Modeling method for piles around tunnel based on multi-source geological data fusion

Abstract

The invention provides a tunnel surrounding pile modeling method based on multi-source geological data fusion, and the method comprises the steps: accessing and integrating multi-source geological data through a BIM software API, constructing a hierarchical topology node system, constructing an initial edge weight in combination with a geological influence coefficient, forming an initial topology network, embedding the initial topology network into a BIM three-dimensional environment, and carrying out the construction of a multi-source geological data model. Performing pre-training through a position-aware dynamic graph neural network to obtain an initial structure model; according to the construction disturbance influence factor and the data credibility, dynamically adjusting the weight of the edge between the influenced nodes, and updating the topology network in real time; according to the activated incremental nodes and the topological neighborhood range thereof, local incremental reasoning is executed to fuse neighborhood node features, uncertainty quantitative evaluation is carried out based on a probabilistic reasoning method, and then an initial structure model is dynamically updated and corrected to form a construction dynamic structure model; and fitting the tunnel initial model and the construction dynamic structure model to obtain a combined model for guiding pile foundation construction.

Description

Technical Field

[0001] This invention relates to the field of pile foundation construction modeling technology, specifically a method for modeling piles around tunnels based on multi-source geological data fusion. Background Technology

[0002] Soil cavities and karst caves pose certain hazards to pile foundation construction, significantly impacting the safety and quality of bored pile construction. However, the foundation engineering in airport expansion differs from general construction safety operations in that airports are high-density pedestrian areas, making the controllability of the entire process of cast-in-place pile construction a crucial control objective for safe construction within airports. This is due to the large size of soil cavities and karst caves on site, the shallow depth of some cavities, and the complex combination of soil cavities and karst caves.

[0003] When constructing large-diameter bored piles in or near tunnels, especially under adverse geological conditions such as karst caves and soil caves, traditional exploration methods struggle to accurately determine the distribution, scale, infill material properties, and spatial relationship with the tunnel. This results in a lack of targeted construction plans and increases the risk of accidents such as borehole collapse, borehole deviation, and grout leakage. In related technologies, pile foundation construction modeling in geology with extremely high karst / soil cave ratios often relies on BIM for geological visualization, but this cannot reflect dynamic changes in construction or the reliability of results, making it difficult to adapt to the construction decision-making needs of complex geological conditions. Summary of the Invention

[0004] The purpose of this invention is to provide a method for modeling piles around tunnels based on the fusion of multi-source geological data, in order to solve the technical problems mentioned in the background.

[0005] To achieve the above objectives, the present invention provides the following technical solution.

[0006] According to an embodiment of the present invention, a method for modeling piles around tunnels based on multi-source geological data fusion is provided, comprising: Multi-source geological data is accessed and integrated through BIM software API, and a hierarchical topology node system is constructed based on the multi-source geological data, including benchmark nodes, core nodes and associated nodes. Core nodes and benchmark nodes are bidirectionally connected, and associated nodes and core nodes are unidirectionally connected. Extract the three-dimensional coordinates of each node, calculate the spatial distance between nodes with connection relationships, construct the initial edge weights by combining the geological influence coefficient, form the initial topology network, embed the initial topology network into the BIM three-dimensional environment, and obtain the initial structural model by pre-training the location-aware dynamic graph neural network. Based on the pile foundation construction sequence, the incremental nodes corresponding to the newly added pile positions are activated in batches. The pile foundation information during the construction process is obtained in real time through the BIM platform, the position and attributes of the newly added nodes are updated, and only the incremental nodes are connected to the existing core nodes of the corresponding level. The relationship between the nodes is visualized and marked in the BIM model. According to the construction disturbance impact factor and data credibility, the weight of the edges between the affected nodes is dynamically adjusted, and the topology network is updated in real time. For the activated incremental node and its topological neighborhood range, local incremental reasoning is performed to fuse the features of the neighborhood node, and uncertainty quantification assessment is performed based on probabilistic reasoning method, thereby dynamically updating and correcting the initial structural model to form a construction dynamic structural model. The initial tunnel model is fitted with the construction dynamic structural model to obtain a combined model. The minimum spatial distance between each pile foundation and the tunnel structure is calculated. Combined with the stratum mechanical parameters and the tunnel safety protection range standard, it is determined whether the pile foundation is within the safe range of the tunnel. Based on the judgment results, risk warnings and adjustment suggestions are output to guide the pile foundation construction.

[0007] Furthermore, the initial tunnel model is obtained through the following steps: Radar wave data of the underground tunnel area is acquired, and the Monte Carlo Dropout method is used to perform multiple forward inferences and uncertainty quantifications on the preset neural network to construct a probability distribution model of the tunnel's location and outline. Based on this model, the three-dimensional location and outline of the tunnel are determined, and an initial tunnel model is generated.

[0008] Furthermore, it also includes: after the pile foundation construction is completed, using a track inspection trolley to continuously or periodically monitor the deformation inside the tunnel and obtain tunnel deformation data; feeding the measured deformation data back to the BIM model, and analyzing and optimizing the combined model.

[0009] Furthermore, the step of forming the initial topology network includes: Identify node pairs with connections, including connections between core nodes and base nodes, and connections between core nodes and associated nodes; The spatial distance between node pairs is calculated using the three-dimensional Euclidean distance formula, expressed as: in, and These are the three-dimensional coordinates of the two nodes in the node pair; The first geological influence coefficient is determined based on the node pair type and geological attributes. With the second geological influence coefficient : The connection between the core node and the baseline node is determined based on the geological risk level corresponding to the core node. ; The connection between core nodes and associated nodes is determined based on their geological connectivity. ; The initial boundary weight W is calculated using spatial distance and geological influence coefficient, and is expressed as: In the formula, K is the corresponding geological influence coefficient. ;d represents the three-dimensional Euclidean distance between node pairs, and the edge weight calculation results are normalized to the range of 0-1.

[0010] Furthermore, the step of pre-training a location-aware dynamic graph neural network to obtain the initial structural model before construction includes: A location encoding module for a geological scene is constructed to generate a fused location feature vector for each node. The location feature vector is composed of absolute location encoding, relative location encoding and geological layer location encoding. A dual attention fusion layer based on location and geology is constructed. Based on the location feature vector and the geological attribute feature vector of the node, the location attention coefficient and the geological attention coefficient are calculated respectively, and the two coefficients are fused to obtain the association attention coefficient. The initial edge weights are enhanced based on the fused association attention coefficient. The positional attention coefficient is calculated based on the positional feature vector and is expressed as follows: In the formula, This represents the positional attention coefficient between node i and node j. The feature vector representing the position of node i. This represents the feature vector representing the position of node j. The dimension of the location feature vector. Represents the normalization function; The geological attention coefficient is calculated based on the geological attribute feature vector of a node, and is expressed as: In the formula, This represents the geological attribute feature vector of node i. This represents the geological attribute feature vector of node j. The dimension representing the geological attribute feature vector; This represents the geological attention coefficient between node i and node j; The enhanced edge weights and node features are input into a dynamic neural network and trained using a dynamic learning strategy to obtain the initial structural model before construction.

[0011] Furthermore, the step of training using a dynamic learning strategy includes: Based on the geological influence coefficient K and the location attention coefficient of the node pair and distance factor Calculate the geological correlation complexity score , is represented as: ,in, For distance-based normalization factor, d is the node distance; The training samples are divided into three levels—simple, medium, and complex—based on a score S, and then input into the model for training in stages. A dual-trigger adaptive learning rate adjustment mechanism is adopted to adjust the learning rate of the position encoding and attention modules; A multi-objective balanced loss function is constructed, and the total loss is composed of a weighted average of feature reconstruction loss, location uncertainty loss, and attention consistency loss.

[0012] Furthermore, the step of activating the incremental nodes corresponding to the newly added pile positions in batches according to the pile foundation construction sequence includes: Establish the mapping relationship between processes and incremental nodes, and determine the activation trigger conditions for each incremental node; Once activated, each incremental node establishes a connection only with existing nodes within its spatial neighborhood.

[0013] Furthermore, the step of performing local incremental inference for the activated incremental node and its topological neighborhood range includes: Based on the type of incremental node and its geological risk level, the neighborhood radius of its local inference is dynamically determined, and existing nodes within the spatial range are selected based on the neighborhood radius to form a local inference subgraph. An improved graph SAGE convolution algorithm is used to reason about the local reasoning subgraph and aggregate neighborhood information; wherein, different attention weights are assigned to different types of neighborhood nodes in the subgraph according to the node level and data credibility. The features of the neighboring nodes are divided into spatial features, geological attribute features and dynamic association features, and then weighted and summed according to preset semantic weights to generate the fusion feature vector of the incremental node. The fused feature vector is updated to the corresponding incremental node in the initial structural model, and the relevant edge weights within the local inference subgraph are adjusted.

[0014] Furthermore, the step of forming a local reasoning subgraph includes: Based on the geological parameters and construction conditions associated with incremental nodes, a two-level risk classification is performed, dividing the risk into high risk, medium risk, and low risk; based on the incremental node type and geological risk level, the neighborhood radius is determined by mapping. Based on the neighborhood radius, a multi-condition neighborhood node screening is performed, including but not limited to spatial distance screening, hierarchical constraint screening, and validity screening, to obtain the screened existing nodes; the incremental nodes and the screened existing nodes are used as subgraph nodes, and their original edge connection relationships are retained to construct a local reasoning subgraph.

[0015] Furthermore, the step of using an improved graph SAGE convolution algorithm to infer the local inference subgraph and aggregate neighborhood information includes: assigning differentiated hierarchical aggregation weights to the neighborhood nodes in the local inference subgraph according to their hierarchical types. When aggregating features of neighboring nodes, the updated dynamic edge weights are used as the basis. The aggregation process, which involves weighting the features of neighboring nodes, can be represented as follows: Where Z is the normalization coefficient. Weights are retained for the incremental node's own characteristics. The hierarchical aggregation weight corresponding to the neighboring node u. Let b be the convolution weight matrix, and b be the bias term. It is a non-linear activation function; The characteristics of the neighboring node u, This represents the feature vector of node v in the previous construction stage. This represents the aggregated feature vector of node v in the current construction phase; This represents the dynamic edge weight between node u and node v; For neighboring nodes with the geological attribute of karst caves, the erosion features representing the degree of erosion are embedded into the vector, and then fused with the original node features with preset weights before participating in the aggregation. For dynamic neighborhood nodes that characterize the state of construction disturbance, the dynamic feature vectors constructed from the construction disturbance influence factor and the data credibility factor are incorporated into the original node features with preset weights and then participated in the aggregation.

[0016] Compared with existing technologies, the beneficial effects of the method for modeling the construction of karst cave pile foundations above tunnels in this invention are: This invention accesses and integrates multi-source geological data through BIM software API, constructs a hierarchical topological node system based on the multi-source geological data, extracts the three-dimensional coordinates of each node, calculates the spatial distance between connected node pairs, constructs initial edge weights in combination with geological influence coefficients, forms an initial topological network, embeds the initial topological network into the BIM three-dimensional environment, and obtains an initial structural model through pre-training with a location-aware dynamic graph neural network, thereby improving the accuracy, efficiency, and understanding of engineering scenarios in geological model construction. This invention activates incremental nodes corresponding to newly added pile positions in batches according to the pile foundation construction sequence. It acquires pile foundation information in real time through a BIM platform, updates the position and attributes of newly added nodes, and connects only the incremental nodes to existing core nodes at the corresponding level. The relationships between nodes are then visualized in the BIM model. Based on construction disturbance impact factors and data reliability, the weights of edges between affected nodes are dynamically adjusted, and the topology network is updated in real time. For activated incremental nodes and their topological neighborhoods, local incremental inference is performed to fuse neighborhood node characteristics, and uncertainty quantification is conducted based on probabilistic inference methods. This dynamically updates and corrects the initial structural model, forming a dynamic construction structural model. By binding construction sequences to node activation and introducing local incremental inference and uncertainty assessment mechanisms, the dynamic updating of the geological structure model is achieved, and the model's response speed to construction disturbances is improved, providing real-time risk prediction capabilities for construction safety. This invention fits the initial tunnel model with the dynamic construction structure model to obtain a combined model, calculates the minimum spatial distance between each pile foundation and the tunnel structure, and combines the geological mechanics parameters and tunnel safety protection range standards to determine whether the pile foundation is within the tunnel's safety range. Based on the judgment results, it outputs risk warnings and adjustment suggestions to guide pile foundation construction. This method, by dynamically fitting the tunnel and pile foundation models and performing real-time analysis, can assess the safety risks of the tunnel adjacent to the pile foundation construction, significantly improving the safety and controllability of the construction process. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] In the attached diagram: Figure 1 This is a flowchart illustrating the implementation of the tunnel perimeter pile modeling method based on multi-source geological data fusion according to the present invention. Figure 2 This is a sub-flowchart of the tunnel surrounding pile modeling method based on multi-source geological data fusion of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] like Figure 1 As shown, in one embodiment of the present invention, a method for modeling piles around tunnels based on multi-source geological data fusion is provided, including the following steps: Step S1: Access and integrate multi-source geological data through BIM software API, and construct a hierarchical topology node system based on the multi-source geological data, including benchmark nodes, core nodes and related nodes. Core nodes and benchmark nodes are bidirectionally connected, and related nodes and core nodes are unidirectionally connected. In step S1, the multi-source geological data of the present invention includes tunnel structure data, geological exploration data and advanced drilling data. Through the structured data reading advantage of BIM software API, it can connect to the standardized formats of various data sources, avoiding the errors caused by manual data entry and the problem of inconsistent data formats. In this invention, scattered multi-source geological data are bound and associated according to data source and node type, and the associated and integrated data provides a standardized input for subsequent extraction of node three-dimensional coordinates and determination of geological influence coefficients. In this embodiment of the invention, the reference node is based on tunnel structure data, and the center point of the key section of the tunnel is selected as the node location; the core node is based on geological exploration data, and the center of the karst cave and the distribution center of the weak interlayer are selected as the node location; the associated node is based on advanced drilling data, and the center of the small fracture and the center of the local loose body are selected as the node location.

[0021] Please continue to refer to Figure 1 This invention is a method for modeling piles around tunnels based on the fusion of multi-source geological data; It also includes the following steps: Step S2: Extract the three-dimensional coordinates of each node, calculate the spatial distance between node pairs with connection relationships, construct the initial edge weights in combination with the geological influence coefficient, form the initial topology network, embed the initial topology network into the BIM three-dimensional environment, and pre-train it through a location-aware dynamic graph neural network to obtain the initial structural model. In step S2, the step of forming the initial topology network in this invention includes: determining the node pairs with interconnected relationships, including the connection between the core node and the reference node, and the connection between the core node and associated nodes; calculating the spatial distance between node pairs using the three-dimensional Euclidean distance formula, expressed as: in, and These are the three-dimensional coordinates of the two nodes in the node pair; the first geological influence coefficient is determined based on the node pair type and geological attributes. With the second geological influence coefficient For the connection between the core node and the baseline node, the risk level of the geological body corresponding to the core node is determined. The connection between core nodes and related nodes is determined based on their geological connectivity. The initial boundary weight W is calculated using spatial distance and geological influence coefficient, and is expressed as: In the formula, K is the corresponding geological influence coefficient. ;d represents the three-dimensional Euclidean distance between node pairs, and the edge weight calculation results are normalized to the range of 0-1.

[0022] The three-dimensional Euclidean distance in this embodiment of the invention is used to quantify the physical proximity between nodes. The closer the distance, the stronger the basic correlation. The geological influence coefficients set include those based on risk level guidance. and connectivity-oriented By injecting geological scene risk attributes, this invention overcomes the limitations of traditional methods that only quantify associations based on distance. It also embeds a topological network into the BIM environment, achieving a combination of topological association and engineering visualization, and providing structured and visualized input data for subsequent model pre-training.

[0023] In one implementation, the present invention pre-trains a location-aware dynamic graph neural network to obtain an initial structural model before construction. This includes: constructing a location encoding module for the geological scene, generating a fused location feature vector for each node, wherein the location feature vector is composed of absolute location encoding, relative location encoding, and geological layer location encoding; constructing a dual-attention fusion layer for location and geology, calculating location attention coefficients and geological attention coefficients based on the location feature vectors and the geological attribute feature vectors of the nodes, and fusing the two coefficients to obtain an association attention coefficient; and enhancing the initial edge weights based on the fused association attention coefficient. The location attention coefficient is calculated based on the location feature vector and is expressed as: In the formula, This represents the positional attention coefficient between node i and node j. The feature vector representing the position of node i. This represents the feature vector representing the position of node j. The dimension of the location feature vector. Represents the normalization function; The geological attention coefficient provided by this invention is calculated based on the geological attribute feature vector of a node, and is expressed as follows: In the formula, This represents the geological attribute feature vector of node i. This represents the geological attribute feature vector of node j. The dimension representing the geological attribute feature vector; This represents the geological attention coefficient between node i and node j; The enhanced edge weights and node features are input into a dynamic neural network and trained using a dynamic learning strategy to obtain the initial structural model before construction.

[0024] As can be seen, in the dual attention fusion layer of the present invention, the location attention is set for spatially nearby nodes, and the geological attention is set for high-risk / strongly correlated nodes. After the two attentions are fused, the edge weights are enhanced, so that the model can give priority to the correlation between spatial proximity and geological key, such as high-risk karst caves and tunnel arch nodes.

[0025] In one implementation of the present invention, the step of training using a dynamic learning strategy includes: Based on the geological influence coefficient K and the location attention coefficient of the node pair and distance factor Calculate the geological correlation complexity score , is represented as: ,in, For distance-based normalization factor, d is the node distance; The training samples are divided into three levels—simple, medium, and complex—based on a score S, and are then input into the model for training in stages. A dual-trigger adaptive learning rate adjustment mechanism is used to adjust the learning rates of the position encoding and attention modules. A multi-objective balanced loss function is constructed, with the total loss consisting of a weighted average of feature reconstruction loss, position uncertainty loss, and attention consistency loss.

[0026] Please continue to refer to Figure 1 The modeling method of the present invention further includes: Step S3: Based on the pile foundation construction sequence, activate the incremental nodes corresponding to the newly added pile positions in batches, obtain the pile foundation information in real time through the BIM platform, update the position and attributes of the newly added nodes, and connect only the incremental nodes with the existing core nodes of the corresponding level, and visualize the relationship between the nodes in the BIM model; dynamically adjust the weight of the edges between the affected nodes according to the construction disturbance impact factor and data credibility, and update the topology network in real time. In step S3, the present invention activates the incremental nodes corresponding to newly added pile positions in batches according to the pile foundation construction procedure, including: establishing a mapping relationship between the procedure and the incremental nodes, and determining the activation trigger conditions for each incremental node; after each incremental node is activated, it only establishes a connection with the existing nodes within its spatial neighborhood.

[0027] This invention breaks down the pile foundation construction process into the following steps, including but not limited to: advanced drilling, borehole formation, hole cleaning, reinforcement cage lowering, and concrete pouring. Each step corresponds to specific geological information or construction parameter supplementation requirements. Each step corresponds to one or more types of incremental nodes, forming a one-to-one or many-to-many mapping between steps and nodes. This invention establishes a structured mapping table through a BIM platform, clearly marking the incremental node ID, node type, and associated data source corresponding to each step, ensuring rapid matching upon activation.

[0028] This invention collects construction progress data and monitoring parameters in real time through a BIM platform. When the triggering conditions are met, the corresponding incremental node is automatically activated without manual intervention, ensuring timeliness.

[0029] This invention activates the model in batches according to the process, allowing the model update to be synchronized with the construction progress. Furthermore, by filtering connection nodes through spatial range and hierarchical constraints, it significantly reduces invalid associations and lowers subsequent calculation costs.

[0030] Please continue to refer to Figure 1 The modeling method of the present invention further includes: Step S4: For the activated incremental node and its topological neighborhood range, perform local incremental reasoning to fuse the features of the neighborhood node, and perform uncertainty quantification assessment based on probabilistic reasoning method, thereby dynamically updating and correcting the initial structural model to form a construction dynamic structural model. Please refer to Figure 2 In one implementation of the present invention, the step of performing local incremental inference for the activated incremental node and its topological neighborhood range includes: Step S41: Based on the type of the incremental node and its geological risk level, dynamically determine the neighborhood radius of its local inference, and filter the existing nodes within the spatial range based on the neighborhood radius to form a local inference subgraph. In the above steps, the present invention determines the base radius according to the type of incremental node. The type of incremental node directly determines its geological influence range. For example, if the core node has a wide influence range on the surrounding geological body and tunnel structure, the base radius can be set to 3-5m; associated nodes only serve as risk transmission carriers and have a limited influence range, so the base radius can be set to 1-2m. After setting the base radius, the present invention adjusts the radius according to the geological risk level. For example, the risk level of an incremental node is determined by its associated geological parameters and construction conditions. For high-risk nodes, the base radius needs to be increased by 0.5-1m to ensure coverage of all potentially affected existing nodes; while for medium / low-risk nodes, the base radius is used to avoid over-calculation.

[0031] In the step of forming a local reasoning subgraph by filtering existing nodes based on radius, this invention filters existing nodes within a set neighborhood radius using multiple conditions to ensure that subgraph nodes only contain valid information strongly related to the incremental node. The filtering of existing nodes can be spatial distance constraint filtering, hierarchical constraint filtering, and validity filtering. In the spatial distance constraint filtering, only nodes whose three-dimensional coordinates and Euclidean distance from the incremental node are not greater than the neighborhood radius are retained, ensuring that the node is physically associated with the incremental node. In the hierarchical constraint filtering, irrelevant nodes are avoided across levels according to preset topology connection rules. In the validity filtering, invalid nodes are removed to ensure the validity of the subgraph node information.

[0032] Specifically, in this embodiment of the invention, the step of forming a local inference subgraph further includes: performing a secondary risk classification based on the geological parameters and construction conditions associated with the incremental nodes, dividing them into high risk, medium risk, and low risk; mapping and determining the neighborhood radius based on the incremental node type and geological risk level; performing multi-condition neighborhood node screening based on the neighborhood radius, including but not limited to spatial distance screening, hierarchical constraint screening, and validity screening, to obtain the screened existing nodes; using the incremental nodes and the screened existing nodes as subgraph nodes, retaining their original edge connection relationships, to construct a local inference subgraph.

[0033] Please refer to Figure 2 The step of performing local incremental inference for the activated incremental node and its topological neighborhood range also includes: Step S42: The improved graph SAGE convolution algorithm is used to reason about the local reasoning subgraph and aggregate neighborhood information; wherein, different attention weights are assigned to different types of neighborhood nodes in the subgraph according to the node level and data credibility. In step S42, the present invention employs an improved graph SAGE convolution algorithm to infer the local inference subgraph and aggregate neighborhood information. This step includes: assigning differentiated hierarchical aggregation weights to the neighborhood nodes in the local inference subgraph based on their hierarchical type. When aggregating features of neighboring nodes, the updated dynamic edge weights are used as the basis. The aggregation process, which involves weighting the features of neighboring nodes, can be represented as follows: Where Z is the normalization coefficient. Weights are retained for the incremental node's own characteristics. The hierarchical aggregation weight corresponding to the neighboring node u. Let b be the convolution weight matrix, and b be the bias term. It is a non-linear activation function; The characteristics of the neighboring node u, This represents the feature vector of node v in the previous construction stage. This represents the aggregated feature vector of node v in the current construction phase; This represents the dynamic edge weight between node u and node v; In this embodiment of the invention, for a neighboring node whose geological attribute is a karst cave, the erosion feature embedding vector representing the degree of erosion is fused with the original node feature with a preset weight and then participates in the aggregation; for a dynamic neighboring node representing the state of construction disturbance, the dynamic feature vector constructed by the construction disturbance influence factor and the data credibility factor is incorporated into the original node feature with a preset weight and then participates in the aggregation.

[0034] Please continue to refer to Figure 2 The step of performing local incremental inference for the activated incremental node and its topological neighborhood range also includes: Step S43: Divide the features of the neighboring nodes into spatial features, geological attribute features and dynamic association features, and perform weighted summation according to preset semantic weights to generate the fusion feature vector of the incremental node; In step S43 of this embodiment, spatial features are used to characterize the spatial positional relationship between incremental nodes and neighboring nodes; geological attribute features are used to characterize the inherent attributes of the geological bodies corresponding to the neighboring nodes; and dynamic association features are used to characterize the dynamic impact of construction disturbances and data quality on node association. Step S44: Update the fused feature vector to the corresponding incremental node in the initial structural model, and adjust the relevant edge weights within the local inference subgraph; after the incremental node features are updated, the association strength between it and its neighboring nodes changes, so the edge weights need to be adjusted to adapt to this change; this invention only applies to the associated edge weights within the local inference subgraph constructed in step S41, and does not involve the global edge weights outside the subgraph, in order to control computational costs; wherein, for the adjustment of edge weights, based on the initial edge weights and dynamic updates, the optimized adjustment formula is expressed as: In the formula, This represents the cosine similarity between the fused feature vector of the incremental node and the feature vectors of its neighboring nodes. Indicates dynamic edge weights. Indicates the adjusted edge weights; Therefore, by binding construction procedures with node activation and introducing local incremental reasoning and uncertainty assessment mechanisms, this invention achieves dynamic updates of the geological structure model, improves the model's response speed to construction disturbances, and provides real-time risk prediction capabilities for construction safety.

[0035] Please continue to refer to Figure 1 The modeling method of the present invention further includes: Step S5: Fit the initial tunnel model with the construction dynamic structure model to obtain a combined model, and calculate the minimum spatial distance between each pile foundation and the tunnel structure. Combine the geological mechanics parameters and the tunnel safety protection range standard to determine whether the pile foundation is within the tunnel safety range. Based on the judgment results, output risk warnings and adjustment suggestions to guide the pile foundation construction.

[0036] When performing model fitting, this invention uses the spatial coordinate system of the BIM 3D environment as a standard to avoid fitting misalignment caused by coordinate deviation. This invention verifies the node ID and spatial position to correspond one-to-one with the incremental pile foundation nodes and geological core nodes in the construction dynamic structural model and the tunnel reference nodes in the initial tunnel model. Then, the topology networks of the two models are superimposed to establish the feature association between the pile foundation nodes and the tunnel nodes to achieve fitting. Furthermore, this invention uses a three-dimensional Euclidean distance formula to calculate the minimum spatial distance between each pile foundation and the tunnel structure.

[0037] In step S5 of the present invention, the initial tunnel model of the present invention is obtained through the following steps: acquiring radar wave data of the underground tunnel area, using the Monte Carlo Dropout method to perform multiple forward inferences and uncertainty quantifications on the preset neural network, constructing a probability distribution model of the tunnel location and contour, and determining the three-dimensional location and contour of the tunnel accordingly to generate the initial tunnel model.

[0038] In one implementation of the present invention, radar wave detection data of underground space within the construction area is collected and preprocessed to obtain a standardized three-dimensional data matrix; a convolutional neural network (CNN) model for processing the three-dimensional data matrix is ​​constructed, wherein the CNN model integrates a Dropout layer. This invention uses a CNN model to perform N forward inferences on a three-dimensional data matrix. During each inference, a Dropout layer is enabled to randomly deactivate some neurons, thereby obtaining N sets of differential prediction results regarding the tunnel's location and contour. Based on the N sets of differential prediction results, the probability value of each coordinate point in the underground space belonging to the tunnel structure is calculated, generating a probability distribution model of the tunnel's three-dimensional space. According to the probability distribution model, the spatial region with a probability value greater than or equal to a first preset threshold is determined as the tunnel's core region. Based on the tunnel's core region, the three-dimensional reference coordinates and three-dimensional contour surface of the tunnel are determined, and the initial tunnel model is output. Preferably, the present invention calculates the dispersion of N sets of differentiated prediction results to quantify the uncertainty of tunnel location and contour; if the dispersion exceeds a preset standard, a data re-acquisition command is triggered.

[0039] This invention calculates the frequency at which each coordinate point in the underground space is identified as belonging to a tunnel structure in N sets of prediction results, and uses this frequency as the probability value of the coordinate point belonging to the tunnel structure. The probability distribution model divides the space into: a tunnel core area with a probability value greater than or equal to a first preset threshold, a contour boundary transition area with a probability value between a second preset threshold and the first preset threshold, and a non-tunnel area with a probability value less than the second preset threshold.

[0040] As can be seen, this invention fits the initial tunnel model with the dynamic construction structure model to obtain a combined model, calculates the minimum spatial distance between each pile foundation and the tunnel structure, and combines the geological mechanics parameters and the tunnel safety protection range standard to determine whether the pile foundation is within the tunnel's safety range. Based on the judgment results, it outputs risk warnings and adjustment suggestions to guide pile foundation construction. This method, by dynamically fitting the tunnel and pile foundation models and performing real-time analysis, can assess the safety risks of the tunnel adjacent to the pile foundation construction, significantly improving the safety and controllability of the construction process.

[0041] The modeling method of the present invention further includes: after the pile foundation construction is completed, a track inspection trolley is used to continuously or periodically monitor the deformation inside the tunnel to obtain tunnel deformation data; the measured deformation data is fed back to the BIM model, and the combined model is analyzed and optimized.

[0042] In summary, this invention enables dynamic updating of the geological structure model and improves the model's response speed to construction disturbances, providing real-time risk prediction capabilities for construction safety. Furthermore, this invention fits the initial tunnel model with the dynamic construction structure model to obtain a combined model, calculates the minimum spatial distance between each pile foundation and the tunnel structure, and, combined with geological mechanics parameters and tunnel safety protection range standards, determines whether the pile foundation is within the tunnel's safety range. Based on the judgment results, it outputs risk warnings and adjustment suggestions to guide pile foundation construction. This method, through dynamic fitting of the tunnel and pile foundation models and real-time analysis, can assess the safety risks of adjacent tunnels during pile foundation construction, significantly improving the safety and controllability of the construction process.

[0043] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0044] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0045] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0046] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for modeling piles around tunnels based on multi-source geological data fusion, characterized in that, include: Multi-source geological data is accessed and integrated through BIM software API, and a hierarchical topology node system is constructed based on the multi-source geological data, including benchmark nodes, core nodes and associated nodes. Core nodes and benchmark nodes are bidirectionally connected, and associated nodes and core nodes are unidirectionally connected. Extract the three-dimensional coordinates of each node, calculate the spatial distance between nodes with connection relationships, construct the initial edge weights by combining the geological influence coefficient, form the initial topology network, embed the initial topology network into the BIM three-dimensional environment, and obtain the initial structural model by pre-training the location-aware dynamic graph neural network. Based on the pile foundation construction sequence, the incremental nodes corresponding to the newly added pile positions are activated in batches. The pile foundation information during the construction process is obtained in real time through the BIM platform, the position and attributes of the newly added nodes are updated, and only the incremental nodes are connected to the existing core nodes of the corresponding level. The relationship between the nodes is visualized and marked in the BIM model. According to the construction disturbance impact factor and data credibility, the weight of the edges between the affected nodes is dynamically adjusted, and the topology network is updated in real time. For the activated incremental node and its topological neighborhood range, local incremental reasoning is performed to fuse the features of the neighborhood node, and uncertainty quantification assessment is performed based on probabilistic reasoning method, thereby dynamically updating and correcting the initial structural model to form a construction dynamic structural model. The initial tunnel model is fitted with the construction dynamic structural model to obtain a combined model. The minimum spatial distance between each pile foundation and the tunnel structure is calculated. Combined with the stratum mechanical parameters and the tunnel safety protection range standard, it is determined whether the pile foundation is within the safe range of the tunnel. Based on the judgment results, risk warnings and adjustment suggestions are output to guide the pile foundation construction.

2. The method for modeling tunnel perimeter piles based on multi-source geological data fusion according to claim 1, characterized in that, The initial tunnel model is obtained through the following steps: Radar wave data of the underground tunnel area is acquired, and the Monte Carlo Dropout method is used to perform multiple forward inferences and uncertainty quantifications on the preset neural network to construct a probability distribution model of the tunnel's location and outline. Based on this model, the three-dimensional location and outline of the tunnel are determined, and an initial tunnel model is generated.

3. The method for modeling tunnel perimeter piles based on multi-source geological data fusion according to claim 2, characterized in that, Also includes: After the pile foundation construction is completed, a track inspection trolley is used to continuously or periodically monitor the deformation inside the tunnel to obtain tunnel deformation data. The measured deformation data is fed back to the BIM model, and the combined model is analyzed and optimized.

4. The method for modeling tunnel perimeter piles based on multi-source geological data fusion according to claim 3, characterized in that, The step of forming the initial topology network includes: Identify node pairs with connections, including connections between core nodes and base nodes, and connections between core nodes and associated nodes; The spatial distance between node pairs is calculated using the three-dimensional Euclidean distance formula, expressed as: in, and These are the three-dimensional coordinates of the two nodes in the node pair; The first geological influence coefficient is determined based on the node pair type and geological attributes. With the second geological influence coefficient : The connection between the core node and the baseline node is determined based on the geological risk level corresponding to the core node. ; The connection between core nodes and associated nodes is determined based on their geological connectivity. ; The initial boundary weight W is calculated using spatial distance and geological influence coefficient, and is expressed as: In the formula, K is the corresponding geological influence coefficient. ;d represents the three-dimensional Euclidean distance between node pairs, and the edge weight calculation results are normalized to the range of 0-1.

5. The method for modeling tunnel perimeter piles based on multi-source geological data fusion according to claim 4, characterized in that, The step of pre-training a location-aware dynamic graph neural network to obtain the initial structural model before construction includes: A location encoding module for a geological scene is constructed to generate a fused location feature vector for each node. The location feature vector is composed of absolute location encoding, relative location encoding and geological layer location encoding. A dual attention fusion layer based on location and geology is constructed. Based on the location feature vector and the geological attribute feature vector of the node, the location attention coefficient and the geological attention coefficient are calculated respectively, and the two coefficients are fused to obtain the association attention coefficient. The initial edge weights are enhanced based on the fused association attention coefficient. The positional attention coefficient is calculated based on the positional feature vector and is expressed as follows: In the formula, This represents the positional attention coefficient between node i and node j. The feature vector representing the position of node i. This represents the feature vector representing the position of node j. The dimension of the location feature vector. Represents the normalization function; The geological attention coefficient is calculated based on the geological attribute feature vector of a node, and is expressed as: In the formula, This represents the geological attribute feature vector of node i. This represents the geological attribute feature vector of node j. The dimension representing the geological attribute feature vector; This represents the geological attention coefficient between node i and node j; The enhanced edge weights and node features are input into a dynamic neural network and trained using a dynamic learning strategy to obtain the initial structural model before construction.

6. The method for modeling tunnel perimeter piles based on multi-source geological data fusion according to claim 5, characterized in that, The steps for training using a dynamic learning strategy include: Based on the geological influence coefficient K and the location attention coefficient of the node pair and distance factor Calculate the geological correlation complexity score , is represented as: ,in, For distance-based normalization factor, d is the node distance; The training samples are divided into three levels—simple, medium, and complex—based on a score S, and then input into the model for training in stages. A dual-trigger adaptive learning rate adjustment mechanism is adopted to adjust the learning rate of the position encoding and attention modules; A multi-objective balanced loss function is constructed, and the total loss is composed of a weighted average of feature reconstruction loss, location uncertainty loss, and attention consistency loss.

7. The method for modeling tunnel perimeter piles based on multi-source geological data fusion according to claim 6, characterized in that, The step of activating the incremental nodes corresponding to newly added pile positions in batches according to the pile foundation construction sequence includes: Establish the mapping relationship between processes and incremental nodes, and determine the activation trigger conditions for each incremental node; Once activated, each incremental node establishes a connection only with existing nodes within its spatial neighborhood.

8. The method for modeling tunnel perimeter piles based on multi-source geological data fusion according to any one of claims 2 to 7, characterized in that, The step of performing local incremental inference for the activated incremental node and its topological neighborhood includes: Based on the type of incremental node and its geological risk level, the neighborhood radius of its local inference is dynamically determined, and existing nodes within the spatial range are selected based on the neighborhood radius to form a local inference subgraph. An improved graph SAGE convolution algorithm is used to reason about the local reasoning subgraph and aggregate neighborhood information; wherein, different attention weights are assigned to different types of neighborhood nodes in the subgraph according to the node level and data credibility. The features of the neighboring nodes are divided into spatial features, geological attribute features and dynamic association features, and then weighted and summed according to preset semantic weights to generate the fusion feature vector of the incremental node. The fused feature vector is updated to the corresponding incremental node in the initial structural model, and the relevant edge weights within the local inference subgraph are adjusted.

9. The method for modeling tunnel perimeter piles based on multi-source geological data fusion according to claim 8, characterized in that, The step of forming a local reasoning subgraph includes: Based on the geological parameters and construction conditions associated with incremental nodes, a two-level risk classification is performed, dividing the risk into high risk, medium risk, and low risk; based on the incremental node type and geological risk level, the neighborhood radius is determined by mapping. Based on the neighborhood radius, a multi-condition neighborhood node screening is performed, including but not limited to spatial distance screening, hierarchical constraint screening, and validity screening, to obtain the screened existing nodes; the incremental nodes and the screened existing nodes are used as subgraph nodes, and their original edge connection relationships are retained to construct a local reasoning subgraph.

10. The method for modeling tunnel perimeter piles based on multi-source geological data fusion according to claim 9, characterized in that, The step of using an improved graph SAGE convolution algorithm to perform inference on the local inference subgraph and aggregating neighborhood information includes: Based on the hierarchical type of the neighboring nodes in the local inference subgraph, assign differentiated hierarchical aggregation weights to them. ; When aggregating the features of neighboring nodes, the updated dynamic edge weights are used as the basis. The aggregation process, which involves weighting the features of neighboring nodes, can be represented as follows: Where Z is the normalization coefficient. Weights are retained for the incremental node's own characteristics. The hierarchical aggregation weight corresponding to the neighboring node u. Let b be the convolution weight matrix, and b be the bias term. It is a non-linear activation function; The characteristics of the neighboring node u, This represents the feature vector of node v in the previous construction stage. This represents the aggregated feature vector of node v in the current construction phase; This represents the dynamic edge weight between node u and node v; For neighboring nodes with the geological attribute of karst caves, the erosion features representing the degree of erosion are embedded into the vector, and then fused with the original node features with preset weights before participating in the aggregation. For dynamic neighborhood nodes that characterize the state of construction disturbance, the dynamic feature vectors constructed from the construction disturbance influence factor and the data credibility factor are incorporated into the original node features with preset weights and then participated in the aggregation.

Images