A foundation pit adjacent to an operating tunnel state prediction method based on a space-time decoupling network

Abstract

The application discloses a kind of based on space-time decoupling network's foundation pit adjacent operating tunnel state prediction method, comprising: obtaining historical tunnel change data and foundation pit real-time monitoring data, corresponding foundation pit feature data is extracted for historical tunnel change time, and the key foundation pit feature set with significant causal relationship with tunnel state change is screened out.Based on the key foundation pit features screened, historical tunnel change data and the space relationship and stress transmission relationship between foundation pit-tunnel, construct time series graph network set.Establish space-time decoupling network model, train using time series graph network set, learn the complex space-time coupling characteristics between foundation pit and tunnel, obtain the trained prediction model.Input the key foundation pit feature data of foundation pit current time, realize the prediction of adjacent operating tunnel future several time deformations, convergences and displacements.The application can effectively capture the dynamic influence of foundation pit construction disturbance on tunnel, improve the accuracy and real-time performance of tunnel safety state prediction.

Description

Technical Field

[0001] This invention belongs to the field of underground engineering safety monitoring and intelligent prediction technology, specifically involving a method for predicting the state of an operational tunnel near a foundation pit based on a spatiotemporal decoupling network. Background Technology

[0002] With the acceleration of urbanization, the development and utilization of underground space is becoming increasingly widespread. Urban rail transit systems (subways), as a crucial component of urban infrastructure, are experiencing continuous growth in line length and station density. When large-scale foundation pit excavation projects are carried out in urban centers or densely built-up areas, factors such as excavation disturbance, support deformation, and changes in groundwater levels can have coupled effects on adjacent operational subway tunnels, manifesting as stress redistribution in the tunnel surrounding rock, deformation of the tunnel retaining structure, convergence, and displacement. Failure to identify and predict these effects in a timely and accurate manner may lead to damage to tunnel function, operational restrictions, and even serious safety hazards. Therefore, effectively monitoring and predicting the impact of foundation pit construction on adjacent operational tunnels has become a crucial issue in the safety management of urban underground engineering.

[0003] Currently, engineering practices addressing the interaction between foundation pits and tunnels primarily rely on the deployment of sensors and the periodic or real-time observation and monitoring of data by engineering technicians, followed by experience-based assessments and decisions. Engineers typically implement control or reinforcement measures based on comparisons between monitored values ​​and preset thresholds, combined with on-site investigations and engineering experience. While this manual monitoring and experience-based judgment method remains indispensable in engineering practice, it has significant shortcomings: First, manual judgment has limited efficiency, making it difficult to achieve high-frequency, low-latency risk identification in large-scale, multi-sensor scenarios; second, engineering judgments are subjective, making it difficult to maintain consistency among different personnel or projects; and third, traditional experience struggles to reveal the complex spatiotemporal coupling mechanism between foundation pit excavation activities and tunnel condition changes, failing to provide quantitative causal evidence for construction control.

[0004] To overcome the limitations of manual methods, advancements in information technology, sensing technology, and computational methods in recent years have driven the application of data-driven approaches in underground engineering. Numerous studies and engineering practices have employed various machine learning algorithms (such as support vector machines, traditional neural networks, and decision trees) and statistical modeling techniques to model and predict monitoring data. These methods have achieved positive results in improving prediction efficiency and, to some extent, accuracy. Furthermore, with the development of graph neural networks, attention mechanisms, and sequence models, some research has begun to attempt to establish spatiotemporal models capable of capturing the relationship between spatial structure and temporal evolution, in order to characterize the diffusion patterns of influences under spatial adjacency or mechanical transmission. For example, in the prior art, Chinese patent CN119202583A discloses a method for predicting foundation pit data based on a spatiotemporal graph convolutional network, including the following steps: acquiring historical foundation pit data; processing the historical foundation pit data to obtain a training dataset; constructing an initial foundation pit data prediction model with a dual-temporal structure based on a graph neural network; training the initial foundation pit data prediction model using the training dataset to obtain the foundation pit data prediction model; and using the foundation pit data prediction model to perform actual foundation pit data prediction.

[0005] However, this method has the following limitations: 1. There are many types of monitoring variables for foundation pits and tunnels. Using all of them directly will increase sensing and computing costs and lead to noise accumulation. 2. Although most existing methods use graph neural networks and time series models, they often directly couple space and time in the modeling, without clearly decomposing long-term trends (and short-term fluctuations). Therefore, it is difficult to explain and utilize the different effects of the two types of dynamics on the tunnel state, and the characterization of delay effects and cumulative effects is also insufficient.

[0006] In summary, existing technologies still have many shortcomings in addressing the impact of foundation pit excavation on adjacent operational tunnels. There is an urgent need for new technologies and methods that can effectively capture spatiotemporal coupling effects, accurately predict tunnel status changes in real time, and have good interpretability, so as to meet the pressing needs for tunnel safety assurance in urban underground engineering construction. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for predicting the state of an operational tunnel near a foundation pit based on a spatiotemporal decoupled network.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] This invention provides a method for predicting the state of an operational tunnel near a foundation pit based on a spatiotemporal decoupled network, comprising the following steps:

[0010] Acquire historical tunnel change data, which includes deformation, convergence, and displacement data of each preset tunnel node and the corresponding change time.

[0011] Real-time monitoring of the foundation pit to obtain foundation pit change data, the foundation pit change data including multiple foundation pit characteristics;

[0012] Based on each moment of change in the historical tunnel change data, obtain the corresponding foundation pit characteristic data;

[0013] Based on historical tunnel change data and corresponding foundation pit feature data, the foundation pit features are filtered using the DirectLiNGAM causal interpretation method to obtain a set of key foundation pit features for modeling.

[0014] Based on the key foundation pit feature set, historical tunnel change data, spatial relationship and force transmission relationship between foundation pit and tunnel, a time series graph network set is constructed.

[0015] A spatiotemporal decoupling network based on graph convolutional networks and Transformer architecture is constructed. The spatiotemporal decoupling network is trained based on a set of temporal graph networks to obtain the trained spatiotemporal decoupling network.

[0016] The system acquires key foundation pit feature data at the current moment, and uses this data, along with a trained spatiotemporal decoupled network, to predict the status of nearby operational tunnels.

[0017] Furthermore, the historical tunnel change data is threshold exceeding event data obtained by filtering conventional monitoring data according to a preset threshold. Specifically, the historical tunnel change data only includes monitoring records when the absolute value of the deformation value, convergence value, or displacement value of any preset tunnel node exceeds the corresponding preset threshold, and the change time corresponding to the monitoring record.

[0018] Furthermore, the characteristics of the foundation pit include support structure displacement, support structure deformation, anchor bolt tension, support structure stress, grouting pressure and grouting volume, excavation depth, surface settlement, lateral displacement, groundwater level, pore water pressure, soil stress, soil strain, and changes in temporary loads or surrounding loads.

[0019] Furthermore, the step of obtaining corresponding foundation pit feature data based on each change moment in the historical tunnel change data specifically includes:

[0020] Starting from each moment of change in the historical tunnel change data, the foundation pit change data within a preset time window is collected. The foundation pit feature values ​​in the foundation pit change data within the preset time window are sorted. For each foundation pit feature, its maximum value within the preset time window is selected as the foundation pit feature data for each moment of change in the historical tunnel change data.

[0021] Furthermore, based on historical tunnel change data and corresponding foundation pit feature data, the foundation pit features are filtered using the DirectLiNGAM causal interpretation method to obtain a set of key foundation pit features for modeling, specifically including:

[0022] Historical tunnel change data and corresponding foundation pit characteristic data are organized into a sample set. :

[0023]

[0024] in, Indicates the first i The feature vector of the foundation pit at each moment corresponds to a change moment in the historical tunnel change data. This represents the total number of changes in historical tunnel data, including the total number of times when changes occurred in the historical tunnels. Indicates the first i At the moment, the _ Individual foundation pit characteristic values, Indicates the total number of features of the foundation pit; Indicates the first i The tunnel change vector at each time step. Indicates the first i At the [time]th moment q Deformation, convergence, and displacement values ​​of each tunnel node; q Indicates the total number of preset tunnel nodes;

[0025] All The feature matrix is ​​formed by stacking rows. ; All of The response matrix is ​​formed by stacking rows. To obtain the joint observation matrix ;

[0026] For joint observation matrix Causal structure learning is performed using the DirectLiNGAM linear non-Gaussian acyclic model to establish a linear causal model:

[0027]

[0028] in, A column vector of random variables, corresponding to the joint observation matrix. Column variables; This is the direct effect coefficient matrix; The perturbation term consists of independent components that follow a non-Gaussian distribution;

[0029] The matrix is ​​obtained by minimizing the residual independence criterion. The estimated value and ensure the matrix Satisfies the acyclicity constraint;

[0030] Based on the estimated direct effects matrix Calculate the total causal effect matrix among variables. :

[0031]

[0032] in, It is the identity matrix. This is the total causal effect matrix;

[0033] For each foundation pit feature variable index Calculate its index set for all tunnel response variables. Overall causal influence:

[0034]

[0035] in, Indicates the first j The total causal influence of each foundation pit feature; Represents the total causal effect matrix The Middle Line 1 Column elements;

[0036] Based on preset screening criteria, a set of key foundation pit features is selected from all foundation pit features. :

[0037]

[0038] in, This is a preset threshold for causal influence.

[0039] Furthermore, the construction of a time-series graph network set based on the key foundation pit feature set, historical tunnel change data, spatial relationship between the foundation pit and the tunnel, and force transmission relationship specifically includes:

[0040] The pit nodes and the pre-defined tunnel nodes are defined as sets of nodes in a graph network, where the pit nodes are denoted as... The set of tunnel nodes is denoted as The set of nodes is ; q Indicates the total number of preset tunnel nodes;

[0041] Based on the spatial relationships and force transmission relationships of each node in the foundation pit and tunnel, a graph network edge set is established. With edge weight set W ;

[0042] For each moment of change Construct a graph network corresponding to each time point, with foundation pit nodes. Node characteristics at each time of change Let the corresponding key foundation pit feature vector be denoted as:

[0043]

[0044] in, Represents a node At any moment t Node characteristics; Indicates the first The moment of the first Key foundation pit characteristic values, This represents the total number of key foundation pit features after screening.

[0045] For each tunnel node The node characteristics of a tunnel node are the corresponding tunnel change data:

[0046]

[0047] in, ; Represents a node At any moment t Node characteristics, Indicates the first At the [time]th moment i Deformation, convergence, and displacement values ​​of each tunnel node;

[0048] For each moment of change t , set of nodes Edge set and the set of node features at that moment Combination constitutes time t Graph networks:

[0049]

[0050] in, Indicates the time of change in the historical tunnel. t The corresponding graph network;

[0051] For all moments of change Construct the corresponding graph network in sequence. And form a time sequence graph network set. :

[0052]

[0053] in, N This represents the total number of change moments in the historical tunnel change data.

[0054] Furthermore, the graph network edge set is established based on the spatial relationship and force transmission relationship between various nodes of the foundation pit and the tunnel. With edge weight set W Specifically, it includes:

[0055] For foundation pit nodes With each tunnel node Establish edges The edge weights are determined based on the spatial distance and force transmission relationship. The calculation formula is:

[0056]

[0057] in, For foundation pit nodes With tunnel nodes Edge weights between them This is the weight normalization coefficient, used to adjust the numerical scale of the overall weights; This is the distance attenuation coefficient, used to control the rate attenuation of the influence intensity due to spatial distance; For foundation pit nodes With tunnel nodes The straight-line distance between them; For the foundation pit node to the tunnel node The force transmission coefficient is used to characterize the intensity of the mechanical influence transmitted from the foundation pit disturbance to the tunnel node;

[0058] For any two tunnel nodes If there is a spatial proximity or force transmission path between two nodes, then an edge is established. Its edge weight The calculation formula is:

[0059]

[0060] in, For tunnel nodes With tunnel nodes Edge weights between them; The weight normalization coefficient between tunnel nodes; This is the distance attenuation coefficient between tunnel nodes; For tunnel nodes and Spatial distance; For tunnel nodes The force transmission coefficient between tunnel structures is used to characterize the degree of force coupling between tunnel structures.

[0061] All established edges and their corresponding edge weights Forming edge sets With edge weight set W ,in, If node If there is no connection between them, then let .

[0062] Furthermore, the spatiotemporal decoupling network includes a graph convolutional network, a time encoder based on a self-attention mechanism, a feature decoupling module, and a prediction output module;

[0063] The temporal encoder includes a multi-head attention layer and a feedforward mapping layer, which are used to model the temporal spatial features output by the graph convolutional network module in the temporal dimension.

[0064] The feature decoupling module includes a trend extraction unit and a fluctuation extraction unit, which are used to decompose the comprehensive features output by the time encoder module: the trend extraction unit uses smoothing filtering and recursive structure to extract long-term change features, which are used to characterize the continuous deformation trend caused by the foundation pit construction process; the fluctuation extraction unit uses residual extraction and short-term convolution structure to extract high-frequency change features, which are used to characterize short-term fluctuations caused by environmental disturbances or transient loads.

[0065] The prediction output module includes a feature fusion layer and an output layer, which are used to fuse trend features and fluctuation features, and output the predicted state of each tunnel node at several future times through a fully connected mapping.

[0066] Furthermore, the step of training the spatiotemporal decoupling network based on a temporal graph network ensemble to obtain the trained spatiotemporal decoupling network specifically includes:

[0067] According to the preset sliding window length and prediction time domain, the temporal graph network set is divided into several training samples by the sliding window. Each training sample is input by the node feature sequence of the graph network at several consecutive time steps and the tunnel node feature sequence at several subsequent time steps as the prediction target. The sliding window length and prediction time domain are adjustable hyperparameters.

[0068] The node feature sequences of each training sample are uniformly processed, including time alignment, missing value imputation, numerical normalization, and dimensionality mapping. The key features of the foundation pit and the deformation / convergence / displacement features of each tunnel node are organized in node order to form the features at each time step. t The node feature matrix;

[0069] For each time step in the input sequence t Edge sets in graph networks With edge weight set W A graph convolutional network from the spatiotemporal decoupling network is used for each time step. t The node feature matrix is ​​aggregated and transformed using neighborhood information to obtain the node feature matrix at each time step. tSpatial representation of each node;

[0070] The spatial representations of each node at each time point are arranged along the time axis to form a time series representation. A time encoder based on a self-attention mechanism is used to extract time features from the time series representation, capture the dynamic dependencies between different time steps, and obtain a node-level spatiotemporal feature representation containing time series information.

[0071] The feature decoupling module separates the trend component and fluctuation component of the node-level spatiotemporal feature representation, and independently models the long-term change trend and short-term disturbance respectively. The trend component is used to capture the overall deformation trend caused by the excavation of the foundation pit and the change of the support structure, while the fluctuation component is used to characterize the local time-varying response under the influence of load changes or environmental disturbances.

[0072] The trend component and the fluctuation component are integrated in the feature fusion layer to form a complete node-level spatiotemporal feature representation, which is then input into the prediction output module. The prediction output module maps the node-level spatiotemporal features into the deformation, convergence and displacement prediction values ​​of each tunnel node at several future times through a fully connected mapping.

[0073] A loss function is constructed based on the difference between the predicted and the true values. The parameters of the spatiotemporal decoupling network are iteratively optimized to obtain the trained spatiotemporal decoupling network.

[0074] Furthermore, the loss function is formulated as follows:

[0075]

[0076] in, For loss function, This is the set of training samples obtained by dividing the data using a sliding window. The total number of training samples, To predict the length of the time domain, Indicates the total number of preset tunnel nodes; The output dimension for each tunnel node includes deformation, convergence, and displacement; Represents a single training sample. Given the input node feature sequence, The corresponding tunnel node feature sequence; The model predicts the value, representing the first... At this moment i The tunnel node, the first u Prediction results for each output dimension; Indicates the first At this moment i The tunnel node, the first u The actual measurement value of each output dimension.

[0077] Compared with the prior art, the present invention has the following advantages:

[0078] (1) Existing technologies typically use all routine monitoring data for modeling, resulting in the model being interfered with by a large number of harmless routine fluctuations during training. This leads to insufficient sensitivity of the model to threshold exceedance events that are truly significant for engineering purposes, making it difficult to provide reliable early warnings in critical risk scenarios. Therefore, this invention employs a historical tunnel change data screening approach based on threshold exceedance events in sample construction. Specifically, the historical tunnel change data only includes monitoring records exceeding a preset threshold, focusing the training / validation samples on hazardous events of engineering concern. After adopting this guiding sample construction, the model's learning signal under critical risk scenarios is significantly enhanced, improving the model's detection sensitivity and early warning reliability for exceedance scenarios, thereby reducing false alarms and false negatives, and improving the timeliness and safety of engineering decisions.

[0079] (2) Existing technologies have key problems in handling the time lag relationship between foundation pit monitoring data and tunnel response: the time lag between foundation pit disturbance and observable changes in the tunnel is usually uncertain and variable; in actual engineering, it is difficult to accurately determine which moment of foundation pit data will ultimately induce subsequent tunnel deformation, so directly using the instantaneous value or average value at the corresponding moment as a feature can easily miss the real trigger signal or be misled by short-term noise, resulting in insufficient capture of causal relationships and decreased prediction performance of the model. To solve this problem, this invention adopts a maximum value aggregation strategy based on a preset time window in sample construction and feature extraction, that is, the maximum value of each foundation pit feature is taken as the foundation pit feature value at the moment of change within the window; when the trigger lag cannot be accurately known, the maximum value within the window can be used as a representative feature to capture the strongest disturbance or extreme excitation event that occurs within the lag interval. Such extreme values ​​are most likely to become the direct cause of tunnel response; at the same time, maximum value aggregation naturally filters out most low-amplitude fluctuations and random noise within the window, retaining the most influential information in engineering. It significantly improves the ability to identify delayed triggering events, enabling the model to more stably correlate foundation pit disturbances with subsequent tunnel changes, and enhances the sensitivity and accuracy of predicting over-limit events.

[0080] (3) The present invention uses the DirectLiNGAM causal interpretation method to screen the features of the foundation pit and form a key foundation pit feature set. By introducing the feature set verified by causal discovery, the model can use causally related inputs as driving variables, which can significantly improve the engineering interpretability of the prediction results and reduce the risk of redundancy and collinearity, thereby enhancing the robustness and generalizability of the model under different construction conditions.

[0081] (4) Existing spatiotemporal models often establish relationships between nodes based on experience or simple topology, neglecting the differences in physical distance and force transmission between the foundation pit and tunnel and between tunnels. This results in a coarse depiction of spatial propagation relationships and makes it difficult to reflect the actual engineering process in terms of propagation paths and intensity. To address this, this invention clarifies the edge connection rules between foundation pit nodes and tunnel nodes, and between tunnel nodes, when constructing the graph network, forming a time-series graph network set, which is then used as the model input. Based on the time-series graph constructed according to time intervals, the model can gradually depict the propagation and accumulation of disturbances on the topology over time, accurately reflecting the delay and accumulation effects, making the prediction closer to the propagation laws of engineering mechanics on site, and facilitating the location of the source of influence and the propagation path.

[0082] (5) In traditional graph networks, edge weights are usually based on a single distance or empirical value, which is difficult to reflect both spatial decay and engineering force transmission characteristics simultaneously. To address this, this invention proposes a specific edge weight construction method, which multiplies the exponential decay of physical distance by the force transmission coefficient to determine the edge weight. With this composite weight, the graph structure is no longer an abstraction of pure topological relationships, but rather a representation that takes into account the influence intensity of spatial geometry and mechanical coupling. This allows graph convolution to be weighted according to the engineering influence intensity when aggregating neighborhoods, thereby improving the physical relevance of the spatial representation and the physical reliability of the prediction, which is beneficial for engineers to understand the model output.

[0083] (6) Existing spatiotemporal modeling often directly couples space and time, making it difficult to distinguish between long-term trends caused by foundation pit construction and short-term environmental disturbances, thus reducing the interpretability of predictions and the ability to model responses at different scales. In this invention, trend units and fluctuation units are extracted separately in the decoupling module, enabling the model to model two types of temporal components from different physical sources, thereby improving prediction accuracy and enhancing the interpretability of the output's engineering semantics. This facilitates the formulation of construction schedules based on trends and the implementation of short-term emergency measures based on fluctuations. Attached Figure Description

[0084] Figure 1 This is a flowchart of the tunnel state prediction method according to an embodiment of the present invention;

[0085] Figure 2 This is a diagram of the spatiotemporal decoupling network model according to an embodiment of the present invention;

[0086] Figure 3 This is a schematic diagram of a tunnel state prediction system according to an embodiment of the present invention. Detailed Implementation

[0087] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0088] Example 1:

[0089] This embodiment provides a method for predicting the state of an operational tunnel near a foundation pit based on a spatiotemporal decoupled network, such as... Figure 1 As shown, it includes the following steps:

[0090] Step S1: Obtain historical tunnel change data, which includes deformation, convergence, and displacement data of each preset tunnel node and the corresponding change time.

[0091] The historical tunnel change data consists of threshold-exceeding event data obtained by filtering conventional monitoring data according to preset thresholds. Specifically, the historical tunnel change data only includes monitoring records when the absolute value of the deformation, convergence, or displacement value of any preset tunnel node exceeds the corresponding preset threshold, along with the corresponding change time of the monitoring record. The deformation, convergence, and displacement data of the tunnel node refer to the following respectively: deformation data is the deformation of the monitoring point on the tunnel structure surface or lining along a specific direction; convergence data is the change in distance between two monitoring points within the tunnel cross-section, used to reflect the deformation trend of the tunnel cross-section; and displacement data is the spatial displacement of the tunnel node relative to the initial reference position, used to characterize the overall movement or settlement.

[0092] Step S2: Real-time monitoring of the foundation pit to obtain foundation pit change data, which includes multiple foundation pit characteristics; among which, the foundation pit characteristics include the following categories: support structure displacement, support structure deformation, anchor bolt tension, support structure stress, grouting pressure and grouting volume, excavation depth, surface settlement, lateral displacement, groundwater level, pore water pressure, soil stress, soil strain, and changes in temporary loads or surrounding loads.

[0093] Step S3: Based on each change moment in the historical tunnel change data, obtain the corresponding foundation pit characteristic data, specifically including:

[0094] Starting from each moment of change in the historical tunnel change data, the foundation pit change data within a preset time window is collected. The foundation pit feature values ​​in the foundation pit change data within the preset time window are sorted. For each foundation pit feature, its maximum value within the preset time window is selected as the foundation pit feature data for each moment of change in the historical tunnel change data.

[0095] Step S4: Based on historical tunnel change data and corresponding foundation pit feature data, the foundation pit features are filtered using the DirectLiNGAM causal interpretation method to obtain a set of key foundation pit features for modeling, specifically including:

[0096] Historical tunnel change data and corresponding foundation pit characteristic data are organized into a sample set. :

[0097]

[0098] in, Indicates the first i The feature vector of the foundation pit at each moment corresponds to a change moment in the historical tunnel change data. This represents the total number of changes in historical tunnel data, including the total number of times when changes occurred in the historical tunnels. Indicates the first i At the moment, the _ Individual foundation pit characteristic values, Indicates the total number of features of the foundation pit; Indicates the first i The tunnel change vector at each time step. Indicates the first i At the [time]th moment q Deformation, convergence, and displacement values ​​of each tunnel node; q Indicates the total number of preset tunnel nodes;

[0099] All The feature matrix is ​​formed by stacking rows. ; All of The response matrix is ​​formed by stacking rows. To obtain the joint observation matrix ;

[0100] For joint observation matrix Causal structure learning is performed using the DirectLiNGAM linear non-Gaussian acyclic model to establish a linear causal model: ;in, A column vector of random variables, corresponding to the joint observation matrix. Column variables; This is the direct effect coefficient matrix; The perturbation term consists of independent components that follow a non-Gaussian distribution;

[0101] The matrix is ​​obtained by minimizing the residual independence criterion. The estimated value and ensure the matrix Satisfies the acyclicity constraint;

[0102] Based on the estimated direct effects matrix Calculate the total causal effect matrix among variables. :

[0103]

[0104] in, It is the identity matrix. This is the total causal effect matrix;

[0105] For each foundation pit feature variable index Calculate its index set for all tunnel response variables. Overall causal influence:

[0106]

[0107] in, Indicates the first j The total causal influence of each foundation pit feature; Represents the total causal effect matrix The Middle Line 1 Column elements;

[0108] Based on preset screening criteria, a set of key foundation pit features is selected from all foundation pit features. : ;in, This is a preset threshold for causal influence.

[0109] Step S5: Based on the key foundation pit feature set, historical tunnel change data, spatial relationship and force transmission relationship between the foundation pit and the tunnel, construct a time series graph network set, specifically including:

[0110] The pit nodes and the pre-defined tunnel nodes are defined as sets of nodes in a graph network, where the pit nodes are denoted as... The set of tunnel nodes is denoted as The set of nodes is ; q Indicates the total number of preset tunnel nodes;

[0111] Based on the spatial relationships and force transmission relationships of each node in the foundation pit and tunnel, a graph network edge set is established. With edge weight set W Specifically, it includes:

[0112] For foundation pit nodes With each tunnel node Establish edges The edge weights are determined based on the spatial distance and force transmission relationship. The calculation formula is:

[0113]

[0114] in, For foundation pit nodes With tunnel nodes Edge weights between them This is the weight normalization coefficient, used to adjust the numerical scale of the overall weights; This is the distance attenuation coefficient, used to control the rate attenuation of the influence intensity due to spatial distance; For foundation pit nodes With tunnel nodes The straight-line distance between them; For the foundation pit node to the tunnel node The force transmission coefficient is used to characterize the intensity of the mechanical influence transmitted from the foundation pit disturbance to the tunnel node;

[0115] For any two tunnel nodes If there is a spatial proximity or force transmission path between two nodes, then an edge is established. Its edge weight The calculation formula is:

[0116]

[0117] in, For tunnel nodes With tunnel nodes Edge weights between them; The weight normalization coefficient between tunnel nodes; This is the distance attenuation coefficient between tunnel nodes; For tunnel nodes and Spatial distance; For tunnel nodes The force transmission coefficient between tunnel structures is used to characterize the degree of force coupling between tunnel structures.

[0118] All established edges and their corresponding edge weights Forming edge sets With edge weight set W ,in, If node If there is no connection between them, then let .

[0119] For each moment of change Construct a graph network corresponding to each time point, with foundation pit nodes. Node characteristics at each time of change Let the corresponding key foundation pit feature vector be denoted as:

[0120]

[0121] in, Represents a node At any moment t Node characteristics; Indicates the first The moment of the first Key foundation pit characteristic values, This represents the total number of key foundation pit features after screening.

[0122] For each tunnel node The node characteristics of a tunnel node are the corresponding tunnel change data:

[0123]

[0124] in, ; Represents a node At any moment t Node characteristics, Indicates the first At the [time]th moment i Deformation, convergence, and displacement values ​​of each tunnel node;

[0125] For each moment of change t , set of nodes Edge set and the set of node features at that moment Combination constitutes time t Graph networks:

[0126]

[0127] in, Indicates the time of change in the historical tunnel. t The corresponding graph network;

[0128] For all moments of change Construct the corresponding graph network in sequence. And form a time sequence graph network set. :

[0129]

[0130] in, N This represents the total number of change moments in the historical tunnel change data.

[0131] Step S6: Construct a spatiotemporal decoupling network based on graph convolutional networks and Transformer architecture, train the spatiotemporal decoupling network based on a set of temporal graph networks, and obtain the trained spatiotemporal decoupling network;

[0132] Among them, such as Figure 2 As shown, the spatiotemporal decoupling network includes a graph convolutional network, a temporal encoder based on a self-attention mechanism, a feature decoupling module, and a prediction output module;

[0133] The temporal encoder includes a multi-head attention layer and a feedforward mapping layer, which are used to model the temporal spatial features output by the graph convolutional network module in the temporal dimension.

[0134] The feature decoupling module includes a trend extraction unit and a fluctuation extraction unit, which are used to decompose the comprehensive features output by the time encoder module: the trend extraction unit uses smoothing filtering and recursive structure to extract long-term variation features, which are used to characterize the continuous deformation trend caused by the foundation pit construction process; the fluctuation extraction unit uses residual extraction and short-term convolution structure to extract high-frequency variation features, which are used to characterize short-term fluctuations caused by environmental disturbances or transient loads.

[0135] The prediction output module includes a feature fusion layer and an output layer, which are used to fuse trend features and fluctuation features, and output the predicted state of each tunnel node at several future time points through a fully connected mapping.

[0136] Step S6 specifically includes:

[0137] According to the preset sliding window length and prediction time domain, the temporal graph network set is divided into several training samples by the sliding window. Each training sample is input by the node feature sequence of the graph network at several consecutive time steps and the tunnel node feature sequence at several subsequent time steps as the prediction target. The sliding window length and prediction time domain are adjustable hyperparameters.

[0138] The node feature sequences of each training sample are uniformly processed, including time alignment, missing value imputation, numerical normalization, and dimensionality mapping. The key features of the foundation pit and the deformation / convergence / displacement features of each tunnel node are organized in node order to form the features at each time step. t The node feature matrix;

[0139] For each time step in the input sequence t Edge sets in graph networks With edge weight set W A graph convolutional network from the spatiotemporal decoupling network is used for each time step. t The node feature matrix is ​​aggregated and transformed using neighborhood information to obtain the node feature matrix at each time step. t Spatial representation of each node;

[0140] The spatial representations of each node at each time point are arranged along the time axis to form a time series representation. A time encoder based on a self-attention mechanism is used to extract time features from the time series representation, capture the dynamic dependencies between different time steps, and obtain a node-level spatiotemporal feature representation containing time series information.

[0141] The feature decoupling module separates the trend component and fluctuation component of the node-level spatiotemporal feature representation, and independently models the long-term change trend and short-term disturbance respectively. The trend component is used to capture the overall deformation trend caused by the excavation of the foundation pit and the change of the support structure, while the fluctuation component is used to characterize the local time-varying response under the influence of load changes or environmental disturbances.

[0142] The trend component and the fluctuation component are integrated in the feature fusion layer to form a complete node-level spatiotemporal feature representation, which is then input into the prediction output module. The prediction output module maps the node-level spatiotemporal features into the deformation, convergence and displacement prediction values ​​of each tunnel node at several future times through a fully connected mapping.

[0143] A loss function is constructed based on the difference between the predicted and the true values. The parameters of the spatiotemporal decoupling network are iteratively optimized to obtain the trained spatiotemporal decoupling network.

[0144] The loss function is expressed as follows:

[0145]

[0146] in, For loss function, This is the set of training samples obtained by dividing the data using a sliding window. The total number of training samples, To predict the length of the time domain, Indicates the total number of preset tunnel nodes; The output dimension for each tunnel node includes deformation, convergence, and displacement; Represents a single training sample. Given the input node feature sequence, The corresponding tunnel node feature sequence; The model predicts the value, representing the first... At this moment i The tunnel node, the first u Prediction results for each output dimension; Indicates the first At this moment i The tunnel node, the first u The actual measurement value of each output dimension.

[0147] Step S7: Obtain key foundation pit feature data at the current moment, and based on the key foundation pit feature data at the current moment and the trained spatiotemporal decoupling network, realize the prediction of the status of the operating tunnel near the foundation pit;

[0148] The key foundation pit feature data at the current moment is the feature set obtained by extracting and filtering the real-time monitoring data of the foundation pit at the current moment;

[0149] The key excavation pit feature data at the current moment is input into the trained spatiotemporal decoupling network. The network first performs feature propagation and aggregation on the spatial relationships between excavation pit nodes and tunnel nodes based on a graph convolutional structure. Then, a temporal encoder models the temporal evolution features of the input sequence. After feature decoupling and fusion, the network outputs predictions of deformation, convergence, and displacement of each tunnel node at several future time points. These predictions can be used to assess the impact of the excavation pit construction phase on adjacent operational tunnels in real time and provide a reference for on-site construction control and safety decisions.

[0150] Example 2:

[0151] This embodiment provides a system for predicting the state of an operational tunnel near a foundation pit based on a spatiotemporal decoupled network, such as... Figure 3 As shown, it includes:

[0152] The data acquisition module is used to acquire historical tunnel change data and real-time monitoring data of the foundation pit;

[0153] The feature filtering module is used to perform causal analysis and filtering of the foundation pit monitoring parameters based on historical tunnel change data and corresponding foundation pit monitoring data at the time, and to determine the set of key foundation pit features that have a significant causal relationship with the tunnel state changes.

[0154] The time sequence diagram construction module is used to construct a set of time sequence diagram networks based on the key foundation pit feature set, historical tunnel change data, and the spatial and force transmission relationships between the foundation pit and the tunnel. Each time sequence diagram network corresponds to a change moment to describe the coupling state of the foundation pit and the tunnel at that moment.

[0155] The spatiotemporal decoupling network module is used to establish a spatiotemporal decoupling network structure including a graph convolutional network, a time encoder based on a self-attention mechanism, a feature decoupling module, and a prediction output module. The graph convolutional network is used to extract spatial correlation features, the time encoder is used to extract time-dependent features, the feature decoupling module is used to separate long-term trend and short-term fluctuation features, and the prediction output module is used to fuse features and output prediction results.

[0156] The model training module is used to train the spatiotemporal decoupling network based on the temporal graph network ensemble. By minimizing the error loss function between the predicted value and the true value, the network parameters are optimized to obtain the trained spatiotemporal decoupling network model.

[0157] The state prediction module is used to acquire key foundation pit feature data at the current moment and input it into the trained spatiotemporal decoupling network model. It outputs the deformation, convergence and displacement prediction values ​​of each tunnel node at several future moments, realizing dynamic prediction and risk warning of the state of the operating tunnels near the foundation pit.

[0158] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0159] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the state of an operational tunnel near a foundation pit based on a spatiotemporal decoupled network, characterized in that, Includes the following steps: Acquire historical tunnel change data, which includes deformation, convergence, and displacement data of each preset tunnel node and the corresponding change time. Real-time monitoring of the foundation pit to obtain foundation pit change data, the foundation pit change data including multiple foundation pit characteristics; Based on each moment of change in the historical tunnel change data, obtain the corresponding foundation pit characteristic data; Based on historical tunnel change data and corresponding foundation pit feature data, the foundation pit features are filtered using the DirectLiNGAM causal interpretation method to obtain a set of key foundation pit features for modeling. Based on the key foundation pit feature set, historical tunnel change data, spatial relationship and force transmission relationship between foundation pit and tunnel, a time series graph network set is constructed. A spatiotemporal decoupling network based on graph convolutional networks and Transformer architecture is constructed. The spatiotemporal decoupling network is trained based on a set of temporal graph networks to obtain the trained spatiotemporal decoupling network. The key feature data of the foundation pit at the current moment is obtained. Based on the key feature data of the foundation pit at the current moment and the trained spatiotemporal decoupling network, the state of the operating tunnel near the foundation pit is predicted. The time-series network set is constructed based on the key foundation pit feature set, historical tunnel change data, spatial relationship between foundation pit and tunnel, and force transmission relationship, specifically including: The pit nodes and the pre-defined tunnel nodes are defined as sets of nodes in a graph network, where the pit nodes are denoted as... The set of tunnel nodes is denoted as The node set is ; q Indicates the total number of preset tunnel nodes; Based on the spatial relationships and force transmission relationships of each node in the foundation pit and tunnel, a graph network edge set is established. With edge weight set W ; For each moment of change Construct a graph network corresponding to each time point, with foundation pit nodes. Node characteristics at each time of change Let the corresponding key foundation pit feature vector be denoted as: in, Represents a node At any moment t Node characteristics; Indicates the first The moment of the first Key foundation pit characteristic values, This represents the total number of key foundation pit features after screening. For each tunnel node The node characteristics of a tunnel node are the corresponding tunnel change data: in, ; Represents a node At any moment t Node characteristics, Indicates the first At the [time]th moment i Deformation, convergence, and displacement values ​​of each tunnel node; For each moment of change t , set of nodes Edge set and the set of node features at that moment Combination constitutes time t Graph networks: in, Indicates the time of change in the historical tunnel. t The corresponding graph network; For all moments of change Construct the corresponding graph network in sequence. And form a time sequence graph network set. : in, N This represents the total number of change moments in the historical tunnel change data.

2. The method for predicting the state of an operational tunnel near a foundation pit based on a spatiotemporal decoupled network according to claim 1, characterized in that, The historical tunnel change data is threshold exceeding event data obtained by filtering conventional monitoring data according to a preset threshold. Specifically, the historical tunnel change data only includes monitoring records when the absolute value of the deformation value, convergence value, or displacement value of any preset tunnel node exceeds the corresponding preset threshold, and the change time corresponding to the monitoring record.

3. The method for predicting the state of an operational tunnel near a foundation pit based on a spatiotemporal decoupled network according to claim 1, characterized in that, The characteristics of the foundation pit include the displacement of the support structure, the deformation of the support structure, the tension of the anchor bolts, the stress of the support structure, the grouting pressure and grouting volume, the excavation depth, the surface settlement, the lateral displacement, the groundwater level, the pore water pressure, the soil stress, the soil strain, and the changes in temporary loads or surrounding loads.

4. The method for predicting the state of an operational tunnel near a foundation pit based on a spatiotemporal decoupled network according to claim 1, characterized in that, The step of obtaining corresponding foundation pit characteristic data based on each change moment in the historical tunnel change data specifically includes: Starting from each moment of change in the historical tunnel change data, the foundation pit change data within a preset time window is collected. The foundation pit feature values ​​in the foundation pit change data within the preset time window are sorted. For each foundation pit feature, its maximum value within the preset time window is selected as the foundation pit feature data for each moment of change in the historical tunnel change data.

5. The method for predicting the state of an operational tunnel near a foundation pit based on a spatiotemporal decoupled network according to claim 1, characterized in that, Based on historical tunnel change data and corresponding foundation pit feature data, the foundation pit features are filtered using the DirectLiNGAM causal interpretation method to obtain a set of key foundation pit features for modeling, specifically including: Historical tunnel change data and corresponding foundation pit characteristic data are organized into a sample set. : in, Indicates the first i The feature vector of the foundation pit at each moment corresponds to a change moment in the historical tunnel change data. This represents the total number of changes in historical tunnel data, including the total number of times when changes occurred in the historical tunnels. Indicates the first i At the moment, the _ Individual foundation pit characteristic values, Indicates the total number of features of the foundation pit; Indicates the first i The tunnel change vector at each time step. Indicates the first i At the [time]th moment q Deformation, convergence, and displacement values ​​of each tunnel node; q Indicates the total number of preset tunnel nodes; All The feature matrix is ​​formed by stacking rows. ; All of The response matrix is ​​formed by stacking rows. To obtain the joint observation matrix ; For joint observation matrix Causal structure learning is performed using the DirectLiNGAM linear non-Gaussian acyclic model to establish a linear causal model: in, A column vector of random variables, corresponding to the joint observation matrix. Column variables; This is the direct effect coefficient matrix; The perturbation term consists of independent components that follow a non-Gaussian distribution; The matrix is ​​obtained by minimizing the residual independence criterion. The estimated value and ensure the matrix Satisfies the acyclic constraint; Based on the estimated direct effects matrix Calculate the total causal effect matrix among variables. : in, It is the identity matrix. This is the total causal effect matrix; For each foundation pit feature variable index Calculate its index set for all tunnel response variables. Overall causal influence: in, Indicates the first j The total causal influence of each foundation pit feature; Represents the total causal effect matrix The Middle Line number Column elements; Based on preset screening criteria, a set of key foundation pit features is selected from all foundation pit features. : in, This is a preset threshold for causal influence.

6. The method for predicting the state of an operational tunnel near a foundation pit based on a spatiotemporal decoupled network according to claim 1, characterized in that, The graph network edge set is established based on the spatial relationship and force transmission relationship between each node of the foundation pit and the tunnel. With edge weight set W Specifically, it includes: For foundation pit nodes With each tunnel node Establish edges The edge weights are determined based on the spatial distance and force transmission relationship. The calculation formula is: in, For foundation pit nodes With tunnel nodes Edge weights between them This is the weight normalization coefficient, used to adjust the numerical scale of the overall weights; This is the distance attenuation coefficient, used to control the rate attenuation of the influence intensity due to spatial distance; For foundation pit nodes With tunnel nodes The straight-line distance between them; For the foundation pit node to the tunnel node The force transmission coefficient is used to characterize the intensity of the mechanical influence transmitted from the foundation pit disturbance to the tunnel node; For any two tunnel nodes If there is a spatial proximity or force transmission path between two nodes, then an edge is established. Its edge weight The calculation formula is: in, For tunnel nodes With tunnel nodes Edge weights between them; This refers to the weight normalization coefficient between tunnel nodes; This is the distance attenuation coefficient between tunnel nodes; For tunnel nodes and Spatial distance; For tunnel nodes The force transmission coefficient between tunnel structures is used to characterize the degree of force coupling between tunnel structures. All established edges and their corresponding edge weights Forming edge sets With edge weight set W ,in, If node If there is no connection between them, then let .

7. The method for predicting the state of an operational tunnel near a foundation pit based on a spatiotemporal decoupled network according to claim 1, characterized in that, The spatiotemporal decoupling network includes a graph convolutional network, a time encoder based on a self-attention mechanism, a feature decoupling module, and a prediction output module. The temporal encoder includes a multi-head attention layer and a feedforward mapping layer, which are used to model the temporal spatial features output by the graph convolutional network module in the temporal dimension. The feature decoupling module includes a trend extraction unit and a fluctuation extraction unit, which are used to decompose the comprehensive features output by the time encoder module: the trend extraction unit uses smoothing filtering and recursive structure to extract long-term change features, which are used to characterize the continuous deformation trend caused by the foundation pit construction process. The fluctuation extraction unit uses residual extraction and short-term convolution structure to extract high-frequency change features, which are used to characterize short-term fluctuations caused by environmental disturbances or transient loads. The prediction output module includes a feature fusion layer and an output layer, which are used to fuse trend features and fluctuation features, and output the predicted state of each tunnel node at several future times through a fully connected mapping.

8. The method for predicting the state of an operational tunnel near a foundation pit based on a spatiotemporal decoupled network according to claim 1, characterized in that, The process of training the spatiotemporal decoupling network based on a temporal graph network ensemble to obtain the trained spatiotemporal decoupling network specifically includes: According to the preset sliding window length and prediction time domain, the temporal graph network set is divided into several training samples by the sliding window. Each training sample is input by the node feature sequence of the graph network at several consecutive time steps and the tunnel node feature sequence at several subsequent time steps as the prediction target. The sliding window length and prediction time domain are adjustable hyperparameters. The node feature sequences of each training sample are uniformly processed, including time alignment, missing value imputation, numerical normalization, and dimensionality mapping. The key features of the foundation pit and the deformation / convergence / displacement features of each tunnel node are organized in node order to form the features at each time step. t The node feature matrix; For each time step in the input sequence t Edge sets in graph networks With edge weight set W A graph convolutional network from the spatiotemporal decoupling network is used for each time step. t The node feature matrix is ​​aggregated and transformed using neighborhood information to obtain the node feature matrix at each time step. t Spatial representation of each node; The spatial representations of each node at each time point are arranged along the time axis to form a time series representation. A time encoder based on a self-attention mechanism is used to extract time features from the time series representation, capture the dynamic dependencies between different time steps, and obtain a node-level spatiotemporal feature representation containing time series information. The feature decoupling module separates the trend component and fluctuation component of the node-level spatiotemporal feature representation, and independently models the long-term change trend and short-term disturbance respectively. The trend component is used to capture the overall deformation trend caused by the excavation of the foundation pit and the change of the support structure, while the fluctuation component is used to characterize the local time-varying response under the influence of load changes or environmental disturbances. The trend component and the fluctuation component are integrated in the feature fusion layer to form a complete node-level spatiotemporal feature representation, which is then input into the prediction output module. The prediction output module maps the node-level spatiotemporal features into the deformation, convergence and displacement prediction values ​​of each tunnel node at several future times through a fully connected mapping. A loss function is constructed based on the difference between the predicted and the true values. The parameters of the spatiotemporal decoupling network are iteratively optimized to obtain the trained spatiotemporal decoupling network.

9. A method for predicting the state of an operational tunnel near a foundation pit based on a spatiotemporal decoupled network according to claim 8, characterized in that, The loss function is defined as follows: in, For loss function, This is the set of training samples obtained by dividing the data using a sliding window. The total number of training samples, To predict the length of the time domain, Indicates the total number of preset tunnel nodes; The output dimension for each tunnel node includes deformation, convergence, and displacement; Represents a single training sample. Given the input node feature sequence, The corresponding tunnel node feature sequence; The model predicts the value, representing the first... At this moment i The tunnel node, the first u Prediction results for each output dimension; Indicates the first At this moment i The tunnel node, the first u The actual measurement value of each output dimension.

Images