Tunnel structure safety monitoring method and system based on multi-source information fusion

Abstract

The application provides a tunnel structure safety monitoring method and system based on multi-source information fusion, and relates to the technical field of tunnel structure health monitoring and safety warning. The method comprises the following steps: obtaining a BIM model of a tunnel to be monitored, dividing components according to stress units, extracting spatial positions and connection relationships, determining connection edge strength according to stress transmission relationships and generating a tunnel topology graph; mapping monitoring data to component nodes, inputting a multi-scale space-time prediction model, fusing local, cross-section and overall spatial features and time sequence features, and predicting component deformation; when a predicted value exceeds a threshold value, combining component distance, force transmission relationships and historical risk propagation to determine risk influence strength, calculating risk increment and a comprehensive risk index and grading warning; according to errors between subsequent measured values and predicted values, updating prediction and risk propagation parameters of abnormal components and neighborhoods, and realizing closed-loop monitoring and warning.

Description

Technical Field

[0001] This invention relates to the field of tunnel structure health monitoring and safety early warning technology, specifically to a tunnel structure safety monitoring method and system based on multi-source information fusion. Background Technology

[0002] Tunnel engineering involves complex geological conditions, numerous construction procedures, and dynamic changes in structural stress state with construction progress. It faces long-term risks such as surrounding rock deformation, abnormal support, lining cracking, arch subsidence, and peripheral convergence. To ensure tunnel structural safety, continuous monitoring and early warning of the tunnel's structural condition are necessary.

[0003] Currently, BIM models and various sensor data have been introduced into tunnel structure safety monitoring. Some solutions use BIM models for visualizing monitoring data; however, the spatial locations, connections, and structural attributes of components in the BIM model are not converted into a topological structure that can participate in calculation and analysis, resulting in the processing of monitoring data being disconnected from the actual stress mechanism of the tunnel structure. Some solutions use time-series prediction models to extrapolate trends from monitoring data; however, such models typically treat the data from each monitoring point as independent time series, failing to consider the spatial correlation between components due to stress transmission, making it difficult to characterize the coordinated deformation features of the tunnel structure under layered construction and dynamic stress conditions.

[0004] Some solutions trigger alarms when monitored values ​​exceed fixed thresholds, but lack quantitative analysis of the propagation path and impact range of abnormal conditions to adjacent components, thus failing to provide early warning. Furthermore, existing solutions often employ models that are trained only once and used permanently, lacking a closed-loop mechanism for continuously correcting model parameters and risk propagation parameters based on measured data. This results in a decrease in model accuracy as surrounding rock conditions and construction stages change.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for monitoring the safety of tunnel structures based on multi-source information fusion, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for monitoring the safety of tunnel structures based on multi-source information fusion, comprising the following steps: Obtain the BIM model of the tunnel to be monitored, divide the tunnel into multiple components according to the stress unit, and extract the spatial position and connection relationship of each component; The strength of the connecting edge is determined based on the force transfer relationship between the components, and a tunnel topology map is generated for calculation and analysis. The connecting edge connects two components that have a force transfer relationship. Acquire the monitoring data corresponding to each component, map the monitoring data to the corresponding component nodes in the tunnel topology diagram, and form the status data of each component node; The state data of each component node is input into the multi-scale spatiotemporal prediction model. Spatial features are aggregated in the adjacent range of the component, the cross-sectional range and the overall range of the tunnel, and the temporal features are encoded. The spatial and temporal features of the three ranges are fused to output the deformation prediction value of each component in the next prediction period. In response to the predicted deformation value of any component exceeding the deformation threshold of that component, the risk impact intensity between two components is determined based on the spatial distance between any two components, the structural force transmission relationship reflected by the strength of the connection edge, and the historical risk propagation. The risk increment of the affected component is calculated based on the risk impact intensity. The comprehensive risk index is determined by combining the deformation degree and deformation growth rate of the affected component, and the warning level is determined based on the comprehensive risk index. The subsequent actual monitoring values ​​are obtained and compared with the deformation prediction values ​​of the corresponding components to obtain the prediction error. In response to the prediction error continuously exceeding the preset threshold, the local prediction parameters and risk propagation parameters of the abnormal components and adjacent components are updated to form a monitoring closed loop.

[0008] Furthermore, the connecting edges include three categories: direct force transmission edges, indirect force transmission edges, and construction disturbance edges; The initial strength of the directly transmitted force edge is greater than that of the indirectly transmitted force edge; the initial strength of the construction disturbance edge is determined based on the time difference of the construction process.

[0009] Furthermore, the direct force transmission edge is established by the connection relationship between adjacent components in the same ring, the connection relationship between corresponding components in the front and rear rings, and the connection relationship between the lining components and the invert arch components; The indirect force transmission edge is established by the connection relationship between the initial support and the secondary lining, as well as the connection relationship between the transverse passage and the main tunnel connecting components; The construction disturbance edge is established based on the current construction stage and construction progress.

[0010] Furthermore, the spatial feature aggregation includes three scales: local component scale, mesoscopic cross-sectional scale, and global tunnel scale; The local component scale is aggregated with directly adjacent components, and fine-grained deformation features of single components and directly adjacent components are extracted; At the mesoscopic cross-sectional scale, components with the same cross-section or adjacent predetermined number of rings are used as the aggregation range to extract multi-ring collaborative deformation features; The global tunnel scale uses all component nodes of the entire tunnel as the aggregation range to extract the risk evolution characteristics of the entire tunnel; The structural force transmission relationship embodied by the strength of the connecting edge is embedded as a physical constraint into the spatial feature aggregation process of the multi-scale spatiotemporal prediction model, so that the strength of the connecting edge varies within a set range near the initial value of engineering experience. After outputting the deformation prediction values, component-level attribution analysis is performed on the multi-scale spatiotemporal prediction model to calculate the contribution of each input feature to the comprehensive risk index, and output the risk contribution factors with the highest contribution ranking and their contribution percentage.

[0011] Furthermore, the spatial features at the three scales are fused and then incorporated into a hierarchical attention mechanism to automatically learn the weight allocation of local component scale, mesoscopic cross-sectional scale, and global tunnel scale under different working conditions. The weights are adaptively adjusted according to the tunnel construction stage and operation status.

[0012] Furthermore, the contribution of spatial distance, structural force transmission relationship reflected by the strength of the connecting edge, and historical risk propagation to the intensity of risk impact adaptively adjusts with changes in the surrounding rock grade and construction stage.

[0013] Furthermore, the comprehensive risk index is determined by the degree of deformation of the affected component itself, the rate of deformation growth, the risk increment, and the degree of historical instability. The warning levels are divided into four levels based on the range of values ​​of the comprehensive risk index: safety level, attention level, warning level, and emergency level.

[0014] Furthermore, when the prediction error obtained by comparing the subsequent actual monitoring value with the deformation prediction value of the corresponding component continuously exceeds the preset threshold, the local subgraph incremental update is triggered. Only the local prediction parameters corresponding to the local subgraph composed of abnormal components and components within the adjacent set ring number range are updated, and the local prediction parameters corresponding to components outside the update range are not updated. The local prediction parameters include the parameters used in spatial feature aggregation and temporal feature encoding for anomalous components and components within the adjacent set ring number range, and update the risk propagation parameters between anomalous components and adjacent components.

[0015] Furthermore, as the tunnel construction progresses, for each set number of construction progresses completed, corresponding component nodes are automatically added, and the connection edges are completed based on the force transfer relationship between the new components and existing components, so that the total number of component nodes in the tunnel topology diagram increases with the tunnel construction progress.

[0016] The present invention also provides a tunnel structure safety monitoring system based on multi-source information fusion, the system being used to execute the above-described tunnel structure safety monitoring method based on multi-source information fusion, comprising: The component division module is used to obtain the BIM model of the tunnel to be monitored, divide the tunnel into multiple components according to the stress unit, and extract the spatial position and connection relationship of each component. The topology construction module determines the strength of the connecting edges based on the force transfer relationship between components and generates a tunnel topology map for calculation and analysis. The connecting edges connect two components that have a force transfer relationship. The status data mapping module is used to obtain the monitoring data corresponding to each component, map the monitoring data to the corresponding component nodes in the tunnel topology diagram, and form the status data of each component node. The multi-scale spatiotemporal prediction module inputs the state data of each component node into the multi-scale spatiotemporal prediction model, aggregates spatial features in the adjacent range of the component, the cross-sectional range and the overall range of the tunnel, encodes the temporal features, and merges the spatial and temporal features of the three ranges to output the deformation prediction value of each component in the next prediction period. The risk propagation early warning module responds when the predicted deformation value of any component exceeds the deformation threshold of that component. Based on the spatial distance between any two components, the structural force transmission relationship reflected by the strength of the connecting edge, and the historical risk propagation situation, it determines the risk impact intensity between the two components, calculates the risk increment of the affected component based on the risk impact intensity, determines the comprehensive risk index by combining the deformation degree and deformation growth rate of the affected component, and determines the early warning level based on the comprehensive risk index. The closed-loop update module is used to obtain subsequent real monitoring values, compare them with the deformation prediction values ​​of the corresponding components to obtain the prediction error, and update the local prediction parameters and risk propagation parameters of the abnormal components and adjacent components in response to the prediction error continuously exceeding the preset threshold, thus forming a monitoring closed loop.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention transforms the force transmission relationships of tunnel components into calculable connection edge strengths, generating a tunnel topology map for computational analysis. This transforms the spatial location and connection relationships of components from mere visual information into a mechanical constraint basis for predictive calculations and risk analysis. Since the connection edge strengths directly originate from the force transmission relationships between components, the processing of monitoring data remains consistent with the actual stress mechanism of the tunnel structure. By aggregating spatial features at the component adjacency range, cross-sectional range, and overall tunnel range, and then encoding and fusing temporal features, the prediction model can simultaneously capture the fine-grained deformation characteristics of individual components and adjacent components, the collaborative deformation characteristics between the same cross-section or adjacent ring segments, and the overall risk evolution characteristics across the entire tunnel. Because the three ranges correspond to different stress influence ranges in tunnel structural analysis, the prediction results can reflect the spatial coupling characteristics under layered construction and dynamic stress conditions, thus outputting component-level deformation prediction values, rather than merely extrapolating trends at the measurement point level.

[0018] This invention determines the risk impact intensity between components based on spatial distance, structural force transmission relationship reflected by the strength of connecting edges, and historical risk propagation. When the deformation prediction value exceeds a threshold, it calculates the risk increment of the affected components and determines a comprehensive risk index by combining the degree of deformation and the deformation growth rate. This expands the analysis from single-point over-limit alarms to quantitative analysis of risk propagation paths and impact ranges. Since the risk impact intensity includes the structural force transmission relationship, the risk propagation calculation results can reflect the differentiated force transmission characteristics between tunnel components, thus improving the accuracy of early warning. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall method flow of the present invention.

[0020] Figure 2 This is a screenshot of a Revit 2021 highway tunnel BIM model in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram illustrating the adaptive comparison of three-scale attention weights in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the decomposition of risk contribution factors in an embodiment of the present invention.

[0023] Figure 5 This is a schematic diagram comparing the prediction error before and after closed-loop incremental learning in an embodiment of the present invention.

[0024] Figure 6 This is a schematic diagram illustrating the early warning levels based on the comprehensive risk index in an embodiment of the present invention.

[0025] Figure 7 This is a structural block diagram of the overall system of the present invention. Detailed Implementation

[0026] 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.

[0027] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0028] Example: Please see Figure 1-6 The present invention provides a technical solution: A method for monitoring the safety of tunnel structures based on multi-source information fusion, comprising the following steps: Step 1: Obtain the BIM model of the tunnel to be monitored, divide the tunnel into multiple components according to the stress unit, and extract the spatial position and connection relationship of each component.

[0029] Please see Figure 6 The structure is divided into multiple ring segments according to the stress unit. The color of each ring segment directly maps the comprehensive risk index of the component. The red rings are concentrated in the middle section of the pipe and gradually transition to blue at both ends, which intuitively presents the risk propagation effect described in this embodiment: the anomaly starts from a certain component and spreads to the adjacent components according to the strength of the connecting edge (force transmission relationship) and the spatial distance, forming a risk gradient distribution with high risk in the middle and low risk at both ends in the longitudinal direction of the tunnel.

[0030] In this step, the BIM model originates from the digital model continuously maintained during the tunnel design and construction phases. Taking the Shirenliang Highway Tunnel as an example, the BIM model includes objects such as the main tunnel, vehicular cross passages, pedestrian cross passages, initial support, secondary lining, invert arch, drainage system, anchor bolts, steel frame, and auxiliary components. Each component is assigned a unique identifier code and its three-dimensional coordinates, ring number, cross section, material type, support type, cross-sectional dimensions, and spatial connection relationships with other components are recorded.

[0031] Dividing a tunnel into multiple components based on load-bearing units means using the units within the tunnel structure that bear independent load-bearing functions as the basis for division, rather than simply dividing them according to geometric shape or modeling convenience. For example, the crown section, waist section, and invert section of a single-ring lining are each treated as independent components because the crown section mainly bears the bending moment and axial force generated by the pressure of the surrounding rock above, the waist section bears the lateral pressure of the surrounding rock and participates in circumferential force transmission, and the invert section bears the bottom reaction force and closes the load-bearing ring of the lining. The three play different mechanical roles in the structural stress. If the entire ring lining is treated as a single component, it is impossible to express the differences in force transmission between the crown section and the waist section, and between the waist section and the invert section, in subsequent topology diagrams, nor is it possible to locate local anomalies at the component level.

[0032] Another reason for dividing the tunnel topology by force-bearing units rather than geometric units is that the strength of the connecting edges in subsequent steps needs to be determined based on the force transmission relationship between components. Only when the component division boundaries correspond to the boundaries of the mechanical force transmission paths can the connection relationships between components truly reflect the force transmission logic within the structure. If components are divided according to geometric rules of equal length or equal arc length, the cutting points may fall within the continuous force-bearing region, causing the originally continuous force transmission path to be artificially truncated. This will result in a large number of connecting edges in the topology diagram that do not have clear mechanical meaning, thus affecting the accuracy of subsequent prediction models and risk propagation calculations.

[0033] When extracting the spatial location and connection relationships of each component, the spatial location includes the three-dimensional coordinates of the component's centroid or key control points. The connection relationships include the joint connection relationships between adjacent components within the same ring, the inter-ring joint connection relationships between corresponding components in front and behind rings, the arch foot connection relationships between lining components and invert arch components, the contact connection relationships between initial support and secondary lining, and the node connection relationships between transverse passages and main opening connecting components. These connection relationships can be obtained in the BIM model through the spatial intersection relationships between components, the constraint relationships during modeling, or the association identifiers in the component attributes.

[0034] As tunnel construction progresses, for each set number of construction progresses completed, corresponding component nodes are automatically added, and connection edges are completed based on the force transfer relationship between the new components and existing components, so that the total number of component nodes in the tunnel topology diagram increases with the tunnel construction progress.

[0035] The determination of the number of rings needs to balance the update frequency and timeliness of the topology map. If the number is too small, the topology map will be updated frequently, and each update requires recalculating the connection relationship between new and existing components and adjusting the map structure, increasing computational overhead. If the number is too large, the topology map will lag behind the actual construction status, and components that have been completed but not yet included in the topology map cannot participate in the prediction calculation, affecting the prediction accuracy. In this embodiment, the completion of 3 rings of lining construction is used as the trigger condition for one topology expansion. 3 rings of lining correspond to the typical progress of one construction cycle. Within this progress, the force transmission relationship between new and existing components is clear, and the disturbance influence range of the tunnel face excavation on adjacent components is basically stable, achieving a balance between update frequency and topology timeliness. In other embodiments, the number of rings can also be adjusted to 1 ring, 2 rings, or 5 rings according to the surrounding rock grade, construction method, and monitoring accuracy requirements, as long as the synchronization requirement between the topology map and the actual construction status is met.

[0036] Step 2: Determine the strength of the connecting edge based on the force transfer relationship between the components, and generate a tunnel topology map for calculation and analysis. The connecting edge connects two components with a force transfer relationship.

[0037] A tunnel topology graph is a graph data structure with components as nodes and the force transmission relationships between components as connecting edges. Connecting edges link two components with a force transmission relationship, and the strength of the edge reflects the tightness of the force transmission between the two components. Higher edge strength indicates that the deformation or change in stress state of one component is more easily transmitted to the other component; lower edge strength indicates a weaker mechanical coupling between the two components.

[0038] Connection edges are classified into three types: direct force transmission edges, indirect force transmission edges, and construction disturbance edges. The initial strength of direct force transmission edges is greater than that of indirect force transmission edges; the initial strength of construction disturbance edges is determined based on the time difference of construction procedures.

[0039] Direct force-transferring edge connections involve direct mechanical contact and a force transmission path between two components, resulting in high force transfer efficiency. For example, the arch crown and arch waist sections within the same ring are in direct contact via lining joints. The bending moment and axial force of one component can be directly transferred to the adjacent component through the joint, without any significant force transmission loss medium in between. Indirect force-transferring edge connections, on the other hand, involve force transmission between two components through an intermediate medium or indirect path, leading to lower force transfer efficiency. For instance, a waterproof layer is typically installed between the initial support and the secondary lining. Force transmission between them requires compression deformation of the waterproof layer and the contact surface, resulting in a significant attenuation of force during transmission due to the presence of a flexible medium in the force transmission path. Therefore, the initial strength of direct force-transferring edges is set to be greater than that of indirect force-transferring edges to reflect the fundamental difference in force transmission efficiency between the two types of connections.

[0040] The initial strength of the construction disturbance edge is determined based on the time difference between construction procedures. The time difference between procedures refers to the interval between the sequential construction of two components. The smaller the time difference between procedures, the stronger the construction disturbance and the greater the edge strength; conversely, the larger the time difference between procedures, the weaker the disturbance and the smaller the edge strength. This is because the disturbance effect of tunneling, blasting, or support operations on adjacent already constructed components at the working face decreases with increasing time and spatial distance, and the time difference between construction procedures directly reflects the temporal proximity of this disturbance.

[0041] The direct force transmission edge is established by the connection relationship between adjacent components in the same ring, the connection relationship between corresponding components in the front and rear rings, and the connection relationship between the lining component and the invert arch component; the indirect force transmission edge is established by the connection relationship between the initial support and the secondary lining, and the connection relationship between the transverse passage and the main tunnel connecting component; the construction disturbance edge is established according to the current construction stage and construction progress.

[0042] The engineering scenarios corresponding to each type of connection edge are explained one by one. Direct force transmission edges between adjacent components within the same ring, such as the direct force transmission between the arch crown and arch waist sections in the same ring lining through a circumferential joint, or between the arch waist section and the invert arch section through an arch foot connection. Direct force transmission edges between corresponding components in preceding and following rings, such as the direct force transmission between the arch crown sections of the Nth ring and the N+1th ring through a circumferential construction joint, where longitudinal axial force and bending moment can be transferred longitudinally between adjacent rings. Direct force transmission edges between lining components and invert arch components, such as the connection between the secondary lining arch wall and the invert arch infill, where the upper load is transferred to the invert arch through a structural connection at the arch foot.

[0043] The indirect force transmission edge between the initial support and the secondary lining reflects the force transmission path through which the initial support indirectly transmits the surrounding rock pressure to the secondary lining via the waterproof layer and contact surface. The indirect force transmission edge between the cross passage and the main tunnel connection components reflects the path through which the disturbance of the cross passage excavation on the main tunnel structure is indirectly transmitted via the connection nodes. The construction disturbance edge is established based on the current construction stage and progress. For example, when the tunnel face reaches a certain mileage, it causes disturbance to both the unexcavated area ahead and the constructed area behind. The construction disturbance edge connects the construction activities near the tunnel face with the adjacent constructed components, reflecting the immediate disturbance impact of the tunneling operation on the constructed components.

[0044] In this embodiment, the initial strength of the direct force transmission edge is set to 0.8, and the initial strength of the indirect force transmission edge is set to 0.3. These initial values ​​are derived from engineering experience regarding the differences in force transmission efficiency between different connection methods; the direct force transmission efficiency is approximately 2 to 3 times that of the indirect force transmission. The initial values ​​can be further calibrated on-site: paired measuring points are set up on both sides of the joint between adjacent components in the same ring, and the synchronous deformation sequence of adjacent components is collected over a period of time. The correlation coefficient and deformation transfer ratio of the deformation of adjacent components are calculated, and the deformation transfer ratio is used as the basis for the measured correction of the direct force transmission edge strength. Pressure cells or displacement gauges are set up between the initial support and the secondary lining to measure the degree of pressure transmission and deformation coordination between the two, thereby calibrating the indirect force transmission edge strength. The initial values ​​after on-site calibration are more consistent with the surrounding rock conditions and support parameters of the specific project. When on-site calibration is not carried out, the above-mentioned empirical initial values ​​can be directly used. Since the connection edge strength is a learnable parameter that is fine-tuned near the initial values ​​during subsequent model training, the initial values ​​only need to ensure reasonable magnitude and correct relative size, rather than pursuing absolute precision. The initial strength of the construction disturbance edge is determined by normalizing the process time difference. In this embodiment, an exponential decay form is adopted, and the normalization formula is: in, The baseline disturbance strength is set to 1.0 when the process time difference is zero. The lower limit of the disturbance intensity is set to 0.1, which is used to preserve the weak construction correlation between components that are far from the working face; This refers to the time difference between the construction procedures of the two components; As the characteristic time constant for disturbance attenuation, this embodiment takes approximately one-third of the maximum reference value, ensuring that the disturbance intensity has sufficiently attenuated and approaches the lower limit when the process time difference reaches the maximum reference value. The maximum reference value is determined based on the typical cycle of tunnel construction, for example, a time span of 10 construction cycles. The reason for using an exponential decay form is that the disturbances generated during tunnel face excavation, blasting, and support operations decay approximately exponentially with distance and time. A negative exponential function can characterize the rapid decay of the disturbance's impact in the early stages and its asymptotically gradual decline in the later stages, naturally reaching its upper limit when the time difference between the procedures is zero. As the time difference between processes approaches infinity, it asymptotically converges to the lower limit. This avoids the problem of negative values ​​or the need for manual truncation in linear attenuation under large time differences between construction processes. When the disturbance attenuation is approximately linear, a linear attenuation form can also be used, i.e., the strength of the construction disturbance edge. according to Calculation, where As the maximum reference value, both forms satisfy the qualitative relationship that the smaller the time difference between processes, the greater the strength of the edge, and the larger the time difference between processes, the smaller the strength of the edge. In engineering, the appropriate form can be selected based on the actual disturbance attenuation characteristics. The normalized strength of the construction disturbance edge is on the same order of magnitude as the initial strength of the directly transmitted and indirectly transmitted edges, and can be directly used in subsequent topology calculations and risk impact strength calculations.

[0045] Step 3: Obtain the monitoring data corresponding to each component, map the monitoring data to the corresponding component nodes in the tunnel topology diagram, and form the status data of each component node.

[0046] The monitoring data originates from monitoring points deployed on the tunnel structural components. These monitoring points are located on the components, and the data is written to the corresponding nodes based on the correspondence between the monitoring points and the components. Monitoring types include at least settlement values ​​and surrounding convergence values, and can be extended to stress, temperature, humidity, and surrounding rock displacement when necessary. Taking the Shirenliang Highway Tunnel as an example, the monitoring data is stored in CSV format, with each record including a timestamp, monitoring point number, monitoring type, and original measurement value.

[0047] Before mapping the monitoring data to the topology graph nodes, the raw monitoring data needs to be preprocessed. Missing values ​​are filled in using linear interpolation, with the following interpolation formula: in, For timestamps of missing data, and These are the timestamps of the most recent valid data before and after the missing time. and These are the monitoring values ​​at the corresponding times. The missing monitoring values ​​are obtained through interpolation. Linear interpolation is used because the settlement and surrounding convergence of the tunnel structure are continuous and gradual deformation processes, without abrupt changes within the sampling interval. The change in monitoring values ​​between two adjacent valid data points is approximately linear. Therefore, using a linear function determined by the line connecting the two points to fill in the missing values ​​reflects this continuous change pattern without introducing additional parameters. It has low computational cost, can be performed online in real time, and avoids overshoot oscillations at data gaps, unlike higher-order polynomial interpolation, thus preventing spurious fluctuations from interpolating and interfering with subsequent predictions. Linear interpolation is suitable for filling in short-term missing data, provided that the change in monitoring data between adjacent valid data points is approximately linear. For long-term or continuously missing data segments, the actual change between adjacent valid data points may deviate from linearity, reducing the accuracy of linear interpolation. In such cases, spatial interpolation can be performed by combining monitoring data from adjacent components, or the missing segments can be directly marked as invalid data and masked during model training to ensure the reliability of the training samples.

[0048] Abnormal abrupt changes are identified and removed using a sliding window. The length of the sliding window can be determined based on the sampling frequency of the monitoring data. For example, a window length of 10 sampling points is used. The mean and standard deviation of the data within the window are calculated. When a sample point deviates from the mean by more than three times the standard deviation, it is identified as an abnormal abrupt change and removed. The removed data points are then interpolated and filled in using the missing value handling method.

[0049] Min-max normalization is used to eliminate dimensional differences for different types of monitoring quantities. The normalization formula is: in, These are the original monitoring values. This is the minimum value for this monitoring type in historical data. This represents the maximum value of this monitoring type in historical data. The value is the normalized value, ranging from [0,1]. Normalization aims to eliminate the adverse effects on subsequent model training caused by differences in dimensions and numerical ranges between settlement, convergence, and stress values, ensuring comparability of different monitoring quantities in the feature space. Min-max normalization is used because it maps monitoring quantities with different dimensions and value ranges to a fixed interval [0,1], maintaining the relative size and distribution of the original data. The calculation relies only on the maximum and minimum values ​​of the historical data for that monitoring quantity, eliminating the need to estimate the mean and variance, allowing for online incremental updates. Furthermore, the value range [0,1] matches the numerical stability interval of subsequent attention mechanisms and activation functions, facilitating model convergence. Compared to zero-mean normalization, which requires data centering, alters the data distribution, and is unsuitable for monitoring quantities that do not conform to a normal distribution, min-max normalization makes no assumptions about the data distribution and is more suitable for various types of monitoring data, such as settlement, convergence, and stress. This normalization method is quite sensitive to outliers. Therefore, this application first performs outlier elimination and then performs min-max normalization. The combination of the two can avoid the normalization boundary being amplified by individual outliers, making the normalization result stable and reliable.

[0050] Node status data includes spatial attributes, structural attributes, monitoring attributes, prediction attributes, and risk attributes. Spatial attributes include the three-dimensional coordinates of the component; structural attributes include ring number, material type, support type, and cross-sectional dimensions; monitoring attributes include settlement value, convergence value, rate of change, and cumulative amount; prediction attributes include predicted future deformation values; and risk attributes include the current risk index. These attributes together constitute a complete state description of the component node in the topology graph, providing input for subsequent multi-scale spatiotemporal prediction and risk propagation calculations.

[0051] Step 4: Input the state data of each component node into the multi-scale spatiotemporal prediction model, aggregate spatial features in the adjacent range of the component, the cross-sectional range and the overall range of the tunnel, encode the temporal features, fuse the spatial features and temporal features of the three ranges, and output the deformation prediction value of each component in the next prediction period.

[0052] The multi-scale spatiotemporal prediction model uses the tunnel topology map as the input structure, aggregates spatial features by utilizing the connection relationships between nodes in the topology map, and encodes temporal features by utilizing the time series of node state data.

[0053] Spatial feature aggregation includes three scales: local component scale, mesoscopic cross-sectional scale, and global tunnel scale. At the local component scale, the aggregation scope is based on directly adjacent components, extracting fine-grained deformation features of single components and directly adjacent components. At the mesoscopic cross-sectional scale, the aggregation scope is based on components of the same cross-section or adjacent rings of a set number, extracting multi-ring collaborative deformation features. At the global tunnel scale, the aggregation scope is based on all component nodes of the entire tunnel, extracting the risk evolution features of the entire tunnel.

[0054] Please see Figure 3 By comparing the weights of the construction and operation phases, the adaptive capability of the hierarchical attention mechanism in this embodiment is explained. The corresponding technical effect is that the model can automatically adjust the contribution weights of local, mesoscopic, and global spatial features under different operating conditions, thereby maintaining its adaptability to deformation modes at different stages.

[0055] There is a correspondence between the three scales and the stress range of the tunnel structure. The local component scale corresponds to the stress influence range of a single component and its directly transmitted force neighborhood. Within this range, the deformation of one component is rapidly transmitted to adjacent components through the direct force transmission edge; therefore, fine-grained deformation features need to be extracted using directly adjacent components as the aggregation range. The mesoscopic cross-section scale corresponds to the collaborative stress range between the same cross-section or adjacent ring segments. Within this range, multi-ring components share the stress collaboratively through inter-ring joints and the overall cross-sectional stiffness; therefore, multi-ring collaborative deformation features need to be extracted using components of the same cross-section or adjacent rings of a predetermined number as the aggregation range. The global tunnel scale corresponds to the overall stress evolution range of the entire tunnel. Within this range, the deformation differences and risk evolution trends of different sections are interrelated; therefore, the overall risk evolution features need to be extracted using all component nodes of the entire tunnel as the aggregation range. In the specific implementation of the aggregation function, graph attention aggregation is used at the local component scale. Learnable attention weights are used to weight and sum the features of the central component and its directly adjacent components, giving higher aggregation weights to adjacent components with greater influence. At the mesoscopic cross-section scale, attention pooling or average pooling is used to aggregate the features of components within the same cross-section or adjacent loops of a set number into cross-section-level features. At the global tunnel scale, global pooling is used to aggregate the features of all component nodes in the entire tunnel into tunnel-wide features. The aggregation results at the three scales are used as spatial features at the local, mesoscopic, and global scales, respectively, for subsequent fusion.

[0056] The spatial features at the three scales are fused and then incorporated into a hierarchical attention mechanism to automatically learn the weight allocation of local component scale, mesoscopic cross-sectional scale, and global tunnel scale under different working conditions. The weights are adaptively adjusted according to the tunnel construction stage and operation status.

[0057] The hierarchical attention mechanism is implemented as follows: First, spatial features at three scales are linearly transformed and mapped to the same feature dimension. Then, the mapped features are input into an attention scoring function to obtain attention scores for each scale. The three scores are normalized using a softmax function to obtain weight coefficients between 0 and 1, summing to 1. Finally, the features at the three scales are weighted and summed according to these weight coefficients to obtain the fused spatial features. The weights are automatically learned during model training. During construction, the immediate impact of tunnel face excavation on surrounding components is significant, and features at the local component scale contribute more to the prediction results; therefore, the local scale weights are higher. During operation, long-term settlement of the entire tunnel and segmental differences become the main concerns, and features at the global tunnel scale contribute more to the prediction results; therefore, the global scale weights are increased. This adaptive adjustment of weights allows the model to maintain high prediction accuracy under different construction stages and operational conditions, without manual intervention, and avoids biases caused by experience differences when manually setting weights.

[0058] Temporal feature encoding employs a self-attention mechanism to encode state data across continuously monitored time points, capturing long-term dependencies over time. The self-attention mechanism automatically identifies historical moments that significantly influence the current prediction time by calculating attention weights between features at different points in the time series, thereby extracting effective temporal features.

[0059] The structural force transmission relationship, embodied by the strength of the connecting edge, is embedded as a physical constraint into the spatial feature aggregation process of the multi-scale spatiotemporal prediction model, so that the strength of the connecting edge varies within a set range near the initial value of engineering experience.

[0060] The physical constraints are implemented by adding a physical constraint regularization term to the loss function during model training. The strength of the constraint connection edges fluctuates within a set range near the initial empirical value. The formula for the physical constraint regularization term is: in, The loss value is the physical constraint regularization term. The regularization coefficient is... The strength of the connection edge between component i and component j is used as a learnable parameter during model training. The initial empirical values ​​for the strength of the connecting edges are 0.8 for direct force transmission edges and 0.3 for indirect force transmission edges. This regularization term adopts... The reason for the norm square form is that, Norm square with respect to It is continuously differentiable everywhere, which facilitates optimization using gradient descent. Furthermore, the squared term imposes a greater penalty on the strength of connecting edges that deviate significantly from the initial value, thus pulling large deviations back to near the initial value more quickly. The norm avoids sparse solutions, ensuring that the strength of connecting edges changes smoothly around the initial value instead of being compressed to zero, thus achieving a balance between data-driven fitting and mechanical priors. This regularization term prevents the strength of connecting edges from drifting against mechanical principles during model training, such as training the strength of directly transmitting edges to be lower than that of indirectly transmitting edges, thereby maintaining the model's learning process in accordance with the force transmission mechanism of the tunnel structure. The set range is determined based on the initial values ​​from engineering experience and the reliability of the force transmission relationship. In this embodiment, it is set to 20% above and below the initial values ​​from engineering experience, i.e., the strength of directly transmitting edges varies between 0.64 and 0.96, and the strength of indirectly transmitting edges varies between 0.24 and 0.36. The larger the value of , the closer the strength of the connection edge is to the initial value based on engineering experience, and the more the model is biased towards physical priors; the smaller the value, the more the strength of the connection edge depends on data-driven approaches, and the larger the adjustable range. Cross-validation can be performed on the training set to determine the validity of the data: Historical monitoring data is divided into a training set and a validation set, and different validation sets are used on the training set. The model is trained using the value that minimizes the prediction error on the validation set. Simultaneously, it checks whether the strength of the connected edges obtained during training satisfies the relative relationship that the strength of the directly transmitting force edge is greater than that of the indirectly transmitting force edge, to ensure that the model maintains physical rationality while achieving fitting accuracy. In this embodiment... Take 0.1.

[0061] After outputting the deformation prediction values, a component-level attribution analysis is performed on the multi-scale spatiotemporal prediction model to calculate the contribution of each input feature to the comprehensive risk index. The top-ranking risk contribution factors and their respective contribution percentages are output. The purpose of attribution analysis is to enhance the interpretability of the early warning results and assist engineers in understanding the causes of risks. For example, when the comprehensive risk index of a component exceeds the standard, attribution analysis can identify which factors among settlement, convergence, stress, or force transmission coefficient contribute the most to the risk index, and the contribution percentage of each factor, thus providing a basis for engineers to formulate remedial measures.

[0062] Step 5: In response to the deformation prediction value of any component exceeding the deformation threshold of that component, determine the risk impact intensity between the two components based on the spatial distance between any two components, the structural force transmission relationship reflected by the strength of the connection edge, and the historical risk propagation. Calculate the risk increment of the affected component based on the risk impact intensity, determine the comprehensive risk index by combining the deformation degree and deformation growth rate of the affected component, and determine the warning level based on the comprehensive risk index.

[0063] Please see Figure 4 , Figure 3 The diagram illustrates the risk contribution structure of an example component, demonstrating that the comprehensive risk index is not triggered by a single threshold, but is jointly determined by the degree of deformation, growth rate, risk increment, and historical instability. This diagram corresponds to the technical effects of "comprehensive risk index interpretability" and "attribution-assisted decision-making" in this embodiment, facilitating the identification of the most critical risk drivers.

[0064] The deformation threshold is the critical value used to determine whether a component has entered an abnormal state. The deformation threshold can be determined based on the statistical distribution of historical monitoring data, allowable deformation values ​​in design specifications, or empirical values ​​corresponding to the surrounding rock grade. When the predicted deformation value of a component exceeds its deformation threshold, the component is identified as an abnormal component, triggering risk propagation analysis.

[0065] Risk impact intensity represents the strength of risk transmission between any two components. The contribution of spatial distance, structural force transmission relationship reflected by the strength of the connecting edges, and historical risk propagation to risk impact intensity adaptively adjusts with changes in surrounding rock grade and construction stage.

[0066] The calculation of the risk impact intensity comprehensively considers three factors: spatial distance, structural force transmission relationship, and historical risk propagation. The calculation formula is as follows: in, The intensity of the risk impact of component i on component j; The spatial distance between component i and component j; The strength of the connection edge between component i and component j is determined by step 2; The historical risk propagation probability between component i and component j; , , These are the weighting coefficients corresponding to spatial distance, structural force transmission relationship, and historical risk propagation, respectively. Spatial distance term uses The reason for using the reciprocal form is that the intensity of risk impact should monotonically decrease with increasing spatial distance. Using this form ensures that the closer the distance, the greater the impact, and the farther the distance, the smaller the impact. Adding 1 to the denominator avoids division by zero when two components overlap spatially. Simultaneously, it keeps this term within a comparable numerical range to the connection edge strength and historical risk propagation probability. Connection edge strength The edge strength of the tunnel topology map established in step 2 is directly reused, ensuring strict consistency between risk propagation and structural stress transmission in the data. Historical risk propagation probability. The statistical method is as follows: Review historical monitoring records, count the number of times component j also experiences an anomaly within a set time period after component i experiences an anomaly, divide this number by the total number of times component i experiences an anomaly, and obtain the conditional frequency of the propagation of the anomaly from component i to component j, as... The initial value is set and continuously updated as monitoring data accumulates. When historical data is insufficient, If we set it to 0, the intensity of the risk impact is mainly determined by spatial distance and structural force transmission relationship. As the number of historical risk records increases during the closed-loop update process, The effect of [the factor] gradually increases. The sum of the weight coefficients of the three factors is 1. , , The initial values ​​can be determined through cross-validation: Abnormal events that have occurred in historical monitoring data are used as samples, and whether the abnormal event propagates along the path between two components is used as a label. The weight combination that maximizes the accuracy of risk propagation prediction is searched in the parameter space as the initial value. When the surrounding rock grade is poor, the self-stabilizing ability of the surrounding rock is weak, and the contribution of force transmission between components to risk propagation is more significant; therefore, the weight of structural force transmission relationship is increased. When construction disturbance is strong, the immediate impact range of tunneling at the face on adjacent components expands, and the contribution of spatial distance to risk propagation is more significant; therefore, the weight of spatial distance is increased. Historical risk propagation reflects the actual frequency and intensity of risk transmission between two components over a period of time. Its weight can be adjusted according to the sufficiency of historical data; the more complete the historical data, the higher the weight. The stronger the statistical significance, the higher the weight of historical risk propagation. The above adjustment rules can be preset with weight adjustment ranges according to the surrounding rock grade and construction stage, so that the weights can adaptively change within the preset range based on the initial values ​​obtained from cross-validation, thereby achieving a balance between data calibration and engineering experience.

[0067] The comprehensive risk index is determined by the degree of deformation of the affected component, the rate of deformation growth, the risk increment, and the degree of historical instability. The warning level is divided into four levels: safety level, attention level, warning level, and emergency level, based on the range of the comprehensive risk index.

[0068] When the predicted deformation value of component k exceeds its deformation threshold, component k is considered an anomaly source component, and its strength is affected by risk. The formula for calculating the risk increment of affected component j is: in, The risk increment for component j; The risk impact intensity of abnormal source component k on component j; Let k be the predicted deformation value of component k. Let k be the deformation threshold for component k. This formula means that the risk increment is related to the degree of exceedance by the abnormal source component. Proportional to, and simultaneously with, the intensity of the risk impact between the two components. The magnitude of the exceedance is directly proportional to the strength of the component; the greater the degree of exceedance and the stronger the impact, the greater the risk increment transmitted to component j. A relative exceedance value is adopted. Rather than absolutely over-limited The reason is that different components have different deformation threshold magnitudes. The relative exceedance can eliminate these differences, making the exceedance levels of different components comparable. This ensures that the risk increment can be uniformly compared and accumulated across different components. The formula for calculating the comprehensive risk index is: in, Let j be the comprehensive risk index of component j. Let j be the predicted deformation value of component j. Let be the deformation threshold of component j, and the ratio of the two values ​​reflects the current degree of deformation of the component. For component j in time interval The amount of deformation growth within, and The ratio represents the rate of deformation growth of the component; The risk increment generated by adjacent components on component j; Let j be the historical stability coefficient of component j; , , , For the weighting coefficients, satisfying Historical stability coefficient The value is determined by the degree of fluctuation in the historical monitoring data of component j, and the calculation formula is as follows: in, Let be the standard deviation of the monitoring data for component j over the historical period. A larger standard deviation indicates more severe fluctuations in the component's historical deformation and poorer stability. The closer it gets to 0, The closer the standard deviation is to 1, the greater its contribution to the overall risk index; the smaller the standard deviation, the more stable the historical deformation of the component. The closer it is to 1, The closer it gets to 0. (Adopt) The normalized form maps the values ​​to the (0,1] interval, making the historical stability comparable to other components, while avoiding division by zero when the standard deviation is 0. The reason for using a weighted summation instead of product or maximum value for the four components is that the degree of deformation, deformation growth rate, risk increment, and historical instability characterize different aspects of component risk. These four components are complementary rather than substitutive. Weighted summation reflects the comprehensive impact of each component simultaneously, and the relative importance of each component is adjusted through weighting coefficients, making it more intuitive and comprehensive than product or maximum value calculation. , , , The value of can be determined through historical data regression analysis: using the actual risk status of components in historical monitoring data as the target variable, and four sub-indicators as independent variables, logistic regression or ordered logistic regression is used to fit the relationship between the four sub-indicators and the risk status, and the normalized regression coefficients obtained from the fitting are used as . to The value of is determined by the following: when historical data is insufficient, expert experience or the analytic hierarchy process (AHP) is used to determine the relative importance of each item, and then normalized to a weighting coefficient. In this embodiment, the degree of deformation of the component itself is considered to be the primary direct representation of risk, the deformation growth rate reflects the speed of risk evolution, the risk increment reflects the degree of anomaly transmitted from adjacent components, and the degree of historical instability reflects the long-term accumulated stability defects. , , , Take values ​​of 0.4, 0.2, 0.2, and 0.2 respectively.

[0069] The warning level is determined based on the range of values ​​for the comprehensive risk index. A value less than 0.4 indicates a safety level, meaning the component is in a normal condition; A value greater than or equal to 0.4 and less than 0.6 indicates a level of alert, signifying that the component exhibits a certain degree of abnormal trend and requires enhanced monitoring. A value greater than or equal to 0.6 and less than 0.8 indicates a warning level, signifying that the component poses a significant safety risk and requires remedial measures. A value of 0.8 or higher indicates an emergency level, signifying a serious safety threat to the component and requiring immediate emergency response measures. The above-mentioned thresholds are determined as follows: when all four sub-indicators are at normal levels, the comprehensive risk index is mainly concentrated below 0.2, and below 0.4 corresponds to a basically normal state for the component; when the deformation degree of the component approaches the threshold... When it approaches 1, the degree of deformation alone is sufficient to make When the deformation reaches a level of 0.4, the component should already be considered a cause for concern; therefore, 0.4 is used as the starting point for the attention level. When the deformation degree further increases and the deformation growth rate and risk increment increase simultaneously, the weighted summation result rapidly rises to above 0.6, corresponding to a state requiring intervention; therefore, 0.6 is used as the starting point for the warning level. When multiple indicators simultaneously approach or exceed their respective critical levels, the comprehensive risk index rises to above 0.8, corresponding to a serious safety threat; therefore, 0.8 is used as the starting point for the emergency level. The classification thresholds can be adjusted online according to the safety level requirements of the engineering project. For example, for high-risk tunnels, the overall classification threshold can be lowered to improve early warning sensitivity.

[0070] The risk propagation results are visualized by overlaying a heat map onto the BIM 3D scene. In the heat map, different colors represent different risk levels, allowing engineers to intuitively identify hazardous areas and their spread trends.

[0071] Step 6: Obtain subsequent actual monitoring values ​​and compare them with the deformation prediction values ​​of the corresponding components to obtain the prediction error. In response to the prediction error continuously exceeding the preset threshold, update the local prediction parameters and risk propagation parameters of the abnormal components and adjacent components to form a monitoring closed loop.

[0072] The formula for calculating prediction error is: in, Let be the mean absolute prediction error at time t. This represents the total number of tunnel components. For monitoring indicator dimensions, Let be the predicted deformation value of the i-th component at time t for the k-th monitoring index. Let be the actual monitoring value of the i-th component at time t for the k-th monitoring indicator. The reason for using mean absolute error (MAE) as the prediction error metric is that individual sudden jumps are inevitable in the monitoring data. Squaring the mean squared error amplifies the impact of these individual large errors on the overall error metric, causing a few sudden jumps to dominate error judgment and mask the overall prediction level. MAE, on the other hand, treats each error with equal weight, is insensitive to individual large errors, and can more robustly reflect the model's overall prediction level across components and monitoring indicators, thus providing a more reliable criterion for whether to trigger a local update. Preset threshold. It is the critical value for judging whether the prediction error is abnormal. The prediction error can be determined by the statistical distribution of prediction errors under historical normal operating conditions: before the model is put into online operation, the mean and standard deviation of the prediction error are statistically analyzed using historical monitoring data under normal operating conditions, and the mean is compared with... The sum of multiple standard deviations is used as the preset threshold. , Typically, a value of 2 to 3 is chosen to ensure that the prediction error under normal operating conditions does not exceed this threshold in most cases. This ensures a low false alarm rate while maintaining sufficient sensitivity to errors exceeding the normal range. In this embodiment... Choose 3. When the working conditions change, the preset threshold can be corrected online to adapt to changes in surrounding rock conditions and construction stages.

[0073] Please see Figure 5 , Figure 5 By comparing error trends, the improvement effect of incremental updates to local subgraphs on model accuracy is illustrated. This figure corresponds to the closed-loop mechanism of "triggering local updates when the prediction error continuously exceeds the threshold" in this embodiment, demonstrating that this mechanism can bring more stable and lower prediction errors, reflecting the engineering value of the self-learning mechanism.

[0074] When the prediction error obtained by comparing the subsequent actual monitoring value with the deformation prediction value of the corresponding component continuously exceeds the preset threshold, the local subgraph incremental update is triggered. Only the local prediction parameters corresponding to the local subgraph composed of the abnormal component and the components within the adjacent set ring number range are updated, and the local prediction parameters corresponding to the components outside the update range are not updated. The local prediction parameters include the parameters used in spatial feature aggregation and temporal feature encoding for the abnormal component and the components within the adjacent set ring number range, and the risk propagation parameters between the abnormal component and the adjacent components are updated.

[0075] The reason for using partial updates instead of full updates is that full updates are computationally expensive and may introduce noise from irrelevant components. When the prediction error of a component consistently exceeds a threshold, the error source is usually concentrated on the local stress state changes of that component and its neighboring components, rather than the systemic changes of all components in the entire tunnel. Updating only the local prediction parameters of abnormal components and components within a set ring number range can reduce computational costs while ensuring model adaptability and avoiding noise interference from irrelevant components.

[0076] The determination of the number of adjacent rings is based on the principle that, centered on the abnormal component, components within a 3-ring radius are most significantly affected by the deformation of the abnormal component, while the impact on components outside this range decreases rapidly. In this embodiment, the number of adjacent rings is 3. In other embodiments, it can be adjusted to 2 or 5 rings depending on the surrounding rock grade and construction method, as long as the main influence range of the abnormal component is covered.

[0077] The loss function for local updates is: in, The loss function value is for incremental learning of the local subgraph. This refers to the batch size in mini-batch gradient descent. To predict the number of steps during the time period, Let be the predicted deformation value for the b-th batch at time t. This represents the actual monitoring value for the b-th batch at time t. The regularization coefficient is . These are the parameters of the local subplot to be updated. The first term of the loss function is the sum of squared errors between the predicted and actual values, used to measure the fitting accuracy; the second term is the regularization term, using parameters... The squared norm constrains the magnitude of the parameters to be updated, preventing overfitting when local data is limited. Regularization coefficient. Cross-validation can be used to determine the optimal training set based on different... During training, values ​​that minimize the prediction error are selected on the validation set. Values, in this embodiment Take 0.001. Batch size. The number of samples within the local subgraph is determined by selecting a subset of the available sample count, such as 32 or 64, to balance update efficiency and stability. While updating the local prediction parameters, the risk propagation parameters between anomalous components and adjacent components are also corrected. Historical risk propagation probability. The corrected formula is: in, This is the corrected historical risk propagation probability. The historical risk propagation probability before correction. For learning rate, The prediction error at the current moment. The preset threshold is used. This correction formula means that when the prediction error exceeds the preset threshold, the ratio of error to threshold is greater than 1, and the historical risk propagation probability is adjusted upwards proportionally, with a larger increase for larger errors; when the prediction error is below the preset threshold, the ratio of error to threshold is less than 1, and the historical risk propagation probability is adjusted downwards proportionally, thus dynamically adjusting according to the actual risk propagation situation. The reason for using multiplicative correction instead of additive correction is that the historical risk propagation probability itself is a probability value between 0 and 1, and multiplicative correction ensures that the corrected value remains between 0 and 1, preventing it from exceeding the limit due to addition or subtraction operations. Learning rate. The magnitude of each correction is controlled by taking a small positive value; in this embodiment, it is 0.05. This ensures a smooth correction process, preventing large fluctuations in the historical risk propagation probability caused by a single error. Through multiple closed-loop corrections, the model gradually approaches the true risk propagation probability. The learning rate can decay as the model deployment time increases, allowing for rapid correction of the historical risk propagation probability in the initial stage and stabilization after long-term operation.

[0078] Through steps 1 to 6 above, a closed-loop mechanism of prediction, feedback, correction, and updating is formed. After outputting the prediction results, the model continuously receives real-world monitoring values ​​from the field. When the prediction error continues to exceed a preset threshold, only the local sub-map within the abnormal component and its adjacent range is incrementally updated, and the risk propagation parameters are corrected simultaneously. This allows the model to adapt to monitoring changes under different surrounding rock grades, different construction stages, and different environmental conditions, avoiding the problem of model failure after a single training session.

[0079] Please see Figure 7 The present invention also provides a tunnel structure safety monitoring system based on multi-source information fusion, the system being used to execute the above-mentioned tunnel structure safety monitoring method based on multi-source information fusion, comprising: The component division module is used to obtain the BIM model of the tunnel to be monitored, divide the tunnel into multiple components according to the stress unit, and extract the spatial position and connection relationship of each component. The topology construction module determines the strength of the connecting edges based on the force transfer relationship between components and generates a tunnel topology map for calculation and analysis. The connecting edges connect two components that have a force transfer relationship. The status data mapping module is used to obtain the monitoring data corresponding to each component, map the monitoring data to the corresponding component nodes in the tunnel topology diagram, and form the status data of each component node. The multi-scale spatiotemporal prediction module inputs the state data of each component node into the multi-scale spatiotemporal prediction model, aggregates spatial features in the adjacent range of the component, the cross-sectional range and the overall range of the tunnel, encodes the temporal features, and merges the spatial and temporal features of the three ranges to output the deformation prediction value of each component in the next prediction period. The risk propagation early warning module responds when the predicted deformation value of any component exceeds the deformation threshold of that component. Based on the spatial distance between any two components, the structural force transmission relationship reflected by the strength of the connecting edge, and the historical risk propagation situation, it determines the risk impact intensity between the two components, calculates the risk increment of the affected component based on the risk impact intensity, determines the comprehensive risk index by combining the deformation degree and deformation growth rate of the affected component, and determines the early warning level based on the comprehensive risk index. The closed-loop update module is used to obtain subsequent real monitoring values, compare them with the deformation prediction values ​​of the corresponding components to obtain the prediction error, and update the local prediction parameters and risk propagation parameters of the abnormal components and adjacent components in response to the prediction error continuously exceeding the preset threshold, thus forming a monitoring closed loop.

[0080] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0081] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0082] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0083] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for monitoring the safety of tunnel structures based on multi-source information fusion, characterized in that, The specific steps include: Obtain the BIM model of the tunnel to be monitored, divide the tunnel into multiple components according to the stress unit, and extract the spatial position and connection relationship of each component; The strength of the connecting edge is determined based on the force transfer relationship between the components, and a tunnel topology map is generated for calculation and analysis. The connecting edge connects two components that have a force transfer relationship. Acquire the monitoring data corresponding to each component, map the monitoring data to the corresponding component nodes in the tunnel topology diagram, and form the status data of each component node; The state data of each component node is input into the multi-scale spatiotemporal prediction model. Spatial features are aggregated in the adjacent range of the component, the cross-sectional range and the overall range of the tunnel, and the temporal features are encoded. The spatial and temporal features of the three ranges are fused to output the deformation prediction value of each component in the next prediction period. In response to the predicted deformation value of any component exceeding the deformation threshold of that component, the risk impact intensity between two components is determined based on the spatial distance between any two components, the structural force transmission relationship reflected by the strength of the connection edge, and the historical risk propagation. The risk increment of the affected component is calculated based on the risk impact intensity. The comprehensive risk index is determined by combining the deformation degree and deformation growth rate of the affected component, and the warning level is determined based on the comprehensive risk index. The subsequent actual monitoring values ​​are obtained and compared with the deformation prediction values ​​of the corresponding components to obtain the prediction error. In response to the prediction error continuously exceeding the preset threshold, the local prediction parameters and risk propagation parameters of the abnormal components and adjacent components are updated to form a monitoring closed loop.

2. The tunnel structure safety monitoring method based on multi-source information fusion according to claim 1, characterized in that: Connection edges include three categories: direct force transmission edges, indirect force transmission edges, and construction disturbance edges; The initial strength of the directly transmitted force edge is greater than that of the indirectly transmitted force edge; the initial strength of the construction disturbance edge is determined based on the time difference of the construction process.

3. The tunnel structure safety monitoring method based on multi-source information fusion according to claim 2, characterized in that: The direct force transmission edge is established by the connection relationship between adjacent components in the same ring, the connection relationship between corresponding components in the front and rear rings, and the connection relationship between the lining components and the invert arch components; The indirect force transmission edge is established by the connection relationship between the initial support and the secondary lining, as well as the connection relationship between the transverse passage and the main tunnel connecting components; The construction disturbance edge is established based on the current construction stage and construction progress.

4. The tunnel structure safety monitoring method based on multi-source information fusion according to claim 1, characterized in that: Spatial feature aggregation includes three scales: local component scale, mesoscopic cross-sectional scale, and global tunnel scale. The local component scale is aggregated with directly adjacent components, and fine-grained deformation features of single components and directly adjacent components are extracted; At the mesoscopic cross-sectional scale, components with the same cross-section or adjacent predetermined number of rings are used as the aggregation range to extract multi-ring collaborative deformation features; The global tunnel scale uses all component nodes of the entire tunnel as the aggregation range to extract the risk evolution characteristics of the entire tunnel; The structural force transmission relationship embodied by the strength of the connecting edge is embedded as a physical constraint into the spatial feature aggregation process of the multi-scale spatiotemporal prediction model, so that the strength of the connecting edge varies within a set range near the initial value of engineering experience. After outputting the deformation prediction values, component-level attribution analysis is performed on the multi-scale spatiotemporal prediction model to calculate the contribution of each input feature to the comprehensive risk index, and output the risk contribution factors with the highest contribution ranking and their contribution percentage.

5. The tunnel structure safety monitoring method based on multi-source information fusion according to claim 4, characterized in that: The spatial features at the three scales are fused and then incorporated into a hierarchical attention mechanism to automatically learn the weight allocation of local component scale, mesoscopic cross-sectional scale, and global tunnel scale under different working conditions. The weights are adaptively adjusted according to the tunnel construction stage and operation status.

6. The tunnel structure safety monitoring method based on multi-source information fusion according to claim 5, characterized in that: The contribution of spatial distance, structural force transmission relationship reflected by the strength of the connecting edge, and historical risk propagation to the intensity of risk impact adaptively adjusts with changes in the surrounding rock grade and construction stage.

7. A method for monitoring the safety of tunnel structures based on multi-source information fusion according to claim 6, characterized in that: The comprehensive risk index is determined by the degree of deformation of the affected component itself, the rate of deformation growth, the risk increment, and the degree of historical instability. The warning levels are divided into four levels based on the range of values ​​of the comprehensive risk index: safety level, attention level, warning level, and emergency level.

8. A method for monitoring the safety of tunnel structures based on multi-source information fusion according to claim 7, characterized in that: When the prediction error obtained by comparing the subsequent actual monitoring value with the deformation prediction value of the corresponding component continuously exceeds the preset threshold, the local subgraph incremental update is triggered. Only the local prediction parameters corresponding to the local subgraph composed of abnormal components and components within the adjacent set ring number range are updated, and the local prediction parameters corresponding to components outside the update range are not updated. The local prediction parameters include the parameters used in spatial feature aggregation and temporal feature encoding for anomalous components and components within the adjacent set ring number range, and update the risk propagation parameters between anomalous components and adjacent components.

9. A method for monitoring the safety of tunnel structures based on multi-source information fusion according to claim 8, characterized in that: As tunnel construction progresses, for each set number of construction progresses completed, corresponding component nodes are automatically added, and connection edges are completed based on the force transfer relationship between the new components and existing components, so that the total number of component nodes in the tunnel topology diagram increases with the tunnel construction progress.

10. A tunnel structure safety monitoring system based on multi-source information fusion, characterized in that, The system is used to execute the tunnel structure safety monitoring method based on multi-source information fusion as described in any one of claims 1-9, including: The component division module is used to obtain the BIM model of the tunnel to be monitored, divide the tunnel into multiple components according to the stress unit, and extract the spatial position and connection relationship of each component. The topology graph construction module determines the strength of the connecting edges based on the force transfer relationship between components and generates a tunnel topology graph for calculation and analysis. The connecting edges connect two components that have a force transfer relationship. The status data mapping module is used to obtain the monitoring data corresponding to each component, map the monitoring data to the corresponding component nodes in the tunnel topology diagram, and form the status data of each component node. The multi-scale spatiotemporal prediction module inputs the state data of each component node into the multi-scale spatiotemporal prediction model, aggregates spatial features in the adjacent range of the component, the cross-sectional range and the overall range of the tunnel, encodes the temporal features, and merges the spatial and temporal features of the three ranges to output the deformation prediction value of each component in the next prediction period. The risk propagation early warning module responds when the predicted deformation value of any component exceeds the deformation threshold of that component. Based on the spatial distance between any two components, the structural force transmission relationship reflected by the strength of the connecting edge, and the historical risk propagation situation, it determines the risk impact intensity between the two components, calculates the risk increment of the affected component based on the risk impact intensity, determines the comprehensive risk index by combining the deformation degree and deformation growth rate of the affected component, and determines the early warning level based on the comprehensive risk index. The closed-loop update module is used to obtain subsequent real monitoring values, compare them with the deformation prediction values ​​of the corresponding components to obtain the prediction error, and update the local prediction parameters and risk propagation parameters of the abnormal components and adjacent components in response to the prediction error continuously exceeding the preset threshold, thus forming a monitoring closed loop.

Images