prediction process, and output the dynamic state vector of the voxel node in the full-time risk perception space;

[0091] The heat regression decoding layer performs nonlinear dimensionality reduction and regression transformation on the dynamic state vector to generate a risk heat value corresponding to each spatial voxel position. The output is a normalized floating-point number, which is used as quantitative input for hot zone identification and auxiliary decision-making.

[0092] The convolutional recurrent neural network model consists of four sequentially connected functional layers. The model input is a time-series feature vector extracted from a three-dimensional geological voxel model. This vector contains the local pressure gradient changes, seepage velocity estimates, and structural response heterogeneity factors for each construction-affected voxel unit over consecutive time steps. This data exhibits temporal continuity and physical coupling. First, the model feeds these multi-channel, equally spaced time-series tensors into a time-series geological disturbance feature extraction layer. This layer employs a one-dimensional convolutional neural network architecture, sliding a window along the time axis to extract short-term dynamic response patterns. Each convolution kernel identifies features such as the disturbance amplitude trend, frequency oscillations, and mutation points over several consecutive time steps. The extracted results are uniformly mapped into a set of disturbance response spectra, which characterize the short-term dynamic behavior of each voxel node under construction disturbance. The output tensor maintains spatial consistency but is expanded to include multiple time-series convolution channels, forming a local disturbance encoding representation for each voxel node.

[0093] Next, these perturbation response spectra are fed into the spatial topology-aware convolutional encoding layer. At this stage, to more realistically reflect the irregular connectivity structure between subsurface voxels, the model incorporates the adjacency relationship between 3D geological voxels as topological prior knowledge. A topological graph structure tensor is extracted from the 3D voxel model constructed in step S101. This tensor explicitly indicates whether each voxel is adjacent to its spatial neighbors and whether there are geological boundaries or fault interfaces. This tensor, input to the model in the form of a sparse adjacency matrix or adjacency index, serves as a structural constraint for the spatial convolution operation. Building on this, the encoding layer employs a 3D irregular convolution operator to incorporate voxel nodes with geological connectivity within the spatial neighborhood into the receptive field and aggregate topological features while maintaining spatial structure stability. The weight sharing mechanism in the encoding process, combined with the topological embedding structure, enables the model to accurately understand the discontinuous interfaces, stratigraphic changes, and structural irregularities of the geological volume. The output is a topological convolution feature map, a high-dimensional spatial representation of each voxel node after considering its geological connectivity.

[0094] After obtaining a spatial feature map with topological semantics, the model enters the attention-regulated recursive modeling phase. This phase is used to model the nonlinear evolution of risk propagation over time. The model employs a bidirectional recurrent structure (such as a Bi-GRU or Bi-LSTM) to simultaneously learn the impact of construction disturbances on risk state evolution in both the forward and reverse directions. Furthermore, to emphasize the information representation capabilities of key areas and time periods, the model introduces a multi-head spatial attention gating mechanism at this layer. By dynamically adjusting the attention weight of each voxel node in the historical state sequence, it emphasizes the most responsive disturbance signal sources and suppresses redundant inputs. This mechanism not only achieves dynamic focusing in the temporal dimension but also enables regional selection in the spatial dimension, thereby forming a risk dynamic state vector that better aligns with geological distribution patterns.

[0095] Finally, the model projects these high-dimensional state vectors into final risk heat values ​​through the heat regression decoding layer. This layer, composed of several fully connected layers and nonlinear activation units, is responsible for reducing the dimensionality of the recursive output and converting it into a normalized risk heat value. The value range is typically limited to 0 to 1, which serves as quantitative input in subsequent hotspot identification modules or auxiliary decision optimization systems.

[0096] Through the cascaded execution of these four layers, the model integrates multi-layer feature fusion from time series disturbance signals with spatial topology, ultimately outputting a stable and interpretable risk heat value. This model's structural design is customized for the spatiotemporal risk coupling characteristics of underground construction. Unlike the generic structures of traditional time series prediction networks or graph neural networks, its input and output paths are clear, constraints are well-defined, and it is both deployable and trainable.

[0097] Furthermore, the time series geological disturbance feature extraction layer is also used to:

[0098] The continuous time series feature vectors of each voxel node are normalized, and a multi-channel normalization mapping mechanism is used to map the pressure gradient change, infiltration velocity estimation, and structural heterogeneity factor to comparable scales in the interval [–1, 1] to ensure that the response weights of different physical quantities in the convolution extraction are balanced.

[0099] A one-dimensional variable convolution kernel structure across time windows is used to simultaneously model disturbance variation patterns within fixed time steps and adaptive time windows. The variable convolution kernel size is dynamically determined by the current construction disturbance activation index. This allows the model to automatically enhance its sensitivity to short-term features during sudden disturbances and improve its ability to extract long-term trends during stable periods.

[0100] The maximum pooling and residual connection structure in the time direction are introduced to the convolution output to enhance the robustness of local patterns and retain the long-term information in the original input features. This enables the output disturbance response spectrum to have both the ability to detect short-term disturbances and maintain long-term change trends, adapting to the semantic alignment requirements of subsequent spatial topology modeling.

[0101] In the underground construction decision-making support method of the present invention, the primary function of the temporal geological disturbance feature extraction layer is to dynamically model the temporal evolution of different geophysical parameters, thereby providing a disturbance response spectrum with expressive power and semantic integrity for subsequent spatial structure perception and risk propagation modeling. In actual construction scenarios, underground structures typically exhibit complex response characteristics characterized by nonlinearity, nonstationarity, and multi-scale coupling. Construction disturbances such as tunneling, blasting, drainage, and grouting can cause pressure redistribution in the underground medium, changes in seepage paths, and disturbances in structural stability. Each spatial location within a geological voxel (i.e., a voxel node) may exhibit significantly different dynamic response trajectories under these disturbances. Therefore, the present invention designs a temporal feature extraction method with multi-channel normalization, a dynamic convolution kernel mechanism, and an information-preserving structure to capture these disturbance evolution characteristics with high resolution.

[0102] First, for each voxel node, it is necessary to input its continuous time series feature vector during the construction process. This vector is composed of multiple physical attributes, mainly including local pressure gradient changes, seepage velocity estimates, and structural response heterogeneity factors. These attributes reflect information sources from multiple dimensions, such as changes in the ground stress field caused by construction, pore water seepage disturbances, and structural continuity disturbances. However, the original physical quantities of different attributes usually have different dimensions and inconsistent numerical ranges. Directly inputting them into the model will cause a certain physical quantity to dominate the network parameter update during training, thereby damaging the overall learning ability and physical interpretability of the model. Therefore, these features need to be standardized before entering the convolutional neural network.

[0103] Normalization utilizes a multi-channel normalization mapping mechanism. Specifically, the maximum and minimum values ​​of each type of physical quantity across all voxel nodes and all time steps are calculated separately, and the physical quantity is uniformly mapped to the interval [–1, 1]. This mapping can be performed using linear transformations or nonlinear mapping functions such as tanh to suppress the impact of outliers on model stability. This normalization not only unifies the scales of multiple source physical quantities but also provides a unified numerical foundation for subsequent feature fusion and convolution calculations. The multi-channel mechanism is reflected in the fact that each physical quantity is treated as a separate input channel and fed into the network in parallel without mixing to avoid semantic ambiguity. Furthermore, the structure of the normalized tensor maintains a three-dimensional form, with each voxel node having an independent time series, with dimensions T × C × N, where T is the time length, C is the number of channels (i.e., the number of physical quantity types), and N is the total number of voxel nodes.

[0104] After completing the standardization, the system enters the convolution operation stage. In the present invention, the convolution structure does not use a one-dimensional convolution kernel with a fixed window size, but introduces a one-dimensional variable convolution kernel structure across time windows to simultaneously model short-term emergencies and long-term trend evolution patterns under construction disturbances. The core design idea of ​​this mechanism is that the impact pattern of construction disturbances varies due to their type, intensity and duration. Strong disturbance events such as blasting or sudden water gushing usually have high frequency, short period and rapid change characteristics, while slow-changing processes such as excavation advancement and groundwater level drop show low frequency, long period and gentle evolution characteristics. Traditional fixed-length convolution kernels can only cover part of the dynamic patterns, which may lead to failure in extracting key disturbance patterns or loss of long-term dependent features.

[0105] To address these issues, the variable convolution kernel mechanism introduces a "disturbance activation index" as a kernel size control signal. This index is generated based on multiple factors, including the current construction phase, the rate of change in sensor feedback data, and sensitive areas predicted by the geological model. Higher values ​​indicate a high-intensity disturbance demand, and the convolution kernel automatically shrinks to enhance its responsiveness to rapidly changing signals. Conversely, when the disturbance intensity is low and the system is in a stable evolutionary phase, the convolution kernel length automatically increases to cover a longer window of historical data and identify potential trend characteristics. The convolution process is performed in a sliding window manner along the time dimension. Each convolution result is aligned with the corresponding voxel node, and the output remains in the form of a tensor of T′ × C′ × N, where T′ may be slightly smaller than T due to boundary effects, and C′ is the new number of convolution channels.

[0106] Since there are differences in the feature semantics corresponding to different convolution kernel lengths, in order to prevent information loss or noise amplification in the convolution operation, the present invention adds maximum pooling and residual connection structures in the time direction after the convolution output. The maximum pooling operation is used to enhance the prominence of local high-response patterns, and can strengthen the key disturbance point signals in the multi-channel output while suppressing background fluctuations. The pooling window length can be dynamically set according to the activation index. A short pooling window is used to extract spike signals in the high disturbance stage, and a long window is used to extract stable change patterns in the low disturbance stage. The role of the residual connection structure is to directly splice or weight the convolution input signal according to the channel dimension to the convolution output, avoid overfitting the model to high-frequency disturbances, and maintain the integrity of the long-term trend in the original time series input.

[0107] Through this combined structure, the disturbance response spectrum output by the extraction layer has the following advantages: first, the response spectrum of each voxel node not only retains its short-term disturbance activation characteristics, but also has the long-term evolution trend, which enables the model to have a stronger temporal understanding of risk evolution; second, the multi-channel input and output mechanism ensures the independent modeling and linkage explanatory power between geological attributes; finally, the entire structure can dynamically adjust the convolution kernel parameters, activation exponential function and pooling window size through error backpropagation during the training phase, thereby realizing adaptive dynamic modeling capabilities without the need for human intervention.

[0108] The resulting disturbance response spectrum is passed to the subsequent spatial topology-aware convolutional coding layer, maintaining high consistency in input dimensions, feature semantics, and spatial node indexing. This ensures seamless structural connectivity and semantic logical continuity within the multi-layer neural network. Each voxel response spectrum output is essentially a high-dimensional disturbance state vector, which can be further embedded in a spatial graph structure to construct propagation channels between irregular geological voxels, thereby reflecting the propagation dynamics of construction disturbances in risk heat modeling.

[0109] In summary, the feature extraction layer structurally integrates normalization, adaptive convolution, maximum pooling, and residual connections. Functionally, it realizes multi-scale dynamic modeling of construction disturbance responses. Semantically, it provides the underlying representation of disturbance responses required for spatial topological analysis. In terms of implementation, it has end-to-end deep trainability, far away from traditional static geological modeling or regular convolution processing paths, and is an important component of the overall architecture of the neural network of the present invention.

[0110] Furthermore, the spatial topology-aware convolutional coding layer is further used to:

[0111] Based on the topological structure tensor constructed in the three-dimensional geological voxel model, a geological structure adjacency matrix and a fault contact surface weight matrix are used to jointly construct a voxel graph representation, wherein each graph edge not only represents a geometric adjacency relationship but also carries a semantic label representing the contact interface type, including a fracture surface, a soft-hard interface, or a homolithic continuous surface;

[0112] The introduction of a projection mechanism for irregular convolution kernels on graph structures allows the convolution operation to adjust the propagation weight according to the attributes of different types of adjacent edges, thus achieving discriminative structural perception.

[0113] The topological embedding vector is fused in the convolution operation. This vector is generated by the geological structure encoding module based on the multi-order adjacency information of the voxel node, indicating the global position role of the node in the spatial structure, including whether it is in the intersection area, closed area, edge area or uniform area, so that the topological convolution feature map has the ability to express multi-scale geological structure context.

[0114] In complex underground construction scenarios, geological structures are often not regular, continuous, homogeneous systems, but rather heterogeneous spatial bodies composed of irregular geological elements such as fracture surfaces, interfaces of different lithologies, and soft-hard transition zones. In order to effectively identify and analyze the nonlinear risk transfer behavior caused by construction disturbances in these structures, spatial topological relationships must be incorporated into the modeling process so that the deep learning model can not only extract local features, but also perceive the topological roles and structural semantics of different regions in space based on the overall layout of the geological structure. Therefore, in the present invention, the spatial topology-aware convolutional coding layer is not only a convolutional feature extraction structure, but also a neural network component that performs structural recognition, attribute transfer, and context fusion on irregular geological voxel graph structures. The core goal of this layer is to achieve deep fusion of geological structure semantics and topological features in the voxel graph, and encode them into high-dimensional feature representations that can be used for subsequent recursive modeling and heat prediction.

[0115] Specifically, the convolutional coding layer first relies on the input 3D geological voxel model, building a topological graph structure tensor based on it. This tensor is a structural representation of spatial topological information, with voxel nodes as graph vertices, and edges between nodes representing adjacent spatial or geological structural connections. Unlike traditional graph neural networks, which define adjacency relationships solely by Euclidean distance or geometric adjacency, the present invention further introduces a geological structural adjacency matrix and a fault contact surface weight matrix to form a more refined graph structure representation with geological semantics. In this structure, each edge not only reflects the geometric contact between two voxels in 3D space but also comes with a semantic label to distinguish whether these adjacencies are due to fracture surface contact, soft-hard lithology boundary, or continuous extension of the same lithology. These semantic labels are typically encoded discretely. For example, a fracture surface can be represented by the label "F", a soft-hard boundary by "SH", and the same lithology by "LL". These labels are used as edge attributes in the graph for subsequent calculations.

[0116] This semantic labeling mechanism significantly improves the model's sensitivity to behavioral differences across different geological regions. For example, voxels on opposite sides of a fracture surface, while potentially geometrically close, have structural discontinuities that lead to distinct disturbance propagation relationships compared to voxels in continuous rock formations. By labeling these edge types and introducing a differentiated weighting mechanism, the convolution operation can dynamically adjust the propagation strength based on the semantics of the edges, thereby achieving structure-aware feature propagation. This mechanism is implemented as a projection operation of irregular convolution kernels onto a graph structure. Traditional convolution kernels are typically applied to regular grids or fixed-size neighborhoods, while irregular convolution kernels can dynamically select or weight different convolution paths based on the actual adjacent edge types in the voxel graph. For example, during a convolution operation, for a given node, the model can set the propagation weight for fracture edges to 0.2, the propagation weight for continuous rock formation edges to 1.0, and the propagation weight for hard-soft boundary edges to 0.5, thereby simulating the differences in the propagation efficiency of actual physical disturbances across different structures.

[0117] While achieving this, the present invention further proposes the concept of a topological embedding vector to enhance the voxel node's understanding of its global role in the entire graph structure. This embedding vector is generated by a geological structure encoding module, which first performs a multi-order adjacency scan on each voxel node to extract its structural statistical information within the k-order neighborhood, such as the stratigraphic category distribution of adjacent nodes, the frequency of fault contacts, and the lithologic variation gradient. These statistical features are then mapped into an embedding vector of fixed dimension, which serves as the "position semantic label" of the node. This topological embedding is not a simple superposition of geometric coordinates or label encodings, but rather a spatial contextual representation formed by combining the node's relative position in the global structure, boundary characteristics, and structural variation trends.

[0118] This vector can be used to characterize the global positional role of a voxel in the spatial structure. Common roles include intersection areas (intersections of multiple faults, where disturbances are concentrated), closed areas (surrounded by structural boundaries, with limited propagation), edge areas (near the stratum termination line), and uniform areas (where structural changes are small and disturbances propagate uniformly). This role information is combined with convolution kernel weights or attention mechanisms in subsequent convolutional feature propagation to adjust feature fusion strategies. For example, the model can use a multi-channel high-pass convolution strategy for voxels in the intersection area to extract complex interference signals, introduce information-preserving paths for voxels in the closed area, use a boundary enforcement mechanism for the edge area, and perform standard convolution aggregation operations on the uniform area.

[0119] After completing the above operations, the convolutional coding layer ultimately outputs a topological convolutional feature map, in which each voxel node is represented as a multidimensional feature vector that simultaneously encodes information such as disturbance response, structural adjacency semantics, and topological role embedding. These features not only demonstrate spatial coherence but also reflect the semantic location of each voxel within the geological structure map, providing excellent contextual interpretation and risk transmission logic. Because these features are constructed and fused within the graph structure, they possess strong interference resistance and structural adaptability, enabling the subsequent recursive modeling module to model and predict disturbance risk over time.

[0120] Furthermore, the entire process does not rely on external graph-building tools. Instead, topological information is directly derived from the voxel model itself, ensuring that its structural representation is consistent with the actual geological data. Edge attributes between voxels are automatically analyzed during the initial modeling process and generated simultaneously during the construction of the geological voxel model using a structural boundary recognition algorithm. Semantic labels can also be manually annotated by geological experts or generated with the assistance of machine learning algorithms, achieving a closed-loop modeling process.

[0121] In summary, the spatial topology-aware convolutional coding layer is not only structurally different from the traditional regular convolutional layer, but also constructs a graph convolution module with multi-scale structure perception capabilities through key technical means such as topological graph tensor construction, introduction of edge attribute semantic labels, irregular convolution kernel projection mechanism and topological embedding vector fusion. It provides high-order feature input with complete structural semantics and real disturbance propagation for subsequent dynamic modeling and decision analysis.

[0122] Furthermore, the attention-regulated recursive modeling layer is further used to:

[0123] A multi-head spatial attention mechanism is introduced to construct three types of attention weight vectors based on spatial position, topological role, and perturbation intensity at each time step. The final comprehensive attention map is obtained through linear weighted aggregation to adjust the information channel distribution of input features in the recursive structure.

[0124] A bidirectional gated recursive structure is used to propagate the disturbance response state bidirectionally along the time dimension with voxel nodes as the unit. The trend of the disturbance evolving from the edge of the hot zone to the core is captured in the forward propagation, and the disturbance convergence or disappearance signal is supplemented in the backward propagation, thereby improving the modeling ability of nonlinear disturbance paths.

[0125] A perturbation state evolution encoder is added to the recursive output stage to record the heat change pattern of each voxel in the entire time series window, including statistical features such as the maximum fluctuation point, continuous perturbation interval, and change slope distribution. The evolution code is then concatenated with the recursive state vector and used as the dynamic state output to enhance the interpretability and prediction accuracy of subsequent heat regression.

[0126] During underground construction, the evolution of geological disturbances is highly nonlinear and regionally heterogeneous, and its propagation path is not only limited by the adjacency of physical locations, but also affected by the complex geological structure, changes in material properties, and the intensity of the disturbance source. Therefore, relying solely on traditional convolution or time series processing structures is often unable to fully model the actual propagation behavior of disturbance risks in three-dimensional space. The present invention sets up an attention-regulated recursive modeling structure in the neural network architecture. Its core design goal is to combine the spatial topological characteristics and dynamic change patterns of construction disturbances, dynamically adjust the information transmission method of the characteristics in the time dimension, and achieve accurate characterization of the disturbance risk transmission path and response intensity, thereby providing a more physically meaningful and explanatory dynamic state representation for the final heat regression.

[0127] In this architecture, each voxel node has a set of perturbation response vectors along the time axis. These vectors, output by the previous feature extraction and topological convolution modules, encode the node's perturbation intensity, structural properties, and role in the geological map structure at each time step. To fully utilize this information, the present invention introduces a multi-head spatial attention mechanism to construct a multidimensional attention map at each time step. This attention mechanism differs from the traditional unified attention weight and instead divides the sources of attention into three complementary dimensions based on the characteristics of the geological structure: spatial position, topological role, and perturbation intensity.

[0128] Spatial position attention refers to the geometric importance weight assigned to voxels based on their position in three-dimensional space at each time step. Generally speaking, voxels close to the boundaries of identified high-risk areas should have higher attention weights because these areas are more susceptible to the spread of disturbances. Topological role attention is based on the structural topological embedding defined in the aforementioned geological voxel model to determine whether the voxel is in a special location such as a fault intersection, structural boundary area, or lithologic transition zone. At these locations, disturbance behavior often exhibits asymmetric or unstable behavior, so it is necessary to enhance the information transmission capability of the recursive channel by increasing its attention coefficient. Disturbance intensity attention dynamically generates weights based on the response amplitude or gradient at a certain time step in the disturbance response spectrum to ensure that high-fluctuation areas or rapidly rising segments can be focused on by the model.

[0129] These three types of attention weights are calculated using attention heads and aggregated linearly to generate a final integrated attention map, which is used to adjust the channel weight distribution of input features in the gated recurrent structure at the current time step. This attention map is automatically updated through backpropagation during model training. Its design aims to guide the model to focus on key sources of perturbations, filter redundant background information, and optimize information pathway selection through multi-dimensional fusion, thereby improving the spatial resolution and semantic consistency of the overall model.

[0130] In terms of recursive structure, the model adopts a bidirectional gated recursive mechanism to enhance the two-way perception of the causes and effects of disturbance propagation. Specifically, each voxel node receives disturbance transmission signals from both the forward and reverse directions in the time dimension. Forward propagation is used to simulate the evolution of disturbances from the boundary into the core of the hot zone. For example, during shield tunneling or grouting construction, disturbances initially appear in front of the construction and gradually spread to the surrounding strata; backward propagation is used to capture the downward trend of disturbances after the peak heat or the self-stabilization process of the structure, such as stress release, hydraulic dissipation, or response attenuation caused by material relaxation. This bidirectional structure ensures that the model can not only predict how disturbances are generated and aggregated, but also understand when disturbances begin to subside, thereby improving the comprehensive perception of dynamic changes in heat.

[0131] Within the gated unit, the model introduces update and reset gates to control the degree of state retention and the magnitude of information updates at each moment, thereby avoiding the vanishing gradient problem in long-sequence modeling. This recursive unit structurally combines the advantages of both LSTM and GRU, offering high computational efficiency and strong semantic preservation, making it suitable for processing continuous but highly nonlinear time series such as underground disturbances.

[0132] In order to further improve the ability to express the evolution characteristics of disturbance risk, the present invention also introduces a disturbance state evolution encoder in the recursive output stage. This encoder runs independently of the recursive process and is specifically used to analyze the heat change trajectory of each voxel in the entire time series window. It receives a complete disturbance response sequence and calculates statistical features including the position of the maximum heat fluctuation point, the disturbance duration interval, the heat change slope distribution, the rise and fall time ratio, etc. These features constitute a fixed-length evolutionary coding vector, which represents the disturbance history pattern of the voxel node in the observation window. By splicing the evolutionary code with the recursive state vector, a richer dynamic state representation can be formed, so that the model output not only contains the state memory at the current moment, but also integrates the full picture of the disturbance behavior of the entire time period.

[0133] The design advantage of the evolutionary encoder is that it significantly enhances the physical interpretability of the model output. For example, when the heat of a voxel node rises sharply in a short period of time and then drops rapidly, the evolutionary code will show a "spike-shaped" disturbance pattern, which may indicate a high-energy short-term disturbance, such as instantaneous deformation caused by blasting. On the other hand, if the evolution process shows a slow increase followed by a long-term high level, it may correspond to persistent instability caused by water pressure accumulation or stress migration. Combining these disturbance pattern encodings with recursive states not only improves the accuracy of subsequent heat predictions, but also provides a clear classification basis for identifying hot zone boundaries and adjusting construction strategies.

[0134] The final dynamic state vector output by the model possesses both a recursive memory structure and an evolving statistical representation, serving directly as the input for the next stage of heat regression decoding. By integrating multi-layered features such as disturbance amplitude, propagation path, temporal changes, and structural semantics, the model can predict future risk heat levels with greater accuracy and accurately depict the spatial distribution of hotspots.

[0135] The entire attention-controlled recursive modeling process has a complete implementation path and logical chain. During the training phase, the model continuously optimizes attention weights, recursive structure parameters, and projection vectors in the evolutionary encoder through backpropagation of heat value prediction errors, enabling it to automatically adapt to the coupling characteristics of different types of construction disturbances and geological structures. During the deployment phase, the module has excellent generalization capabilities and can adapt to dynamic risk perception tasks in different sites, construction methods, and disturbance patterns.

[0136] This structure breaks through the limitations of traditional unidirectional RNN or fixed attention mechanisms in its design. By combining multi-head attention with bidirectional recursion, it enhances the multidimensional modeling capability of disturbance perception. By introducing the evolutionary coding module, the time series modeling results have the interpretability of disturbance behavior. It is an indispensable key module for realizing intelligent hot spot identification and decision optimization in underground construction.

[0137] Step S104: Based on the distribution characteristics of risk heat values ​​in three-dimensional space, high-risk hot spots that are continuous and dynamically changing in space are identified through heat gradient clustering and neighborhood structure consistency screening algorithms to form a dynamic risk hot spot map.

[0138] In step S104, based on the risk heat values ​​of each voxel in the three-dimensional space output from step S103, the system further identifies high-risk areas that are spatially continuous, structurally stable, and exhibit a clustering trend over time. To achieve this goal, this step introduces a heat gradient clustering algorithm and a neighborhood structure consistency screening mechanism to construct a three-dimensional risk hotspot map that can be dynamically updated during the construction process, and provides a specific, implementable algorithm path.

[0139] First, for the 3D voxel set within the entire construction impact range, the system calculates the risk heat value of each voxel node. and its three-dimensional coordinates in the geological model Binding constructs a risk field with spatial information. The risk heat value comes from the neural network prediction result in step S103 and has been normalized to a continuous value between 0 and 1 to express the risk level.

[0140] The system then performs the first step of processing, which is spatial gradient clustering based on risk heat values. This process is achieved by first sorting all voxels in descending order of heat values ​​and extracting those that exceed the set threshold. The voxels are used as the initial candidate point set, and the threshold is usually set to 0.7. , take it as the center, perform a three-dimensional local clustering operation, and check whether the risk heat of its 26 neighboring voxels (i.e. all directly contacting voxels in the three-dimensional direction) is also higher than ,in The value can be 0.1. If more than half of the adjacent voxels meet the condition, it is considered as the hot zone growth core.

[0141] Based on the above judgment, each hotspot core can be expanded through a simple connectivity growth algorithm. For example, the breadth-first clustering logic described by the following code can be used:

[0142] def grow_hot_zone(seed_voxel, threshold, delta):

[0143] queue = [seed_voxel]

[0144] visited = set()

[0145] hot_zone = set()

[0146] while queue:

[0147] current = queue.pop(0)

[0148] if current in visited:

[0149] continue

[0150] visited.add(current)

[0151] R_current = get_risk_value(current)

[0152] if R_current<threshold:

[0153] continue

[0154] hot_zone.add(current)

[0155] for neighbor in get_26_neighbors(current):

[0156] if neighbor not in visited:

[0157] R_neighbor = get_risk_value(neighbor)

[0158] if R_neighbor>= threshold - delta:

[0159] queue.append(neighbor)

[0160] return hot_zone

[0161] The core function of this code is to implement a connectivity identification algorithm for spatial hot spots. Its processing logic is based on breadth-first search (BFS). The purpose is to identify a whole high-risk connected area with a similar heat value and spatially adjacent to a known high-risk core voxel in a three-dimensional geological voxel model, thereby forming a hot spot object in the construction risk hot spot map.

[0162] Specifically, the algorithm first uses the input seed voxel (that is, the voxel initially determined to be high-risk) as the starting point for expansion and places it into a queue for subsequent processing. The system maintains two sets: a visited set, which records all processed voxels to avoid repeated visits; and a hot zone set, which stores all voxel nodes ultimately identified as belonging to the current hot zone.

[0163] Each time the algorithm takes a voxel node from the queue, it first checks whether it has been visited. If so, it skips it; otherwise, it marks it as visited. Then, the system calls get_risk_value(current) to get the risk heat value of the voxel. , and compares it with the set heat threshold threshold. If the voxel's heat value is less than the threshold, it indicates that its risk level is not high enough to be included in the hot zone determination range and is skipped directly. If the heat value is greater than or equal to the threshold, it is added to the current hot zone set.

[0164] The system then retrieves the voxel's 26 adjacent voxels in three-dimensional space—namely, those in the upper, lower, front, back, left, and right directions, as well as in each diagonal direction—and determines whether their risk heat values ​​are greater than or equal to threshold minus delta. Delta is a buffer tolerance used to smooth the hotspot boundary, allowing the hotspot to expand slightly at the edge to areas slightly below the threshold to improve connectivity and the naturalness of the hotspot structure. If a neighboring voxel has not been visited and meets the heat requirement, it is added to the processing queue for subsequent expansion.

[0165] The algorithm loop continues until the queue is empty, indicating that all high-heat voxels connected to the hot zone core have been identified. The hot_zone set finally returned is the complete 3D space voxel set of the current connected hot zone.

[0166] The get_risk_value() function returns the risk heat value of a voxel node, while get_26_neighbors() provides the indices of its 26 neighboring voxels. Each execution of this function generates a complete hotspot. The system executes this function for all seed points, merging intersecting regions to form multiple connected hotspot clusters.

[0167] After completing the initial hotspot clustering, it is necessary to perform spatial structure consistency screening. This step is to avoid "noise points" caused by local prediction errors from being mistakenly identified as hotspots. The specific method is to calculate the number of voxels in each candidate hotspot , spatial extension rate (i.e. maximum boundary length) and average thermal gradient. If the number of voxels in a region is lower than the minimum hot zone size threshold If the average thermal gradient changes too dramatically (for example, more than 0.5), the spatial structure of the region is considered unstable and should be removed. The thermal gradient can be estimated by the following formula:

[0168]

[0169] in is the hot zone voxel set, For nodes The risk heat value of The formula is used to calculate the average maximum thermal gradient of all voxel nodes in a 3D hotspot, which is used to measure the smoothness of the thermal distribution within the hotspot and thus determine its spatial consistency. Its physical meaning is: within each hotspot, the maximum change in thermal value between each voxel point and its neighboring voxels is examined, and then the average of all voxels is taken to reflect the overall level of the maximum local thermal mutation within the hotspot.

[0170] Indicates the average thermal gradient amplitude of the current hot zone (the unit is the same as the thermal value, generally dimensionless). The larger the value, the more uneven the thermal value distribution within the hot zone, the more drastic the local changes, and there may be a risk of misidentification. The hotspot voxel set currently being analyzed is a set of adjacent, connected voxels with high heat values ​​in three-dimensional space. For example, a set of spatially continuous voxels obtained through a clustering algorithm. Indicates the total number of voxel nodes contained in the hot zone Z, which is a positive integer. For example, a hot zone may consist of 12 consecutive high-temperature voxels, then . Represents each voxel node in the hot zone Traverse, where is the node number. Representation and voxel A set of adjacent voxels. Usually a 26-neighborhood definition is used, which is the same as All voxels that are in contact in three-dimensional space include 26 possible neighboring voxels in the up and down, front and back, left and right, and diagonal directions.

[0171] Ultimately, the retained set of hot zone voxels is assigned a unique identifier and constructed into a structured "dynamic risk hot zone map", which not only contains the voxel composition, spatial boundaries, center position and average heat value of each hot zone, but also records its evolution trend over time (such as volume change rate, heat growth rate, etc.) for subsequent dynamic decision-making and hot zone level classification.

[0172] The entire calculation process of step S104 can be encapsulated and executed in a graph space analysis framework, supporting timing linkage with the main control system and visual interface output, and having real-time update capabilities.

[0173] Furthermore, based on the distribution characteristics of risk heat values ​​in three-dimensional space, the heat gradient clustering and neighborhood structure consistency screening algorithm are used to identify spatially continuous and dynamically changing high-risk hot spots, forming a dynamic risk hot spot map, including:

[0174] Based on the risk heat values, a risk heat change map is constructed, the risk heat value of each voxel at the current moment is subtracted from its risk heat history sequence in the previous time period, and the time change rate of the heat of each voxel is extracted. This change rate is used as an initial screening basis to identify a set of voxels with significant changes in heat;

[0175] The above voxel set is used as the seed voxel set. Based on the consistency of the heat change rate direction and the three-dimensional spatial adjacency relationship, a point-by-point connection and expansion operation is performed. Between each voxel and its adjacent voxels, if the heat change rate direction is the same and the risk heat value is not lower than the preset threshold, the adjacent voxels are classified into the same candidate hot zone, forming a spatially continuous voxel subset as a candidate high-risk hot zone;

[0176] For each candidate high-risk hotspot, the three-dimensional topological adjacency consistency index between each voxel is calculated. Based on the number of connections, direction differences, and relative changes in heat at the boundaries of the candidate hotspot, voxel points with unstable boundaries or inconsistent responses are identified. These boundary voxels are then removed from the candidate hotspot, retaining the main hotspot area with stable topological structure.

[0177] Match the main hot zone area with the structural information in the three-dimensional geological voxel model to identify the spatial overlap between the main hot zone and the geological unit boundary, fracture interface or construction disturbance boundary, and exclude voxel groups that do not have geological structural continuity or do not meet the geological model adjacency logic to ensure that the spatial structure of each hot zone is physically reasonable;

[0178] The spatial center position, voxel boundary range, heat change curve and three-dimensional spatial volume information of each main hot zone after screening and structural matching are extracted, and the above content is uniformly stored in the dynamic risk hot zone map, which serves as the input basis for path adjustment, construction rhythm control and power limit configuration in subsequent construction auxiliary decision-making.

[0179] During the implementation of the present invention, in order to achieve dynamic perception and accurate identification of risk status in underground construction environments, it is necessary to further analyze and structure the distribution characteristics of the risk heat value output by the neural network model in three-dimensional space.

[0180] First, the basic data source for hotspot identification should be clarified, namely the risk heat value distribution generated by the aforementioned convolutional recurrent neural network model. The risk heat value is a numerical expression of the risk level of each voxel unit in three-dimensional space under the current construction disturbance environment. After normalization, its value range usually falls between 0 and 1. The higher the value, the more likely the spatial location is to be in an accident-prone or engineering-instability area. Because construction disturbance is a dynamic process, the risk heat value also changes over time. Therefore, hotspot identification cannot rely solely on the heat distribution at a single moment. Instead, the changing trend of the heat along the time axis should be fully considered.

[0181] In this context, identifying voxel points where heat has changed significantly is the first step in the entire hot zone growth analysis. In order to quantify the temporal changes in risk heat, the present invention constructs a heat change map on each voxel unit, that is, by comparing the risk heat value at the current moment with the historical heat sequence in the previous time period (which can be set to the previous round of construction interval, such as 30 minutes, 1 hour, etc.), the change rate is calculated. Specifically, for each voxel , set its current heat to , the previous moment's heat was , then the rate of change of the voxel's heat can be defined as the difference To enhance noise immunity, a weighted sliding window approach can be used in practical implementations to calculate the mean or trend line over several past moments, followed by a difference operation. By setting a threshold for the rate of change in heat (e.g., 0.05), a set of voxels with significant heat changes can be screened out. These voxels form the seed point set for subsequent hotspot identification.

[0182] After the seed point set is established, the present invention adopts a spatial voxel connection expansion strategy based on local consistency conditions to achieve the structure growth of the hot zone in three-dimensional space. The specific method is: for any voxel node in the seed set , the system will check each neighboring voxel one by one within its 26 neighboring voxels , to determine whether it meets the following two conditions: one is the risk heat value of the neighboring voxel Above a preset threshold (such as 0.7), ensuring that it has a high risk level; secondly, the direction of its heat change is consistent with the current voxel, that is, , indicating that both are in the same disturbance trend. These two conditions together ensure the spatial connectivity of the hotspot and the consistency of its evolution over time. All neighboring voxels that meet these conditions will be included in the current candidate hotspot and serve as new expansion centers. The same logic will be further executed until all reachable voxels have been traversed, forming a preliminary candidate high-risk hotspot.

[0183] While preliminary hotspots demonstrate spatial connectivity and consistent thermal trends, they may contain noisy voxels or thermal transitions in their boundary regions due to factors such as model prediction errors and uncertainty in local geological disturbances. To further identify primary hotspots with stable structures and consistent internal responses, the present invention introduces a spatial topological adjacency consistency metric within candidate hotspots. Specifically, for each boundary voxel node in a candidate hotspot, the number of connections between it and its surrounding neighbors (e.g., the number of directions with high-heat neighbors), the thermal gradient difference (e.g., the difference in thermal values ​​between the voxel and its neighbors), and the distribution of structural directions (e.g., whether connected voxels are concentrated on one side or evenly distributed) are calculated. If a voxel is found to have too few connections, a large thermal difference, or an isolated structural direction, it is considered to be in a boundary transition zone or an isolated noise point and is removed from the hotspot. This process can be implemented using a simple scoring function, for example, by calculating a structural consistency score for each voxel and removing those below a threshold.

[0184] After completing the boundary cleaning, the hot zone obtained is called the main hot zone, and its internal voxels are connected in topological structure and consistent in thermal changes, forming a high-risk area that is independent in space and has a clear physical meaning. In order to further ensure the consistency of the hot zone with the actual geological environment, the present invention also introduces a structural matching process with the three-dimensional geological voxel model. The system will perform spatial matching between the main hot zone boundary and key structural surfaces such as fault zones, interlayer interfaces, and construction disturbance boundaries in the geological voxel model to analyze whether its boundary is continuous with the actual geological unit, whether it spans different lithologic areas, and whether it falls on the construction interference path. If it is found that some hot zones do not have clear structural continuity, such as falling into a known stable rock area in the geological model, or not coinciding with the actual cross-section profile, it can be considered to be a pseudo-hot zone formed by model misjudgment, and the system will eliminate it.

[0185] After the aforementioned multi-level processing, including heat screening, spatial aggregation, boundary consistency judgment, and structural matching, the final set of main hot spots will be recorded in the dynamic risk hot spot map. The map stores each main hot spot in a standardized manner, including the spatial center coordinates (such as the center of mass), the three-dimensional voxel boundary range (which can be represented as a minimum bounding box or voxel set), the risk heat value distribution curve (including maximum, minimum, average, and change trend), and overall volume information (converted into cubic meters using the number of voxels). Furthermore, its evolutionary characteristics during the construction process can be recorded, such as the shape expansion trend of the hot spot over the past few time periods and the drift path of the heat center, providing a basis for subsequent dynamic adjustments during construction.

[0186] The dynamic risk hotspot map ultimately serves as a key input to the decision-making support module, directly contributing to the development of core parameters such as path planning optimization, cadence control scheduling, and power limit configuration. By structuredly representing hotspot spatial information, the construction system can flexibly control the route and cadence of advancement while ensuring safety, avoiding prolonged operations in high-risk areas and effectively mitigating risks such as geological disasters and equipment overloads.

[0187] Step S105: In combination with the current construction stage, equipment deployment plan and sensor feedback information, a multi-objective auxiliary decision-making function is constructed based on the dynamic risk hot zone map, and construction efficiency, personnel safety and energy consumption constraints are comprehensively considered. Through iterative parameter optimization, auxiliary decision-making results including operation path reconstruction, construction rhythm adjustment and equipment power limit control suggestions are generated.

[0188] In step S105, the system integrates the identified dynamic risk hotspot map with the construction plan, equipment deployment information, and sensor feedback data to generate intelligent, assisted decision-making for the construction process. This assisted decision-making must not only consider safety risk avoidance but also balance construction efficiency, energy usage, and equipment deployment. Therefore, a multi-objective optimization model with adjustable objective functions, scalable parameter constraints, and an iterative solution mechanism must be constructed.

[0189] The system first reads the hotspot map generated in step S104. This map includes data such as each hotspot's number, voxel set, center coordinates, boundary range, average risk heat value, and its evolutionary trend over time. Simultaneously, the system also obtains information on the current construction progress (such as construction start and end locations, time window, and advancement rate), equipment deployment plans (such as the initial position, range of motion, and operating power limit of the shield machine or robotic arm), and real-time data from front-end sensors (such as soil pressure, vibration, groundwater flow rate, and power load). This information collectively constitutes the environmental input parameter set for the decision-making problem.

[0190] Based on the above input, the system needs to define an auxiliary decision function, the goal of which is to make the path reasonable, the rhythm controllable, and the energy consumption minimized while maintaining construction safety. Specifically, the operation path of a certain construction section is represented as an ordered voxel sequence , then for each path node , the system needs to adjust its operation speed (unit: ), power setting (Unit: kW) for optimized configuration.

[0191] The decision function can be formalized as the following objective combination:

[0192]

[0193] in, Represents a path node The risk heat value of the voxel (output by S103); Indicates the voxel operation advancement speed; Indicates the device power set at this voxel; are weight coefficients, representing the trade-off priorities among safety, efficiency, and energy consumption (dynamically adjustable, with recommended values ​​of 0.1, 0.5, and 0.2). Indicates the total number of voxel nodes contained in the current construction operation path, that is, the total number of three-dimensional space locations to be traversed on the path. In order to meet the actual engineering constraints, the system also introduces the following boundary conditions:

[0194] Each propulsion speed The rated limits of the equipment must not be exceeded;

[0195] power The sum must not exceed the transformer capacity or energy budget;

[0196] Changes in operating rhythm between adjacent voxels The permitted dynamic difference must not be exceeded to prevent the equipment from starting and stopping frequently;

[0197] If a voxel belongs to the severe level area in the hot zone map, its advancement speed must be below a set threshold (e.g. ), and trigger the "check-and-pass" strategy.

[0198] To solve the above objective function, the system uses a multi-objective iterative optimization algorithm. This can be implemented using methods such as the modified particle swarm optimization (MOPSO), the non-dominated sorting genetic algorithm (NSGA-II), or fast simulated annealing (Fast SA). One simple and easy-to-implement strategy is the greedy multi-objective equilibrium iteration method, as shown in the following code:

[0199] def optimize_path(path_voxels, risk_map, speed_limit, power_limit):

[0200] s = [initial_speed for _ in path_voxels]

[0201] p = [initial_power for _ in path_voxels]

[0202] for iteration in range(max_iterations):

[0203] for i in range(len(path_voxels)):

[0204] if risk_map[i]>high_risk_threshold:

[0205] s[i] = max(min_speed, s[i]- adjustment_rate)

[0206] p[i] = min(p[i], reduced_power_limit)

[0207] else:

[0208] s[i] = min(s[i]+ adjustment_rate, speed_limit)

[0209] total_power = sum(p)

[0210] if total_power>power_limit:

[0211] scale_factor = power_limit / total_power

[0212] p = [val * scale_factor for val in p]

[0213] return s, p

[0214] The algorithm dynamically adjusts the speed and power according to the risk level of each node, prioritizes reducing the propulsion intensity in high-risk sections, and performs constrained scaling of the overall power to ensure that the system load operates within a safe range.

[0215] This code implements a multi-objective balance optimization strategy for underground construction paths. Its core goal is to allocate appropriate propulsion speeds and equipment power to each construction location along the path, given a known risk distribution. This approach balances risk control with construction efficiency and energy constraints. This strategy, based on a greedy iteration approach, offers clear logic and is easy to implement, making it suitable for rapid parameter adjustments and local optimization during project deployment.

[0216] Specifically, path_voxels represents the planned construction path, which is composed of a series of spatial voxels; risk_map is a sequence of risk heat values ​​corresponding to the path positions (output by step S103); speed_limit is the upper limit of the propulsion speed, and power_limit is the upper limit of the total power allowed per unit time for the entire path.

[0217] The program first initializes a propulsion speed s and power p for each voxel position on the path, with the initial values ​​being initial_speed and initial_power respectively. It then enters the main iteration loop, where each position on the path is evaluated as follows:

[0218] If the risk heat value of the location exceeds the preset high risk threshold high_risk_threshold, it means that it belongs to the high risk section. At this time, the system will reduce the propulsion speed s[i] (but not lower than min_speed) and limit the device power p[i] to no more than reduced_power_limit.

[0219] If the risk is low, increase the advancement speed appropriately (without exceeding speed_limit) to improve construction efficiency.

[0220] After completing a round of speed and power updates, the program checks whether the total power (total_power) of all locations along the path exceeds the global power budget (power_limit). If so, all powers are proportionally scaled down using a linear scaling method to ensure that the total power meets the constraints. This logic ensures that the optimization results meet risk mitigation requirements without causing energy configuration overload. After the entire algorithm iterates max_iterations times, it returns the updated lists s and p, representing the optimized propulsion speed distribution and equipment power setting, respectively. In engineering, this result can be used to generate construction scheduling instructions, adjust equipment parameters, or serve as the initial solution for more complex optimization models.

[0221] The advantages of this algorithm lie in its simple implementation and fast computational speed, making it easy to embed into real-time scheduling systems or link with dynamic hotspot update models. It is suitable for fast-response parameter adaptive adjustment scenarios in high-risk construction environments. Ultimately, the system encapsulates the optimized path speed and power configuration results into a decision output, which contains the following information fields: propulsion speed recommendations for each path node, power setting recommendations, whether to recommend early maintenance, and whether to recommend avoiding this section of the area. Furthermore, the system can also visualize the change curves of key parameters during the optimization process, allowing decision makers to intuitively grasp the risk-benefit balance of the current strategy.

[0222] Through this complete multi-objective decision-making process, the system can not only respond to risk changes under complex geological conditions in real time, but also achieve dynamic optimization of construction rhythm, energy consumption allocation and operation path without sacrificing construction efficiency, providing highly reliable and highly feasible intelligent support means for high-risk underground construction.

[0223] A second embodiment of the present application provides an electronic device, comprising:

[0224] processor;

[0225] The memory is used to store a program, which, when read and executed by the processor, executes an underground construction decision-making method based on three-dimensional geological modeling and risk hot zone identification provided in the first embodiment of the present application.

[0226] The third embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the program executes an underground construction decision-making method based on three-dimensional geological modeling and risk hot zone identification provided in the first embodiment of the present application.

[0227] Although the present application is disclosed as above with the preferred embodiments, it is not intended to limit the present application. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. A method for underground construction decision-making based on three-dimensional geological modeling and risk hotspot identification, characterized in that: include: Acquire multi-source geological data including borehole data, geological radar images, and seismic reflection layer information to construct a three-dimensional geological voxel model with spatial topological constraints. The three-dimensional geological voxel model uses an irregular grid reconstruction method to describe the distribution characteristics of each geological unit; Based on the three-dimensional geological voxel model, extracting time series characteristic indicators under the influence of construction disturbance, including local pressure gradient changes, seepage velocity estimation and structural response heterogeneity factors, to form a continuous time series characteristic vector; Inputting the continuous time series feature vector into a trained convolutional recurrent neural network model, which has a spatial attention aggregation mechanism and deep memory units, is used to model the nonlinear risk propagation behavior during the construction evolution process and output the risk heat value of each spatial location; Based on the distribution characteristics of risk heat values ​​in three-dimensional space, the heat gradient clustering and neighborhood structure consistency screening algorithm are used to identify spatially continuous and dynamically changing high-risk hot spots, forming a dynamic risk hot spot map. Combining the current construction stage, equipment deployment plan and sensor feedback information, a multi-objective auxiliary decision-making function is constructed based on the dynamic risk hot zone map. Taking construction efficiency, personnel safety and energy consumption constraints into comprehensive consideration, an auxiliary decision-making result including operation path reconstruction, construction rhythm adjustment and equipment power limit control suggestions is generated through iterative parameter optimization.

2. The underground construction decision-making method based on three-dimensional geological modeling and risk hot zone identification according to claim 1 is characterized in that: The convolutional recurrent neural network model includes a temporal geological disturbance feature extraction layer, a spatial topology perception convolutional encoding layer, an attention regulation recursive modeling layer, and a heat regression decoding layer; The time-series geological disturbance feature extraction layer is used to receive continuous time series feature vectors extracted based on the three-dimensional geological voxel model, and the input is a multi-channel time series tensor containing pressure gradient, seepage velocity and structural heterogeneity factor. It uses a one-dimensional time series convolution method to realize dynamic pattern recognition across time windows and output the local disturbance response spectrum corresponding to each voxel. The spatial topology-aware convolutional coding layer is used to integrate the geological structural adjacency relationship between voxel nodes. The input is the disturbance response spectrum and the topological graph structure tensor. The topological embedding coding method is combined with the three-dimensional irregular convolution operator to achieve modeling of spatially irregular geological connection relationships. The output is a topological convolution feature map. The topological graph structure tensor is constructed by the spatial adjacency relationship and the geological structural boundary relationship between voxels in the three-dimensional geological voxel model. It is used to clarify the topological connection mode of each voxel node and serves as the adjacency input in the spatial convolution operation. The attention control recursive modeling layer is used to receive the topological convolution feature map, combine the historical disturbance evolution state, and construct a multi-head spatial attention gating mechanism and a bidirectional recurrent unit to dynamically adjust the risk dependency weight of each region voxel in the multi-step prediction process, and output the dynamic state vector of the voxel node in the full-time risk perception space; The heat regression decoding layer performs nonlinear dimensionality reduction and regression transformation on the dynamic state vector to generate a risk heat value corresponding to each spatial voxel position. The output is a normalized floating-point number, which is used as quantitative input for hot zone identification and auxiliary decision-making.

3. The underground construction decision-making method based on three-dimensional geological modeling and risk hot zone identification according to claim 2 is characterized in that: The time series geological disturbance feature extraction layer is also used to: The continuous time series feature vectors of each voxel node are normalized, and a multi-channel normalization mapping mechanism is used to map the pressure gradient change, infiltration velocity estimation, and structural heterogeneity factor to comparable scales in the interval [–1, 1] to ensure that the response weights of different physical quantities in the convolution extraction are balanced. A one-dimensional variable convolution kernel structure across time windows is used to simultaneously model disturbance variation patterns within fixed time steps and adaptive time windows. The variable convolution kernel size is dynamically determined by the current construction disturbance activation index. This allows the model to automatically enhance its sensitivity to short-term features during sudden disturbances and improve its ability to extract long-term trends during stable periods. The maximum pooling and residual connection structure in the time direction are introduced to the convolution output to enhance the robustness of local patterns and retain the long-term information in the original input features. This enables the output disturbance response spectrum to have both the ability to detect short-term disturbances and maintain long-term change trends, adapting to the semantic alignment requirements of subsequent spatial topology modeling.

4. The underground construction decision-making method based on three-dimensional geological modeling and risk hot zone identification according to claim 2 is characterized in that: The spatial topology-aware convolutional coding layer is further used for: Based on the topological structure tensor constructed in the three-dimensional geological voxel model, a geological structure adjacency matrix and a fault contact surface weight matrix are used to jointly construct a voxel graph representation, wherein each graph edge not only represents a geometric adjacency relationship but also carries a semantic label representing the contact interface type, including a fracture surface, a soft-hard interface, or a homolithic continuous surface; The introduction of a projection mechanism for irregular convolution kernels on graph structures allows the convolution operation to adjust the propagation weight according to the attributes of different types of adjacent edges, thus achieving discriminative structural perception. A topological embedding vector is fused in the convolution operation. The topological embedding vector is generated by the geological structure encoding module based on the multi-order adjacency information of the voxel node, indicating the global position role of the node in the spatial structure, including whether it is in an intersection area, a closed area, an edge area or a uniform area, thereby enabling the topological convolution feature map to have the ability to express multi-scale geological structure context.

5. The underground construction decision-making method based on three-dimensional geological modeling and risk hot zone identification according to claim 2 is characterized in that: The attention-regulated recursive modeling layer is further used to: A multi-head spatial attention mechanism is introduced to construct three types of attention weight vectors based on spatial position, topological role, and perturbation intensity at each time step. The final comprehensive attention map is obtained through linear weighted aggregation to adjust the information channel distribution of input features in the recursive structure. A bidirectional gated recursive structure is used to propagate the disturbance response state bidirectionally along the time dimension with voxel nodes as the unit. The trend of the disturbance evolving from the edge of the hot zone to the core is captured in the forward propagation, and the disturbance convergence or disappearance signal is supplemented in the backward propagation, thereby improving the modeling ability of nonlinear disturbance paths. A perturbation state evolution encoder is added to the recursive output stage to record the heat change pattern of each voxel in the entire time series window, including statistical features such as the maximum fluctuation point, continuous perturbation interval, and change slope distribution. The evolution code is then concatenated with the recursive state vector and used as the dynamic state output to enhance the interpretability and prediction accuracy of subsequent heat regression.

6. The underground construction decision-making method based on three-dimensional geological modeling and risk hot zone identification according to claim 1 is characterized in that: According to the distribution characteristics of risk heat values ​​in three-dimensional space, the heat gradient clustering and neighborhood structure consistency screening algorithm are used to identify spatially continuous and dynamically changing high-risk hot spots, forming a dynamic risk hot spot map, including: Based on the risk heat value, a risk heat change map is constructed, the risk heat value of each voxel at the current moment is subtracted from its risk heat historical sequence in the previous time period, the heat time change rate of each voxel is extracted, and the heat time change rate is used as an initial screening basis to identify a set of voxels with significant heat changes; The voxel set is used as the seed voxel set. Based on the consistency of the heat change rate direction and the three-dimensional spatial adjacency relationship, a point-by-point connection and expansion operation is performed. Between each voxel and its adjacent voxels, if the heat change rate direction is the same and the risk heat value is not lower than the preset threshold, the adjacent voxels are classified into the same candidate hot zone, forming a spatially continuous voxel subset as a candidate high-risk hot zone; For each candidate high-risk hotspot, the three-dimensional topological adjacency consistency index between each voxel is calculated. Based on the number of connections, direction differences, and relative changes in heat at the boundaries of the candidate hotspot, voxel points with unstable boundaries or inconsistent responses are identified. These boundary voxels are then removed from the candidate hotspot, retaining the main hotspot area with stable topological structure. Match the main hot zone area with the structural information in the three-dimensional geological voxel model to identify the spatial overlap between the main hot zone and the geological unit boundary, fracture interface or construction disturbance boundary, and exclude voxel groups that do not have geological structural continuity or do not meet the geological model adjacency logic to ensure that the spatial structure of each hot zone is physically reasonable; The spatial center position, voxel boundary range, heat change curve and three-dimensional spatial volume information of each main hot zone after screening and structural matching are extracted, and the spatial center position, voxel boundary range, heat change curve and three-dimensional spatial volume information are uniformly stored in the dynamic risk hot zone map, which serves as the input basis for path adjustment, construction rhythm control and power limit configuration in subsequent construction auxiliary decision-making.

7. The underground construction decision-making method based on three-dimensional geological modeling and risk hot zone identification according to claim 1 is characterized in that: The method of acquiring multi-source geological data including borehole data, geological radar images, and seismic reflection layer information to construct a three-dimensional geological voxel model with spatial topological constraints includes: Based on the borehole data, geological radar images and seismic reflection layer information, a geological profile set is generated and a preliminary model of a three-dimensional geological structure is constructed to drive the division boundary of the three-dimensional geological voxel grid; Based on the fault zones, lithologic transition zones, and structural boundary areas identified in the preliminary 3D geological structure model, a voxel refinement strategy is implemented to locally densify the grid in highly heterogeneous areas, thereby forming an irregular grid partitioning structure with controlled partition density. The geological information entropy distribution map is calculated based on multi-source geological data to evaluate the geological complexity of each region in three-dimensional space. Based on this, the density of the voxel grid is dynamically adjusted to refine the information-complex areas and coarsen the information-uniform areas. For each voxel node after irregular grid division, geological unit number information and spatial topology label are added. The former is used to identify the stratigraphic category and lithologic type to which it belongs, and the latter is used to identify the spatial connection relationship and structural boundary attributes between it and adjacent voxels. A three-dimensional geological voxel model with spatial topological constraints was constructed. The model uses irregular grid reconstruction to describe the distribution characteristics of various geological units. It has the ability to express fracture structures, stratigraphic boundaries and spatial discontinuities, providing a structural basis for subsequent underground construction risk analysis.

Images