A large foundation pit stress state intelligent calculation and grading early warning safety evaluation method

Abstract

This invention relates to the field of geotechnical engineering safety monitoring and intelligent early warning technology, specifically a method for intelligent calculation and hierarchical early warning safety assessment of stress state in large foundation pits. The method includes: acquiring initial environmental parameters and a three-dimensional structural model of the target large foundation pit; determining initial material properties and graph edge constitutive parameters; and converting these into a heterogeneous multiphysics tensor graph. It also involves acquiring multi-source physical monitoring data and updating the status of observation anchor nodes; evolving hidden computational nodes based on a physical information-constrained spatiotemporal tensor graph assimilation network to determine the full-map stress tensor field; extracting the Laplace matrix spectral features and the discrete residuals of the stress equilibrium partial differential equations; determining the algebraic connectivity drift rate; and generating safety assessment and early warning results and structural adjustment instructions. This achieves global stress state reconstruction, early identification of hidden risks, and automatic generation of support adjustments.

Description

Technical Field

[0001] This invention relates to the field of geotechnical engineering safety monitoring and intelligent early warning technology, specifically to a method for intelligent calculation and graded early warning safety assessment of stress state in large foundation pits. Background Technology

[0002] Existing safety assessment technologies for large foundation pit construction typically involve collecting monitoring data such as displacement around the foundation pit, axial force of supports, pore water pressure, and surface settlement, and combining this data with a three-dimensional structural model of the foundation pit, stratum parameters, or numerical analysis models to determine the stability during the construction period. In related technologies, in order to identify abnormal risks during the construction of large foundation pits, a monitoring and alarm method based on single measuring point thresholds can be used to determine whether displacement, stress or pore pressure data exceed the limits; or, finite element analysis, empirical models or conventional data interpolation methods can be used to calculate and predict the stress and deformation state of local areas of the foundation pit. However, given the diverse types of sensors, inconsistent sampling frequencies, dynamic changes in construction conditions, and complex spatial structures of the foundation pit, the former method struggles to reflect the overall stress transfer state in non-observed areas, while the latter method struggles to simultaneously address the fusion of multi-source monitoring data, global mechanical consistency, and early identification of hidden instability trends. Consequently, the safety assessment methods for large foundation pits in related technologies struggle to balance the ability to reconstruct the stress state across the entire map with the accuracy of graded early warning. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an intelligent calculation and graded early warning safety assessment method for the stress state of large foundation pits. Specifically, the technical solution of this invention includes: The initial environmental parameters and three-dimensional structural model of the target large foundation pit are obtained, and the initial material properties and constitutive parameters of the graph edge are determined accordingly. Based on the initial material properties and graph edge constitutive parameters, the initial environmental parameters and three-dimensional structural model are transformed to determine a heterogeneous multiphysics tensor graph containing observation anchor nodes, hidden computation nodes, and edges connecting the two. Acquire multi-source physical monitoring data of the target large foundation pit at the current moment, including stress and displacement monitoring data; The state of the observation anchor node is updated based on the multi-source physical monitoring data to determine the state of the target observation anchor node containing energy stress characteristics. Based on the target observation anchor node state and the preset physical information constrained spatiotemporal tensor graph assimilation network, the hidden computation nodes are evolved to determine the full graph stress tensor field; wherein, the network uses a preset physical conservation law as the loss function for weight constraints and includes a multiphysics coupled graph operator; Extract the features of the stress tensor field of the whole image, determine the Laplace matrix spectrum features and the discrete residuals of the stress equilibrium partial differential equations based on the physical conservation law; Based on the spectral characteristics of the Laplace matrix, determine the algebraic connectivity drift rate; Based on the algebraic connectivity drift rate and the discrete residual of the stress balance partial differential equation, the safety assessment and early warning results are determined. Based on the safety assessment and early warning results, a structural adjustment instruction for the target large foundation pit is determined.

[0004] Optionally, the step of transforming the initial environmental parameters and the three-dimensional structural model based on the initial material properties and graph edge constitutive parameters to determine a heterogeneous multiphysics tensor graph containing observation anchor nodes, hidden computation nodes, and edges connecting the two includes: Based on the initial environmental parameters, determine the spatial coordinates of the graph nodes; Based on the spatial coordinates of the graph nodes, the initial properties of the material, and the constitutive parameters of the graph edges, an initial tensor graph containing the nodes corresponding to the spatial coordinates of the graph nodes is constructed. Obtain preset physical sensor location information; Based on the physical sensor location information, the nodes of the initial tensor graph are classified to determine the observation anchor node and the hidden computation node; The heterogeneous multiphysics tensor graph is generated based on the observation anchor node, the hidden computation node, and the initial tensor graph.

[0005] Optionally, the process of evolving the hidden computation nodes to determine the full-graph stress tensor field based on the target observation anchor node state and a preset physical information-constrained spatiotemporal tensor graph assimilation network includes: Based on the multiphysics coupling graph operator, the energy stress characteristics in the target observation anchor node state are extracted; The energy stress features are propagated along the edge to the hidden computing node to determine the node feature update sequence; The node feature update sequence is subjected to local iterative update processing to determine the full-image stress tensor field.

[0006] Optionally, the step of extracting the features of the full-image stress tensor field, determining the Laplace matrix spectrum features and the discrete residuals of the stress equilibrium partial differential equations based on the physical conservation laws includes: The full-image stress tensor field is subjected to matrix decomposition to determine the target Laplace matrix; Spectral analysis is performed on the target Laplacian matrix to determine its spectral characteristics; Based on the stress equilibrium partial differential equation corresponding to the preset physical conservation law, residual calculation processing is performed on the stress tensor field of the whole figure to determine the discrete residual of the stress equilibrium partial differential equation.

[0007] Optionally, determining the algebraic connectivity drift rate based on the spectral characteristics of the Laplace matrix includes: The second smallest eigenvalue is extracted from the spectral features of the Laplacian matrix to determine the target algebraic connectivity. Obtain the historical time series corresponding to the target large foundation pit, and extract the historical algebraic connectivity corresponding to the historical time series; Based on the historical algebraic connectivity and the target algebraic connectivity, time derivative calculation is performed to determine the algebraic connectivity drift rate.

[0008] Optionally, determining the safety assessment and early warning result based on the algebraic connectivity drift rate and the discrete residual of the stress balance partial differential equation includes: Obtain the preset critical phase transition point threshold; Construct a multidimensional phase space with the algebraic connectivity drift rate and the discrete residuals of the stress balance partial differential equation as dimensions; When the algebraic connectivity drift rate is greater than the critical phase transition point threshold, the corresponding multidimensional phase space hierarchical early warning mechanism is triggered according to the preset boundary division in the multidimensional phase space to generate an early warning level, and the first safety assessment early warning result is determined by combining the discrete residual of the stress balance partial differential equation. If the algebraic connectivity drift rate is less than or equal to the critical phase transition point threshold, the current monitoring status is maintained, and the second safety assessment warning result is determined by combining the discrete residual of the stress balance partial differential equation. The first security assessment warning result or the second security assessment warning result shall be used as the security assessment warning result.

[0009] Optionally, the step of triggering a corresponding multi-dimensional phase space hierarchical early warning mechanism based on the preset boundary division within the multi-dimensional phase space to generate an early warning level, and determining the first safety assessment early warning result in conjunction with the discrete residual of the stress balance partial differential equation, includes: Based on the stress balance partial differential equation, the residuals are discretized to generate a residual heatmap; The residual heatmap is processed to locate abnormal regions, and regions with residual values ​​greater than a preset residual threshold are extracted. Combined with the equivalent stress and displacement gradient data extracted from the stress tensor field of the whole map and the multi-source physical monitoring data, stress concentration areas and displacement anomaly areas are determined respectively. The warning level, the residual heat map, the stress concentration area, and the displacement anomaly area are combined to determine the first safety assessment warning result.

[0010] Optionally, determining the structural adjustment instruction for the target large foundation pit based on the safety assessment and early warning results includes: Obtain a pre-defined structural support system adjustment strategy library; If the safety assessment warning result is the first safety assessment warning result, the first safety assessment warning result and the structural support system adjustment strategy library are matched to determine the target reinforcement command or the target decompression command. If the security assessment warning result is the second security assessment warning result, a status quo maintenance order is determined; The target reinforcement command, the target decompression command, or the status quo maintenance command are output as the structural adjustment command.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. This method constructs a heterogeneous multiphysics tensor graph containing observation anchor nodes and hidden computation nodes, and evolves the hidden computation nodes based on the state of the target observation anchor nodes and a pre-defined physical information-constrained spatiotemporal tensor graph assimilation network. This mechanism can uniformly map discrete, multi-source physical monitoring data with inconsistent sampling frequencies into a graph structure with mechanical constraints, effectively overcoming the blind spot problem that traditional single-point monitoring cannot cover the whole, and realizing real-time and accurate reconstruction of the stress tensor field of the non-observation area and the globally consistent graph. 2. This method extracts the stress tensor field characteristics of the whole image, determines the Laplace matrix spectrum characteristics and the discrete residuals of the stress equilibrium partial differential equation, and further calculates the algebraic connectivity drift rate by combining historical time series. This method can simultaneously evaluate the health status of the system from two dimensions: the dynamic decay trend of global topological connectivity and the degree of imbalance of local physical equilibrium. It effectively solves the problem that existing conventional data interpolation or finite element analysis methods are difficult to take into account global continuity and are prone to missing hidden risks, thus enabling early identification of asymptotic instability processes. 3. This method overcomes the limitations of traditional single-point fixed threshold alarms, which are prone to false alarms or missed alarms. It constructs a multi-dimensional phase space with algebraic connectivity drift rate and discrete residuals of stress equilibrium partial differential equations as dimensions. By combining preset critical phase transition point thresholds, it triggers a multi-dimensional phase space hierarchical early warning mechanism and generates early warning levels. It integrates the overall evolution trend of the system and local physical anomalies into a joint judgment framework, thereby providing more stable and hierarchical safety assessment and early warning results for the construction period of large foundation pits. 4. After obtaining the safety assessment and early warning results, this method further combines the preset structural support system adjustment strategy library for matching processing, and automatically determines the target reinforcement command, target pressure reduction command, or maintain the status quo command. This mechanism directly transforms the abstract calculation and early warning results into executable on-site commands with spatial positioning and disposal priority, shortening the response link that traditionally relies on manual review and judgment, and realizing the automatic closed-loop execution of foundation pit safety monitoring, intelligent calculation, risk early warning, and structural adjustment measures output. Attached Figure Description

[0012] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0014] like Figure 1 As shown, a method for intelligent calculation and graded early warning safety assessment of stress state in large foundation pits is proposed. The method includes: The initial environmental parameters and three-dimensional structural model of the target large foundation pit are obtained, and the initial material properties and constitutive parameters of the graph edge are determined accordingly. Based on the initial material properties and graph edge constitutive parameters, the initial environmental parameters and three-dimensional structural model are transformed to determine the heterogeneous multiphysics tensor graph containing observation anchor nodes, hidden computation nodes, and the edges connecting the two. Acquire multi-source physical monitoring data of the target large foundation pit at the current moment, including stress and displacement monitoring data; Update the state of the observation anchor node based on multi-source physical monitoring data, and determine the state of the target observation anchor node containing energy stress characteristics. Based on the target observation anchor node state and the pre-defined physical information constraint spatiotemporal tensor graph assimilation network, the hidden computation nodes are evolved to determine the full graph stress tensor field; the network uses a pre-defined physical conservation law as the loss function for weight constraint and includes a multi-physics coupled graph operator. Extract the features of the stress tensor field of the whole image, determine the Laplace matrix spectrum features and the discrete residuals of the stress equilibrium partial differential equations based on the physical conservation laws; Determine the algebraic connectivity drift rate based on the spectral characteristics of the Laplace matrix; Based on the algebraic connectivity drift rate and the discrete residuals of the stress equilibrium partial differential equation, the safety assessment and early warning results are determined. Based on the safety assessment and early warning results, a structural adjustment instruction for the target large foundation pit was determined.

[0015] This embodiment provides a mechanism for intelligent calculation and graded early warning safety assessment of stress state in large foundation pits. Specifically, it uses the safety monitoring of a deep foundation pit during the construction period of an urban rail transit transfer station as the main scenario. The foundation pit has an excavation depth of 28 meters and a planar area of ​​approximately 120 meters × 65 meters. There are existing subway sections, storm and sewage pipelines, and adjacent commercial buildings in the surrounding area. The support system includes underground continuous walls, steel supports, local ecological composite soil nails, and a dewatering well system. Axial force gauges, inclinometers, surface settlement points, pore water pressure gauges, and displacement sensors are deployed on site. An edge graphics processor server is set up in the monitoring room at the construction site to receive monitoring data and perform real-time calculations. The process is described below: Initial environmental parameters and a 3D structural model are obtained. Initial environmental parameters may include stratum thickness, natural soil unit weight, cohesion, internal friction angle, elastic modulus, permeability coefficient, groundwater level, rainfall intensity, excavation stage number, and surrounding load information. The 3D structural model can be formed by combining a building information model, survey profiles, and support construction drawings. Taking a simplified computational space as an example, if the local area of ​​the foundation pit is discretized into 6 spatial points N1 to N6, where N1 and N2 are located at the top of the retaining wall, N3 and N4 are located in the middle soil layer, and N5 and N6 are located in the area adjacent to the bottom of the pit, then each point has spatial coordinates and material labels. For example, N1 corresponds to coordinates 0,0,0, and N4 corresponds to coordinates 6,0,-12. Soil layer labels can be distinguished as silty clay layer, sand layer, fill layer, etc. Initial material properties are determined based on initial environmental parameters, and constitutive parameters at the graph edges are determined based on a three-dimensional structural model. These initial material properties are not single strength indices, but rather composite eigenvectors used in subsequent tensor graph calculations. For example, the soil layer at a given node can be characterized by an elastic modulus of 30 MPa, a Poisson's ratio of 0.33, and a permeability coefficient of... An initial combination of m / s and a natural water content of 24% is used; edge constitutive parameters are used to describe the strength of the interaction between two nodes, and can be composed of equivalent stiffness, seepage connectivity coefficient, and contact influence coefficient; taking 6 nodes as an example, if N3 and N4 are located in the same continuous soil layer, their edge weights can be set as [stiffness 0.8, seepage 0.6, contact 0.9]; if there is a supporting member between N2 and N5, their edge weights can be set as [stiffness 0.7, seepage 0.2, contact 0.5]; in this way, the physical relationships in continuous space are transformed into computable graph edge relationships; Furthermore, based on the initial material properties and graph edge constitutive parameters, the initial environmental parameters and the three-dimensional structural model are transformed to form a heterogeneous multiphysics tensor graph. Here, heterogeneity is manifested in different node types, different physical meanings of edges, and different dimensions of state quantities. For example, observation anchor nodes can directly receive sensor data, while hidden computation nodes are updated through graph network deduction. Edges can be categorized into soil transfer edges, support force transfer edges, and seepage coupling edges. The node states in the tensor graph can at least include stress components, displacement components, pore pressure components, and energy stress characteristics. As an example, a single node state can be set as a vector of length 5, such as […]. , , , , ], representing horizontal stress, vertical stress, displacement, pore pressure, and energy per unit volume, respectively; After acquiring multi-source physical monitoring data, the status of the observed anchor nodes is updated. Different sensors in the field typically have different sampling frequencies; for example, displacement data is collected every 5 minutes, axial force data every 1 minute, and pore pressure data every 10 minutes. Therefore, time alignment, unit unification, and outlier removal are performed first. Taking a certain time t as an example, if the observed axial force equivalent stress of anchor node A1 is 120 kPa, the horizontal displacement is 3.2 mm, and the pore pressure is 18 kPa, then the original node status can be updated from […]. Updated to [] ]; The energy stress characteristics can be generated by normalizing and weighting stress, displacement and pore pressure. Specifically, the normalized and weighted processing method is as follows: the stress, displacement and pore pressure components are respectively normalized to their maximum reference value to eliminate the difference in dimensions, and then multiplied by the preset physical quantity weighting coefficient and summed. For example, after normalization and weighting, a value of 0.53 can be obtained. If multiple sensors are mapped to the same node at the same time, a weighted average or confidence-first mechanism can be used for updating. For example, if the confidence of the stress sensor is 0.9 and the confidence of the displacement sensor is 0.8, the fusion result will be more biased towards the high-confidence data. After obtaining the state of the target observation anchor node, the hidden computation node is evolved by using the pre-defined physical information constrained spatiotemporal tensor graph assimilation network to obtain the full graph stress tensor field; the main basic architecture of the spatiotemporal tensor graph assimilation network includes at least one graph convolutional network layer or graph attention network layer for extracting spatial topological node features, and at least one long short-term memory network layer for extracting temporal evolution features. This network does not simply fit sensor values. Instead, during training or online fine-tuning, it incorporates stress balance, mass conservation, and boundary constraints as loss terms. The specific constraint logic is as follows: part of the loss constraint requires the predicted value to be close to the sensor value, while another part of the loss constraint requires that the mechanical transmission between adjacent nodes does not violate the conservation law. For example, in a local 3-node chain A1-H1-H2, if the energy stress characteristics of A1 increase, H1 and H2 will not experience a sudden increase in values ​​without physical constraints. Instead, they will gradually propagate according to the edge weights and conservation constraints, eventually forming a continuous, smooth, and support force transmission path full-map stress tensor field. Feature extraction is performed on the stress tensor field of the entire graph to obtain the spectral features of the Laplace matrix and the discrete residuals of the stress equilibrium partial differential equations. Here, the entire graph can be regarded as a physical network with edge weights. Taking a 4-node subgraph as an example, if its adjacency relationship is A1 connecting H1, H1 connecting H2, and H2 connecting A2, then the corresponding graph Laplace matrix can be generated. After spectral analysis of this matrix, the second smallest eigenvalue is extracted as the basic quantity of algebraic connectivity. At the same time, the stress and displacement of the nodes of the entire graph are substituted into the discretized equilibrium relationship to calculate the residual value of each node or element. If the residual at H2 is greater than a preset multiple of the mean residual of the surrounding nodes, for example, the residuals of the surrounding nodes are all between 0.03 and 0.07, while H2 reaches 0.21, it indicates that there may be a stress mutation or material state anomaly that is not explained by the model at this point. The algebraic connectivity drift rate is determined based on the spectral characteristics of the Laplace matrix. The second smallest eigenvalue at the current time can be compared with historical values; for example, if it is 1.20 at time t1, 1.05 at time t2, and 0.78 at time t3, it indicates that the system connectivity is gradually decreasing. Let the algebraic connectivity drift rate be... Its general calculation formula is: ,in, and These represent the second smallest eigenvalues ​​of the Laplace matrix at the current time and the previous historical time, respectively; If the drift rate is calculated per unit time, the drift rate within the time window from t2 to t3 can be expressed as: ,in, The time step difference corresponding to the aforementioned time window corresponds to a negative trend. When the rate of change continues to increase or the absolute value exceeds the preset safety threshold range, it indicates that the overall force transmission topology of the foundation pit is approaching the instability boundary, rather than the accidental measurement noise of a single sensor. Based on the algebraic connectivity drift rate and the discrete residuals of the partial differential equation of stress balance, the safety assessment early warning results are determined, and structural adjustment instructions are further determined. For example, if the drift rate is in the warning zone and the residual heat map shows that the corner of the pit bottom and the junction of the third support are abnormal at the same time, a comprehensive instruction of secondary early warning + local grouting reinforcement + reduction of excavation step distance is output; if the drift rate is stable and the residual does not exceed the limit, an instruction to maintain the status quo and continue monitoring is output. In special or abnormal situations, such as missing data, sensor drift, and limited computing resources, the following processing logic can be adopted: If an observation anchor node has no data for two consecutive sampling periods, the most recent reliable state is retained and the weight of the node is reduced, rather than directly deleting the node; If a sensor of a certain type has a numerical difference between adjacent sampling periods that is greater than the set jump threshold, such as the displacement jumping from 3mm to 35mm within 5 minutes, but there is no coordinated change in adjacent nodes, it is first marked as a suspected anomaly and a review is initiated, without directly driving the entire map into a high-level warning; If the numerical difference is less than or equal to the set jump threshold, the status of the observation anchor node is updated to the tensor map according to the normal process; If the graph size is too large and causes the edge server to time out in a single calculation, the graph is divided into blocks to calculate the local subgraph first, and then consistency correction is performed at the block interface; When the excavation of the fourth layer of earthwork in the foundation pit of the rail transit transfer station was carried out, the two inclinometers in the southwest corner only showed a displacement increment less than the preset displacement alarm threshold, which did not reach the traditional threshold alarm condition. However, after the system synchronized this slight increment to the tensor map, it found that multiple hidden calculation nodes connected to it formed a continuously rising discrete residual band in the local area. At the same time, the second smallest eigenvalue dropped from 0.96 to 0.71 in three consecutive time windows. At this time, the system determined that there was a deep soil softening trend in the area, issued a graded warning in advance, and generated structural adjustment instructions for grouting the passive zone at the bottom of the pit in the southwest corner, temporarily reducing the dewatering rate, and suspending the continued excavation of adjacent blocks. The purpose of this step is to map discrete, multi-source, and different frequency field monitoring data into a spatiotemporal diagram structure with mechanical constraints, thereby enabling real-time reconstruction of the global stress state, early identification of hidden risks, and automatic generation of support adjustment measures. In this embodiment, based on the initial material properties and graph edge constitutive parameters, the initial environmental parameters and the three-dimensional structural model are transformed to determine a heterogeneous multiphysics tensor graph containing observation anchor nodes, hidden computation nodes, and edges connecting the two, including: Determine the spatial coordinates of the graph nodes based on the initial environmental parameters; Based on the spatial coordinates of the graph nodes, the initial material properties, and the constitutive parameters of the graph edges, an initial tensor graph containing the nodes corresponding to the spatial coordinates of the graph nodes is constructed. Obtain preset physical sensor location information; Based on the location information of the physical sensors, the nodes of the initial tensor graph are classified to determine the observation anchor nodes and hidden computation nodes; A heterogeneous multiphysics tensor graph is generated based on the observation anchor node, hidden computation node, and initial tensor graph.

[0016] This embodiment provides a mechanism for constructing a heterogeneous multiphysics tensor graph. Specifically, in the aforementioned deep foundation pit scenario of a transfer station, if modeling is directly based solely on existing sensor locations, the spatial coverage density variance will exceed a set distribution threshold: there are many sensors near the retaining wall, but few sensors at the bottom of the pit and in the intersection area of ​​the supports, which will lead to blind spots during subsequent full-map stress reconstruction. Therefore, this embodiment further projects the initial environmental parameters, the three-dimensional structural model, and the sensor location information into the graph space to form a heterogeneous tensor graph that simultaneously contains real measurement points and virtual calculation points. The following is a detailed description: The spatial coordinates of the graph nodes are determined based on the initial environmental parameters. The node coordinates can be directly derived from the grid nodes of the Building Information Model (BIM), or they can be generated jointly by the stratum layer lines, key points of the support components, and the locations of the dewatering wells. To avoid the graph being too dense, a discrete spacing can be set according to the engineering dimensions, for example, one point every 3 meters horizontally and one point every 2 meters vertically. For example, in a simplified profile that is 12 meters long and 6 meters deep, 9 nodes can be discretized: 3 for the top layer, 3 for the middle layer, and 3 for the bottom layer. If the top layer nodes are located at 0,0,0, 6,0,0, and 12,0,0 respectively, the middle layer nodes are located at 0,0,-3, 6,0,-3, and 12,0,-3, and the bottom layer nodes are located at 0,0,-6, 6,0,-6, and 12,0,-6, then these coordinates become the spatial basis of the graph nodes. An initial tensor graph is constructed based on the spatial coordinates of the graph nodes, initial material properties, and constitutive parameters of the graph edges. One feasible implementation is as follows: if adjacent nodes satisfy geometric proximity and are located in the same continuous soil layer or can transmit force through support components, then an edge is established; the properties of the edge are jointly determined by material properties and structural relationships; for example, the equivalent stiffness between two adjacent nodes located in a silty clay layer can be taken as 0.75; if a steel support passes through the middle, the stiffness increases to 0.92; if it crosses a cutoff wall, the seepage coefficient decreases; the initial tensor graph formed in this way preserves the spatial location and also encodes the engineering physical relationships. Obtain the preset physical sensor location information; field sensors usually have an installation log, including number, coordinates, burial depth, monitoring type, and installation date; when mapping this information to graph nodes, nearest neighbor matching or radius matching can be used; for example, if sensor P1 has coordinates of 6.2, 0, -3.1, then it is closest to node 6, 0, -3, so it can be mapped to that node first; if the difference between the distances of a sensor and multiple nodes is less than the set distance tolerance, it can be further determined based on its type, such as inclinometers being mapped to nodes adjacent to the retaining wall first, and pore pressure gauges being mapped to soil layer nodes first; After completing the point mapping, the nodes of the initial tensor graph are classified to determine observation anchor nodes and hidden computation nodes. The classification criteria here are not based on the importance of the nodes, but on whether the field observation value can be directly obtained at the current moment. If a node is mapped by at least one sensor and can provide valid data, it is marked as an observation anchor node. If the node does not have a direct observation input, but is located within the physical network and needs to participate in the solution, it is marked as a hidden computation node. Taking a local model consisting of 9 nodes as an example, if the top layer midpoint, left middle layer, and bottom layer midpoint are mapped to actual sensors, then these 3 nodes belong to observation anchor nodes, and the remaining 6 nodes belong to hidden computation nodes. A heterogeneous multiphysics tensor graph is generated based on two types of nodes and the initial tensor graph. Specifically, node type identifiers can be added to the node attributes, such as marking observation anchor nodes as 1 and hidden computation nodes as 0. Simultaneously, the same data dimension space is reserved in the node state vector to ensure compatibility during subsequent network propagation. If the observation anchor node contains measured values, the corresponding dimension is filled with valid values. If the hidden computation node has not yet been observed, an initial value or prior estimate can be filled in. For example, the state of the observation anchor node can be set to […]. The hidden compute node state can be set to []. ]; In special or abnormal situations, if a sensor is located at an equidistant position between two nodes and its physical influence range covers both, a multi-node weighted mapping method can be used to distribute the measured value to the two nodes according to the distance weight. If a node is mapped by two similar sensors at the same time and the results are contradictory, the data of the device with the more recent calibration time and smaller drift history will be retained first. If a certain area is not covered by sensors and the resolution of the geological data is lower than the preset accuracy standard, the nodes in that area can still be retained, but their initial uncertainty coefficient will be increased, so that the subsequent network will reduce the confidence dependence of the results in that area during propagation. On the west side of the transfer station, near the existing subway section, multiple inclinometers are installed along the retaining wall, while only two pore pressure gauges are installed in the passive zone at the bottom of the pit. If the former is used as the main input, the model will be biased towards the wall response and will be unable to reflect the deep softening at the bottom of the pit. With this embodiment, hidden calculation nodes can be added to the passive zone at the bottom of the pit according to the stratigraphic grid, and seepage coupling edges can be established through anchor nodes adjacent to the pore pressure gauges, thereby forming a complete graph structure with a small number of measured points and a large number of virtual calculation points, which provides a basis for subsequent full-map stress inversion. The purpose of this step is to encode the engineering space, material properties, structural force transmission relationship and sensor deployment status into the same graphical model, so as to realize the collaborative modeling of the observation area and the non-observation area, and provide a stable data foundation for the subsequent global state evolution. In this embodiment, based on the target observation anchor node state and a preset physical information-constrained spatiotemporal tensor graph assimilation network, the hidden computation nodes are evolved to determine the full graph stress tensor field, including: Based on the multiphysics coupling graph operator, the energy stress characteristics in the target observation anchor node state are extracted; The driving energy stress features are propagated along the edge to the hidden computation nodes to determine the node feature update sequence; The node feature update sequence is locally iteratively updated to determine the stress tensor field of the whole graph.

[0017] This embodiment provides an evolution mechanism for a spatiotemporal tensor graph assimilation network based on physical information constraints. Specifically, after the graph is constructed, if the measured values ​​of the observed anchor nodes are directly interpolated to the surrounding hidden nodes, although the computation time is lower than the first preset time requirement, simple interpolation will mask the real mechanical abrupt changes when there are support transformations, sudden precipitation changes, or the presence of local weak interlayers. Therefore, this embodiment introduces a multiphysics coupled graph operator, so that the energy stress characteristics consider the coupling relationship between stress, displacement, and seepage during the propagation process. The details are as follows: Energy stress features in the state of the target observation anchor node are extracted using the multiphysics coupling graph operator; this extraction can be understood as re-encoding the original state vector; assuming the state of anchor node A is [ New coupling features can be obtained through a set of preset linear transformations and nonlinear activations. The specific operations of the preset linear transformations and nonlinear activations are as follows: the original node state vector is dimensionless to eliminate the difference in dimensions, then multiplied by the preset multiphysics feature mapping weight matrix, and after adding the bias vector, it is input into the linear rectifier function or the S-shaped nonlinear activation function for calculation. like The first dimension primarily characterizes the structural stress response, the second dimension primarily characterizes the seepage effect, and the third dimension primarily characterizes the displacement energy. For another anchor node B, if its pore pressure is high but its displacement is low, then it is possible to obtain... After this processing, different physical quantities are compressed into a unified feature space, which facilitates propagation along the graph edges. The driving energy stress features are propagated along the edges to the hidden computing nodes to obtain the node feature update sequence. This process can be performed in a layer-by-layer propagation manner. Taking a 4-node chain A-H1-H2-B as an example, in the first round of propagation, H1 receives information from A and H2, and H2 receives information from H1 and B. If the edge stiffness from A to H1 is greater than the first stiffness threshold and the seepage coefficient is in the preset intermediate value range, while the edge seepage coefficient from H2 to B is greater than the first seepage threshold, then H1 integrates the stress-dominant feature of A with a ratio greater than the second preset weight, and H2 integrates the pore pressure-dominant feature of B with a ratio greater than the second preset weight. The specific deduction process is as follows: the updated value of H1 is... The old value is composed of H2, while the updated value of H2 is composed of H2. The old values ​​are formed; after the first round, H1 and H2 form new features respectively; the second and third rounds continue to iterate until the update amount is small enough; it should be noted that the weight coefficients in the above propagation process, such as 0.6 and 0.4, are not fixed empirical constants, but are dynamically calculated by the multiphysics coupled graph operator based on the constitutive parameters of the graph edge; specifically, the multiphysics coupled graph operator reads the equivalent stiffness between nodes, the seepage connectivity coefficient, and the displacement gradient of the current state, and assigns higher transmission weights to edges with high stiffness and unobstructed seepage channels, so as to ensure that the propagation trajectory of energy stress features strictly conforms to the physical conservation and transmission law of engineering practice; Furthermore, the node feature update sequence is locally iteratively updated to determine the stress tensor field of the entire image. The significance of local iteration is that it avoids the need to globally recalculate the complete tensor map containing all nodes in a single iteration. Instead, it prioritizes updating only the local sub-maps whose weights affected by new observations exceed a set threshold. For example, when a new set of pore pressure anomalies is added in the southwest corner, the 2-hop or 3-hop neighborhood of that region can be locked for rapid evolution before being fused with the entire image. This ensures both real-time performance and spatial continuity. The final output stress tensor field of the entire image can be the multi-component stress result at each node, such as each node containing σx and σz components representing horizontal and vertical stresses, or their corresponding principal stress indices. To clarify the calculation process, a simplified example of local quantization derivation can be provided; assuming that the initial coupling characteristics of H1 and H2 are respectively [ ]and[ The characteristics of A and B are respectively [] ]and[ After the first update, H1 becomes [ H2 becomes [ After the second update, H1 became [ H2 becomes [ When the variation between the two rounds is less than 0.03, it can be determined that the local convergence has been basically achieved; then the convergence feature is mapped back to the stress space, for example, H1 corresponds to The pore pressure coupling stress index corresponding to H2 increased from 0.43 to 0.58 when the pressure was updated from 108 kPa to 126 kPa. In special or abnormal situations, if fluctuations exceeding the preset oscillation threshold occur during the propagation of a local area, such as the difference between the values ​​of H1 between two iterations being greater than the set jump threshold, it indicates that the edge weights in that area may be too strong or there may be conflicts in the observation data. In this case, the propagation step size in that area can be automatically reduced, or an edge weight smoothing strategy can be enabled. If a hidden node relies solely on connections with a path length greater than the set number of jumps and edge weights lower than the set lower limit for a long period of time without near-field anchor support, it can be marked as a low-confidence node, and the low reliability of the result can be separately indicated in the warning report. If the on-site working conditions suddenly change, such as the addition of a support or local backfilling, the constitutive parameters of the relevant edges should be updated first, and then incremental propagation should be started, rather than forcibly evolving using the old edge weights. After the tensioning of the third steel support in the pit of the transfer station was completed, the displacement increment at the top of the eastern retaining wall did not exceed the preset displacement observation threshold, but the pore pressure of the sand layer in the middle continued to rise within two sampling windows. The system identified the composite characteristics of stable displacement and rising pore pressure through the coupled graph operator, and propagated along the support-soil-pit bottom path to the hidden nodes, eventually forming a continuous high-energy stress zone at the bottom of the pit in the southeast corner. This result indicates that the risk is not manifested as immediate wall inclination, but more likely as seepage uplift at the bottom of the pit and local unloading imbalance. The purpose of this step is to rapidly diffuse the composite mechanical information carried by discrete measuring points into the entire graph structure through local propagation and iterative updates, thereby achieving continuous reconstruction of the stress state in the hidden area while taking into account both computational efficiency and physical rationality. In this embodiment, the features of the stress tensor field of the entire image are extracted, and the Laplace matrix spectrum features and the discrete residuals of the stress equilibrium partial differential equations based on the physical conservation laws are determined, including: The stress tensor field of the whole image is subjected to matrix decomposition to determine the target Laplace matrix; Spectral analysis is performed on the target Laplacian matrix to determine its spectral characteristics. Based on the stress equilibrium partial differential equation corresponding to the preset physical conservation law, the residual calculation of the stress tensor field of the whole figure is performed to determine the discrete residual of the stress equilibrium partial differential equation.

[0018] This embodiment provides a mechanism for spectral feature extraction and discrete residual calculation of the stress tensor field of the whole graph. Specifically, after obtaining the stress tensor field of the whole graph, if only the absolute value of the stress at a single node is observed, it is easily affected by local noise or transient construction disturbances. For example, at the moment of tensioning of a certain support, the stress value may increase for a short time, but this does not mean that the whole system is unstable. Therefore, this embodiment analyzes the system state from the perspectives of graph topological spectral features and physical equilibrium residuals to avoid distortion by a single index. The process is described below: Matrix decomposition is performed on the stress tensor field of the entire graph to determine the target Laplace matrix. Specifically, the stress tensor field containing the characteristics of each node can be reorganized into a global state matrix, and singular value decomposition or principal component decomposition is performed on this global state matrix to extract the correlation coefficient matrix reflecting the dominant stress modes of the entire graph, which serves as the basis for dynamically adjusting edge weights. Generally, a weighted adjacency matrix W can be constructed first in conjunction with the basic topology, where the matrix element Wij represents the comprehensive connectivity strength between node i and node j. Here, i and j represent the index numbers of any two different nodes in the graph. This connectivity strength can be jointly formed by the initial equivalent stiffness, seepage coupling strength, and the current stress compatibility extracted through the above matrix decomposition, i.e., the degree of stress co-evolution between nodes. Then construct the Laplace matrix based on the degree matrix D and the adjacency matrix W. ; where the degree matrix D is a diagonal matrix, and the value of each element on its diagonal is equal to the sum of all elements in the corresponding row or column of the weighted adjacency matrix W; For ease of understanding, assume there are 3 nodes C1, C2, and C3, where C1 is connected to C2 with a weight of 0.8; C2 is connected to C3 with a weight of 0.6; and C1 and C3 are not directly connected. Then W can be understood as a matrix with 0 on the diagonal and the off-diagonal positions recording the connection strength, and D records the total connection strength of each node; further subtraction yields L. Spectral analysis is performed on the target Laplacian matrix to obtain its spectral characteristics. These characteristics can include at least several eigenvalues ​​and their corresponding eigenvectors, where the second smallest eigenvalue typically reflects the algebraic connectivity of the graph. For example, if the sequence of eigenvalues ​​at a certain moment is […] If 0.92 is the key indicator to watch later, then the corresponding feature vector can also be used to locate which region contributes the most to the decline in connectivity. For example, if the vector corresponding to the second smallest feature value shows a larger value at the node in the southwest corner region, it indicates that the region plays a key role in the overall instability trend. Furthermore, based on the stress equilibrium partial differential equation corresponding to the physical conservation law, residual calculation is performed on the stress tensor field of the entire graph to obtain discrete residuals. Here, it is not necessary to elaborate on the continuous formula derivation, but to verify whether the internal forces, external loads and boundary constraints are in equilibrium at the discrete node or element level. This can be expressed by the node equilibrium difference method: for a certain node, the resultant force from adjacent nodes and external loads is reduced and compared with the predicted stress state of the node, and the difference between the two is the residual. The specific structured calculation process is as follows: extract the stress tensor transmitted from adjacent nodes to the target node according to the graph topology, and decompose the internal forces along the normal and tangential directions of the graph edges. By superimposing the surrounding soil and water pressure and external loads defined by the initial environmental parameters, the resultant force feedback vector at the node is obtained. Based on the central difference scheme of the partial differential equation of stress equilibrium, the residual distribution sequence between the resultant force feedback vector and the predicted stress state of the node itself is calculated, thereby accurately characterizing the degree of mechanical imbalance. Taking the simplified 3-node chain as an example, if the equivalent force transmitted from the left and right sides of C2 should theoretically be close to equilibrium, but the actual state difference obtained by tensor inversion is 0.18, while the adjacent nodes C1 and C3 are only 0.04 and 0.06 respectively, then C2 can be regarded as a local anomaly. After combining multiple anomalies, a residual heat map can be further formed. It should be noted that spectral features reflect the overall structural trend of the network, while residuals reflect the degree of local physical mismatch. The two are complementary. The former is suitable for identifying whether the system is heading towards instability, while the latter is suitable for locating where the anomalies mainly occur. Therefore, when the absolute stress value of a node is not the highest, but the region where it is located causes a decrease in spectral connectivity and the residuals are consistently higher than the preset residual threshold, the actual risk is often higher than that of a region with high stress but normal equilibrium. If a local subgraph is found to be approximately isolated during matrix decomposition, resulting in an additional zero eigenvalue in the Laplacian matrix, it indicates that there may be broken edges or disconnected monitoring areas during the graph construction stage. In this case, it is advisable to check whether the corresponding edge weights and node classifications have been omitted. If boundary conditions are missing during residual calculation, such as the latest excavation load not yet being recorded for a newly excavated area, the residual results for that area are only for reference and should not be directly used in high-level early warning determination. If the spectral analysis and residual conclusions are inconsistent, such as overall connectivity being stable but local residuals being suddenly high, it is advisable to enter observation mode first and increase the sampling frequency for that area. After localized dewatering adjustments were made at the northwest corner of the transfer station's foundation pit, the new full-map stress tensor field generated by the system showed that the overall stress peak was still located near the second steel support. However, Laplace spectral analysis revealed that the second smallest eigenvalue decreased from 1.01 to 0.83. At the same time, the residual heat map showed a continuous high-value area at the northwest corner of the pit bottom. Combining these two results, it can be determined that the risk does not originate from the steel support itself, but is more likely from the disruption of the balance between the soil at the pit bottom and the support system after the dewatering was changed. The purpose of this step is to evaluate the system state simultaneously from both global topology and local physical paths, thereby achieving the synchronous identification of the overall instability trend and the location of local anomalies. In this embodiment, the algebraic connectivity drift rate is determined based on the spectral characteristics of the Laplace matrix, including: Extract the second smallest eigenvalue from the spectral features of the Laplacian matrix to determine the target algebraic connectivity; Obtain the historical time series corresponding to the target large foundation pit, and extract the historical algebraic connectivity of the historical time series; Based on the historical algebraic connectivity and the target algebraic connectivity, the time derivative is calculated to determine the algebraic connectivity drift rate.

[0019] This embodiment provides a mechanism for determining the drift rate of algebraic connectivity. Specifically, comparing only the second smallest eigenvalue at a certain moment can reflect the connectivity level of the current graph structure, but it is difficult to distinguish between two completely different engineering states: stable low-level operation and rapid deterioration in a short period of time. Therefore, this embodiment further introduces historical time series and uses the rate of change of algebraic connectivity to identify the instability trend. The details are as follows: Extract the second smallest eigenvalue from the spectral features to determine the target algebraic connectivity; if the eigenvalue sequence corresponding to the current analysis time is […]. If the target algebraic connectivity is 0.74, then the historical time series is obtained and the historical algebraic connectivity is extracted. The historical time series can be constructed according to a fixed time window, such as the most recent 6 sampling windows, denoted as T1 to T6, with corresponding algebraic connectivity of 1.12, 1.06, 0.98, 0.91, 0.83, and 0.74. The algebraic connectivity drift rate is obtained by calculating the time derivative based on the historical and target algebraic connectivity. In engineering practice, complex high-order derivatives are unnecessary; a sliding window difference method can be used instead. For example, the average rate of change can be calculated using the three most recent windows. If T4 to T6 correspond to 0.91, 0.83, and 0.74 respectively, then the average drift rate can be understood as... At each time step, a negative downward trend is obtained. To avoid being misled by fluctuations at a single sampling point, the sequence can be smoothed first. If the sequence continues to decline after smoothing, it indicates that the graph structure is continuously losing its cohesion. Furthermore, the drift rate can be compared with the historical baseline. Assuming that the foundation pit was in the normal construction phase in the first two weeks, the absolute value of the drift rate usually does not exceed 0.03. However, the current window reaches 0.085, which can be regarded as an abnormal acceleration phase. Its beneficial effect is that the absolute algebraic connectivity of different foundation pits and different phases may not be directly comparable, but it is more meaningful for early warning relative to the rate of change of its own historical state. A simplified extrapolation example can be given: if the normal stage sequence is 1.20→1.18→1.17→1.16, the drift rate is close to a slow decrease; if the abnormal stage sequence becomes 1.10→0.98→0.84→0.70, the drift rate is significantly amplified; the former may just be a gradual redistribution of the structure caused by conventional excavation, while the latter may correspond to the precursors of local softening, abnormal seepage, or support failure. If the historical time series is insufficient, for example, if the system has just been deployed and has not yet accumulated enough windows, the simulation sequence in the design phase or the standard reference curve under similar working conditions can be used as a temporary baseline; if the sampling time interval is uneven, the actual time difference should be used instead of the default equal interval when calculating the drift rate; if the second smallest feature value shows a short-term rebound, for example, if 0.82 rises to 0.87 for a short time and then continues to decline, it should be determined first whether it is caused by human intervention such as local support tensioning, so as to avoid misjudging the short-term repair after effective reinforcement as abnormal noise; After continuous rainfall during the rainy season at the transfer station foundation pit, although the single displacement increment in the southwest corner area was still small, the algebraic connectivity decreased from 1.03 to 0.76 in five consecutive windows, and the drift rate increased from -0.02 to -0.09. Based on this, the system judged that the current situation was not a normal construction disturbance, but that the overall force transmission network was degrading at an accelerated rate, and entered a higher priority risk verification process. The purpose of this step is to characterize the health of the system by changing trends rather than static values, thereby enabling earlier identification of gradual instability processes. In this embodiment, the safety assessment and early warning results are determined based on the algebraic connectivity drift rate and the discrete residuals of the stress balance partial differential equation, including: Obtain the preset critical phase transition point threshold; Construct a multidimensional phase space with algebraic connectivity drift rate and discrete residuals of stress equilibrium partial differential equations as dimensions; When the algebraic connectivity drift rate is greater than the critical phase transition point threshold, the corresponding multidimensional phase space hierarchical early warning mechanism is triggered according to the preset boundary division in the multidimensional phase space to generate an early warning level, and the first safety assessment early warning result is determined by combining the discrete residual of the stress balance partial differential equation. If the algebraic connectivity drift rate is less than or equal to the critical phase transition point threshold, the current monitoring status is maintained, and the second safety assessment warning result is determined by combining the discrete residuals of the stress balance partial differential equation. The first or second security assessment warning result shall be used as the security assessment warning result.

[0020] This embodiment provides a hierarchical early warning mechanism based on multidimensional phase space. Specifically, the drift rate and discrete residuals can be obtained in the aforementioned steps. If only one of these indicators is used to set a fixed threshold, two types of problems may easily occur: one is that the overall connectivity decreases by more than the preset decrease threshold, but the local residuals have not yet been amplified, resulting in missed alarms; the other is that the local residuals temporarily increase due to construction transient interference, but the overall topology remains stable, resulting in false alarms. Therefore, this embodiment maps both to a multidimensional phase space for hierarchical judgment. The steps are as follows: First, a preset critical phase transition point threshold is obtained. This threshold is used to identify whether the system has entered a critical stage of transition from stability to instability. The threshold can be determined jointly by historical engineering data, numerical simulation, and expert verification. For example, in a certain type of stratum and support system, the absolute value of the algebraic connectivity drift rate of 0.06 is set as the critical phase transition point. Then, a multidimensional phase space is constructed, where one dimension represents the drift rate and the other dimension represents the discrete residual. Pore pressure anomaly index or displacement gradient index can also be added as additional dimensions in extended scenarios. For example, taking a two-dimensional phase space: the horizontal axis represents the algebraic connectivity drift rate, and the vertical axis represents the discrete residuals of the stress equilibrium partial differential equation; preset boundaries can divide the plane into a safe zone, a zone of interest, a warning zone, and a danger zone; if the drift rate is 0.03 and the residual is 0.05 at a certain moment, it falls into the safe zone; if the drift rate is 0.07 and the residual is 0.09, it may enter the warning zone; if the drift rate is 0.10 and the residual is 0.18, it falls into the danger zone; in this way, different indicators are no longer judged in isolation, but jointly represent the system state within a unified coordinate system; When the drift rate exceeds the critical phase transition point threshold, it indicates that the system has experienced significant topological degradation. At this point, a graded early warning mechanism is triggered based on the phase space boundary to generate an early warning level, and the first safety assessment early warning result is determined in conjunction with the discrete residuals. In addition to assessing the magnitude of the residuals, it is also necessary to determine whether their spatial distribution is concentrated and whether they coincide with the location of critical structures. For example, if the residual is 0.12, it can be classified as a level two early warning if it is dispersed in multiple low-risk areas; if it is concentrated at the bottom corner of the pit, the end of the support, or in the area adjacent to the existing subway section, the level can be upgraded. When the drift rate is less than or equal to the critical phase transition point threshold, it indicates that the system as a whole has not yet entered a significant phase transition stage. At this time, the current monitoring status is maintained, and the second safety assessment warning result is determined in combination with the discrete residuals. Maintaining the current monitoring status does not indicate the cancellation of residual anomalies, but rather means temporarily delaying the triggering of high-level responses and only entering a process of continuous observation, local verification, or increasing the sampling frequency. For example, if the drift rate is 0.04 but the local residual is 0.11, the output can be to maintain the current construction pace and increase the intensity of monitoring in the abnormal area, rather than immediately requiring a complete shutdown. A further microscopic deduction can be given: assuming the critical phase transition threshold is 0.06, the drift rate at the current time M1 is 0.08 and the residual is 0.07, falling within the yellow boundary of the phase space, then the warning level is level two; at time M2, the drift rate rises to 0.11 and the residual increases to 0.16, falling into the red boundary, then it is upgraded to level one; conversely, if the drift rate at time M3 is only 0.03, but the residual is 0.12, then an observational warning can be generated, indicating a local anomaly but no signs of system-level topological instability are observed; If the drift rate is close to the threshold boundary, such as fluctuating repeatedly between 0.059 and 0.061, a hysteresis zone can be introduced. The warning will only be upgraded if multiple consecutive windows exceed the boundary, thus avoiding high-frequency oscillations in the warning state. If the overall discrete residual is high but it is known to be in a large-scale manual loading and unloading phase, the phase space boundary can be temporarily adjusted within the construction window. If both the drift rate and residual are missing or the confidence level is too low, the system will not output a clear warning level, but will output a message indicating insufficient data quality and require manual review. During the nighttime construction phase following continuous torrential rain at the transfer station's foundation pit, the system calculated a drift rate of 0.087 for the southwest corner area, with a local residual peak of 0.143, exceeding the phase transition point threshold and falling into the warning zone. At this point, the system generated a first-class safety assessment result, marked as a level-three response warning, recommending a halt to further excavation in that area and an immediate review of the precipitation parameters. After local water release was completed the following morning, the drift rate dropped to 0.048, but the residual was still 0.091. The system then switched to a second-class safety assessment result, maintaining the current monitoring status and continuing to sample and recheck the bottom of the southwest corner pit every 5 minutes. The purpose of this step is to integrate system-level trends and local anomalies into the phase space analysis framework, thereby achieving a more stable and hierarchical early warning judgment. In this embodiment, a corresponding multi-dimensional phase space hierarchical early warning mechanism is triggered based on a preset boundary division within the multi-dimensional phase space to generate an early warning level. The first safety assessment early warning result is then determined by combining the discrete residuals of the stress balance partial differential equation, including: Discretize the residuals based on the partial differential equation of stress equilibrium and generate a residual heatmap; Anomaly region localization processing is performed on the residual heat map to extract regions where the residual value is greater than the preset residual threshold. Combined with the equivalent stress and displacement gradient data extracted from the stress tensor field of the whole map and the multi-source physical monitoring data, stress concentration areas and displacement anomaly areas are determined respectively. The warning level, residual heat map, stress concentration area and displacement anomaly area are combined and processed to determine the first safety assessment warning result.

[0021] This embodiment provides a refined generation mechanism for the first safety assessment result. Specifically, in the previous layer scheme, although the warning level can be output, if only a level two or level one warning is given, it is still difficult for on-site technicians to quickly locate the risk location and formulate a response plan. Therefore, this embodiment further visualizes the discrete residual as a spatial heat map and automatically locates the stress concentration area and the displacement anomaly area, thereby making the warning result executable. The details are as follows: A residual heatmap is generated based on the discrete residuals; the residual value of each node or unit can be mapped to a color level, for example, 0 to 0.05 is blue, 0.05 to 0.10 is yellow, 0.10 to 0.15 is orange, and values ​​above 0.15 are red; if in a simplified 3×3 node plane, the residual matrix is ​​as follows: [...] ], second line [ ], third line [ If the two nodes in the lower middle section are displayed as red or orange, it will visually indicate a local concentration of abnormalities. Anomaly regions are located in the residual heatmap to identify stress concentration areas and displacement anomaly areas. Stress concentration areas can be determined by a combination of high residual and high equivalent stress, while displacement anomaly areas can be determined by a combination of high residual and abrupt displacement gradient changes. For example, if the equivalent stress in a certain area is significantly higher than that in the surrounding area when the residual is 0.14, it is classified as a stress concentration area. If the stress in another area does not reach the preset stress peak threshold, but the displacement difference between adjacent nodes is greater than the preset displacement gradient threshold within a preset spatial distance, it is classified as a displacement anomaly area. This achieves spatial decoupling and differentiation of different types of anomaly risks. Furthermore, the warning level, residual heat map, stress concentration area, and displacement anomaly area are merged to form the first safety assessment result. The merged result can be presented as a structured report, including: the current warning level, the planar location and depth range of the anomaly area, the anomaly type, and the recommended priority inspection objects. For example, a level two warning can be output: there is a continuous stress concentration area in the range of -18m to -22m at the bottom of the pit in the southwest corner; there is a displacement anomaly area near the end of the third support; it is recommended to prioritize checking the signs of sudden surge at the bottom of the pit and the stress state of the support nodes. To enhance the operability of the project, a micro-merging logic can also be provided. Assuming the warning level is level two, there are two red connected regions R1 and R2 in the heat map. R1 satisfies the high stress condition and R2 satisfies the high displacement gradient condition. In the report, they are marked as stress concentration area and displacement anomaly area, respectively. If a region satisfies both conditions, it can be marked as a composite anomaly area, and its handling priority is higher than that of a single anomaly area. If a scattered single-point anomaly appears on the heat map while the surrounding nodes are normal, it is first identified as an isolated anomaly and a retest is required, rather than being immediately identified as a continuous risk area. If the overlap between the stress concentration area and the displacement anomaly area is less than the set area threshold, they are output separately to avoid excessive merging that could lead to ambiguous positioning. If the heat map resolution is too low and the positioning accuracy is insufficient, a higher-density local sub-map is called to recalculate the area before the final report is output. After two consecutive rounds of early warning upgrades at the southwest corner of the transfer station's foundation pit, the system generated a residual heat map, showing a red strip-shaped area extending along the inner side of the underground continuous wall at the bottom corner of the pit, while orange high-gradient patches appeared near the end of the third steel support. After automatic positioning, the former was identified as a stress concentration area, and the latter as a displacement anomaly area. The final first-class safety assessment result not only included a level-three early warning, but also clearly pointed out two key inspection areas: the passive area at the southwest corner of the pit bottom and the western end node of the third support. The purpose of this step is to transform the abstract early warning judgment into an on-site report with spatial location, anomaly type and handling priority, thereby realizing the transformation from qualitative early warning of risk status to spatial location of abnormal area; In this embodiment, based on the safety assessment and early warning results, a structural adjustment instruction for the target large foundation pit is determined, including: Obtain a pre-defined structural support system adjustment strategy library; If the safety assessment warning result is the first safety assessment warning result, the first safety assessment warning result and the structural support system adjustment strategy library are matched to determine the target reinforcement instruction or the target decompression instruction. If the security assessment warning result is the second security assessment warning result, the order to maintain the status quo shall be determined; The target reinforcement command, target decompression command, or status quo maintenance command are output as structural adjustment commands.

[0022] This embodiment provides a mechanism for generating structural adjustment instructions. Specifically, a safety assessment result can be generated in the aforementioned process. However, if the result only stays at the level of a risk report, technicians still need to manually consult the handling manual, which results in a response delay. In highly sensitive conditions such as heavy rain, nighttime construction, or nearby operating subways, manual judgment may be delayed. Therefore, this embodiment automatically matches the early warning result with a preset structural support system adjustment strategy library to directly generate executable on-site adjustment instructions. The details are as follows: A pre-defined structural support system adjustment strategy library is obtained; this strategy library can establish multi-dimensional indexes according to anomaly type, anomaly area, warning level, and construction stage; for example, for a stress concentration zone at the bottom of the pit + high pore pressure + level three warning, corresponding measures include local grouting reinforcement, mitigating dewatering, and suspending excavation of adjacent blocks; for anomaly zone at the support end displacement + level two warning, corresponding measures include reviewing prestress, adding temporary supports, and shortening the monitoring cycle; for residual anomalies that are generally stable and only locally below warning level two, the corresponding measures are to maintain the status quo and increase the frequency of observation. If the safety assessment result is classified as Category I, it is matched with the strategy library to determine the target reinforcement command or the target decompression command. Reinforcement commands typically target load-bearing and deformation control, such as adding temporary steel supports, local grouting, adding steel girders, and increasing the soil reinforcement range. Decompression commands typically target anomalies in pore pressure and soil and water pressure, such as reducing the rate of precipitation, adding water release wells, staged excavation and unloading, and temporarily suspending heavy machinery loading. In addition to introducing the warning level as an assessment dimension, the location and type of the anomaly must also be considered for comprehensive judgment during matching. For example, if both are Level III warnings, and the anomaly is located in the passive zone at the bottom of the pit and pore pressure is dominant, then the decompression command is given priority; if the anomaly is located at a support node and stress concentration is significant, then the reinforcement command is given priority. If the safety assessment result is Category II, a "maintain status quo" instruction is issued. Maintaining status quo here does not mean no action, but rather keeping the current support system unchanged and requiring continued monitoring at the established frequency or slightly more frequent. For example, the output could be "maintain status quo"; the sampling cycle for the southwest corner could be shortened from 10 minutes to 5 minutes; and the stress tensor field of the entire map could be recalculated before the next excavation step. The matching process can be illustrated with a simplified matching example; assuming there are three rules in the strategy library: Rule 1 corresponds to a stress concentration area located at the support end and a warning level of... Level 2, output verification of axial force + addition of temporary supports; Rule 2 corresponds to an abnormal displacement zone located at the bottom corner of the pit and residual Output: Grouting at the bottom of the pit + Pause excavation in this area; Rule 3 corresponds to a drift rate that does not exceed the threshold and the residual... The output remains unchanged; when the input result is a level 2 warning, an abnormal displacement zone at the southwest corner of the pit bottom, and a residual of 0.14, the system will prioritize rule 2 and generate a target reinforcement instruction. If an assessment result matches multiple strategies, they can be merged according to the priority matrix. For example, high-priority instructions such as shutdown / relief can be executed first, followed by supplementary instructions such as reinforcement. If there is no complete match in the strategy library, suggested instructions are generated based on the closest anomaly type and region, and a manual review is required. If an instruction has already been executed on-site and the monitoring results have not improved, the handling level will be automatically upgraded in the next round of matching to avoid repeatedly outputting ineffective measures. After a Level 3 warning, a stress concentration zone at the bottom of the pit, and a continuous increase in pore pressure were identified at the southwest corner of the transfer station's foundation pit, the system queried the strategy library and automatically generated a set of structural adjustment instructions. These instructions included: First, immediately suspending further excavation of the adjacent earthwork at the southwest corner; Second, reducing the pumping rate of the dewatering wells in the area and activating the release wells; Third, implementing segmented grouting reinforcement in the passive zone at the bottom of the pit; Fourth, verifying the axial force at the western end node of the third support. If the drift rate and residual simultaneously decrease within the next hour, the next output will switch to an instruction to maintain the status quo and continue monitoring. The purpose of this step is to directly link the risk identification results to engineering response actions, thereby achieving a closed-loop execution of monitoring, calculation, early warning and adjustment.

[0023] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent calculation and graded early warning safety assessment of stress state in large foundation pits, characterized in that, The method includes: The initial environmental parameters and three-dimensional structural model of the target large foundation pit are obtained, and the initial material properties and constitutive parameters of the graph edge are determined accordingly. Based on the initial material properties and graph edge constitutive parameters, the initial environmental parameters and three-dimensional structural model are transformed to determine a heterogeneous multiphysics tensor graph containing observation anchor nodes, hidden computation nodes, and edges connecting the two. Acquire multi-source physical monitoring data of the target large foundation pit at the current moment, including stress and displacement monitoring data; The state of the observation anchor node is updated based on the multi-source physical monitoring data to determine the state of the target observation anchor node containing energy stress characteristics. Based on the target observation anchor node state and the preset physical information constrained spatiotemporal tensor graph assimilation network, the hidden computation nodes are evolved to determine the full graph stress tensor field; wherein, the network uses a preset physical conservation law as the loss function for weight constraints and includes a multiphysics coupled graph operator; Extract the features of the stress tensor field of the whole image, determine the Laplace matrix spectrum features and the discrete residuals of the stress equilibrium partial differential equations based on the physical conservation law; Based on the spectral characteristics of the Laplace matrix, determine the algebraic connectivity drift rate; Based on the algebraic connectivity drift rate and the discrete residual of the stress balance partial differential equation, the safety assessment and early warning results are determined. Based on the safety assessment and early warning results, a structural adjustment instruction for the target large foundation pit is determined.

2. The intelligent calculation and graded early warning safety assessment method for the stress state of a large foundation pit according to claim 1, characterized in that, The process of transforming the initial environmental parameters and the three-dimensional structural model based on the initial material properties and graph edge constitutive parameters to determine a heterogeneous multiphysics tensor graph containing observation anchor nodes, hidden computation nodes, and edges connecting the two includes: Based on the initial environmental parameters, determine the spatial coordinates of the graph nodes; Based on the spatial coordinates of the graph nodes, the initial properties of the material, and the constitutive parameters of the graph edges, an initial tensor graph containing the nodes corresponding to the spatial coordinates of the graph nodes is constructed. Obtain preset physical sensor location information; Based on the physical sensor location information, the nodes of the initial tensor graph are classified to determine the observation anchor node and the hidden computation node; The heterogeneous multiphysics tensor graph is generated based on the observation anchor node, the hidden computation node, and the initial tensor graph.

3. The intelligent calculation and graded early warning safety assessment method for the stress state of a large foundation pit according to claim 1, characterized in that, The method for evolving the hidden computation nodes to determine the full-graph stress tensor field, based on the target observation anchor node state and a preset physical information-constrained spatiotemporal tensor graph assimilation network, includes: Based on the multiphysics coupling graph operator, the energy stress characteristics in the target observation anchor node state are extracted; The energy stress features are propagated along the edge to the hidden computing node to determine the node feature update sequence; The node feature update sequence is subjected to local iterative update processing to determine the full-image stress tensor field.

4. The intelligent calculation and graded early warning safety assessment method for stress state of large foundation pits according to claim 1, characterized in that, The step of extracting the features of the full-image stress tensor field, determining the Laplace matrix spectral features, and the discrete residuals of the stress equilibrium partial differential equations based on the physical conservation laws includes: The full-image stress tensor field is subjected to matrix decomposition to determine the target Laplace matrix; Spectral analysis is performed on the target Laplacian matrix to determine its spectral characteristics; Based on the stress equilibrium partial differential equation corresponding to the preset physical conservation law, residual calculation processing is performed on the stress tensor field of the whole figure to determine the discrete residual of the stress equilibrium partial differential equation.

5. The intelligent calculation and graded early warning safety assessment method for the stress state of a large foundation pit according to claim 1, characterized in that, The step of determining the algebraic connectivity drift rate based on the spectral characteristics of the Laplace matrix includes: Extract the second smallest eigenvalue from the spectral features of the Laplacian matrix to determine the target algebraic connectivity; Obtain the historical time series corresponding to the target large foundation pit, and extract the historical algebraic connectivity corresponding to the historical time series; Based on the historical algebraic connectivity and the target algebraic connectivity, time derivative calculation is performed to determine the algebraic connectivity drift rate.

6. The intelligent calculation and graded early warning safety assessment method for stress state of large foundation pits according to claim 1, characterized in that, The determination of safety assessment and early warning results based on the algebraic connectivity drift rate and the discrete residuals of the stress balance partial differential equation includes: Obtain the preset critical phase transition point threshold; Construct a multidimensional phase space with the algebraic connectivity drift rate and the discrete residuals of the stress balance partial differential equation as dimensions; When the algebraic connectivity drift rate is greater than the critical phase transition point threshold, the corresponding multidimensional phase space hierarchical early warning mechanism is triggered according to the preset boundary division in the multidimensional phase space to generate an early warning level, and the first safety assessment early warning result is determined by combining the discrete residual of the stress balance partial differential equation. If the algebraic connectivity drift rate is less than or equal to the critical phase transition point threshold, the current monitoring status is maintained, and the second safety assessment warning result is determined by combining the discrete residual of the stress balance partial differential equation. The first security assessment warning result or the second security assessment warning result shall be used as the security assessment warning result.

7. The intelligent calculation and graded early warning safety assessment method for stress state of large foundation pits according to claim 6, characterized in that, The step of triggering a corresponding multi-dimensional phase space hierarchical early warning mechanism based on a preset boundary division within the multi-dimensional phase space to generate an early warning level, and determining the first safety assessment early warning result by combining the discrete residual of the stress balance partial differential equation, includes: Based on the stress balance partial differential equation, the residuals are discretized to generate a residual heatmap; The residual heatmap is processed to locate abnormal regions, and regions with residual values ​​greater than a preset residual threshold are extracted. Combined with the equivalent stress and displacement gradient data extracted from the stress tensor field of the whole map and the multi-source physical monitoring data, stress concentration areas and displacement anomaly areas are determined respectively. The warning level, the residual heat map, the stress concentration area, and the displacement anomaly area are combined to determine the first safety assessment warning result.

8. The intelligent calculation and graded early warning safety assessment method for stress state of large foundation pits according to claim 6, characterized in that, The step of determining the structural adjustment instructions for the target large foundation pit based on the safety assessment and early warning results includes: Obtain a pre-defined structural support system adjustment strategy library; If the safety assessment warning result is the first safety assessment warning result, the first safety assessment warning result and the structural support system adjustment strategy library are matched to determine the target reinforcement command or the target decompression command. If the security assessment warning result is the second security assessment warning result, a status quo maintenance order is determined; The target reinforcement command, the target decompression command, or the status quo maintenance command are output as the structural adjustment command.

Images