A method for monitoring road slope data under complex geological conditions

Abstract

The present invention provides a method for monitoring road slope data under complex geological conditions, which relates to the technical field of road slope monitoring. The method includes: performing temperature compensation on the change rate, calculating the deformation amount caused by temperature change by analyzing the thermal expansion characteristics of materials, and correcting the monitored side lengths to generate temperature-compensated side lengths; according to the temperature-compensated side lengths and the surrounding rock rheological equation, correcting two adjacent side lengths through the strain difference between adjacent sides to obtain the final side lengths; combining the final side lengths with the reference interior angles to generate a topological structure; based on the topological structure, constructing a three-dimensional deformation index matrix including the diagonal displacement difference, the adjacent side curvature ratio, and the interior angle coefficient of variation; inputting the index matrix, the support structure strain spatio-temporal matrix, and the rock mass fragmentation tensor into a spatio-temporal attention graph neural network, and outputting the deformation evolution trend including the diagonal displacement difference, the adjacent side curvature ratio, and the interior angle coefficient of variation. The present invention can achieve high-precision real-time monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of road slope monitoring, and particularly to a method for monitoring road slope data under complex geological conditions. Background Art

[0002] In the construction of road projects in complex geological regions, slope stability monitoring is the core technical problem to ensure project safety. Especially in rock mass fracture zones, high in-situ stress soft rocks, and regions with significant temperature alternation, some traditional monitoring methods face the following technical bottlenecks:

[0003] For example, existing monitoring technologies rely on the layout mode of discrete sensors (such as single-point displacement gauges and inclinometers), which is essentially a data acquisition logic based on "point-like observations".

[0004] Such methods have the following technical limitations:

[0005] Single-point sensors can only reflect the local deformation state of their installation positions and cannot construct a continuous displacement field distribution model on the slope surface. For example, in a rock mass fracture zone, potential slip surfaces often exhibit heterogeneous expansion characteristics, and the deformation gradient information between discrete data points is missing, resulting in the inability to accurately reconstruct the spatial shape of the slip surface through interpolation algorithms.

[0006] For example, in progressive failure, dynamic characteristics such as the phase difference and direction synergy of displacement rates at different positions are key indicators for judging the penetration of the slip surface. However, due to the lack of real-time cross-correlation calculation of multi-node data in existing technologies, it is difficult to capture such co-deformation laws; the sparse layout of discrete sensors is likely to cause monitoring blind spots. Especially at the intersection of structural planes or the interface of soft interlayers, local strain concentration may appear prior to macroscopic displacement, and traditional methods are not easy to detect such precursors of hidden damage due to the lack of distributed strain sensing capabilities. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method for monitoring road slope data under complex geological conditions, which can achieve high-precision real-time monitoring.

[0008] To solve the above technical problem, the technical solution of the present invention is as follows:

[0009] A method for monitoring road slope data under complex geological conditions, the method comprising:

[0010] Measuring the lengths and interior angles of each side of a quadrilateral in real time to generate a reference topology;

[0011] Based on the benchmark topology and the real-time distances between the vertices of the quadrilateral and the tunnel face, calculate the weight of each vertex; when the weight of a vertex is greater than the threshold, monitor the change rate of the corresponding side in real time; perform temperature compensation on the change rate, analyze the thermal expansion characteristics of the material, calculate the deformation caused by temperature change, and correct the monitored side length to generate the temperature-compensated side length; according to the temperature-compensated side length and the surrounding rock rheological equation, correct the two adjacent side lengths by the strain difference between adjacent sides to obtain the final side length; combine the final side length with the benchmark interior angles to generate the topological structure.

[0012] Based on the topological structure, construct a three-dimensional deformation index matrix including the diagonal displacement difference, the adjacent side curvature ratio, and the interior angle coefficient of variation; input the index matrix, the support structure strain spatio-temporal matrix, and the rock mass fragmentation tensor into the spatio-temporal attention graph neural network, and output the deformation evolution trend including the diagonal displacement difference, the adjacent side curvature ratio, and the interior angle coefficient of variation.

[0013] Execute multi-level early warning decisions according to the component parameters of the deformation evolution trend.

[0014] The above solution of the present invention has at least the following beneficial effects:

[0015] By constructing the benchmark topology through real-time measurement of the side lengths and interior angles of the quadrilateral unit, it breaks through the discreteness limitation of traditional single-point monitoring and realizes the systematic description of the "areal deformation" of the slope. For example, when the joint surface inside the slope undergoes dislocation, the quadrilateral vertex weight calculation mechanism can automatically identify the key deformation areas near the tunnel face (such as the vertices of the free face triggered by the weight threshold), and track the length change rate of the corresponding side in real time, which can improve the ability to identify the local landslide risk of the slope.

[0016] Based on the thermal expansion characteristics of the material, establish a temperature-deformation coupling model, calculate the false deformation caused by temperature change in real time, eliminate the interference of environmental factors on the monitoring data, and improve the side length monitoring accuracy from ±2mm to ±0.5mm; combined with the Burgers surrounding rock rheological equation, use the strain difference between adjacent sides to dynamically adjust the influence of long-term creep, and solve the defect that the traditional method does not consider the time-dependent deformation of the rock mass. For example, in the shale-sandstone interlayer slope, it can accurately capture the progressive influence of the decrease in cohesion caused by groundwater softening during the rainy season on the slope deformation.

[0017] Construct a topological deformation index including the diagonal displacement difference (resolution 0.1mm), the adjacent side curvature ratio (accuracy 0.01), and the interior angle coefficient of variation (sensitivity 0.5°), and combine multi-source parameters such as the support structure strain (such as the change rate of anchor cable tension 1%) and the rock mass fragmentation tensor (the resolution of the volume ratio of fragmented blocks 5%) to form a full-element monitoring matrix.

[0018] Component parameters based on the deformation evolution trend (such as the displacement difference rate of 0.5 mm / h and the curvature ratio growth rate of 10% / h) can achieve multi-level responses from yellow warning (local deformation) to red warning (overall instability), reducing the false alarm rate and missed alarm rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic flow chart of a method for monitoring road slope data under complex geological conditions provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] Hereinafter, exemplary embodiments of the present disclosure will be described in more detail with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully conveyed to those skilled in the art.

[0021] As Figure 1 shown, an embodiment of the present invention provides a method for monitoring road slope data under complex geological conditions, and the method includes the following steps: <{

[0022] Measure the lengths and interior angles of each side of the quadrilateral in real time to generate a reference topology;

[0023] Based on the reference topology and the real-time distances between each vertex of the quadrilateral and the tunnel face, calculate the weight of each vertex; when the weight of a vertex is greater than the threshold, monitor the change rate of the corresponding side in real time; perform temperature compensation on the change rate, analyze the thermal expansion characteristics of the material, calculate the deformation amount caused by temperature change, and correct the monitored side length to generate temperature-compensated side lengths; according to the temperature-compensated side lengths and the surrounding rock rheological equation, correct the lengths of two adjacent sides through the strain difference between adjacent sides to obtain the final side lengths; combine the final side lengths with the reference interior angles to generate a topological structure;

[0024] Based on the topological structure, construct a three-dimensional deformation index matrix including diagonal displacement difference, adjacent side curvature ratio, and interior angle coefficient of variation; input the index matrix, the strain spatio-temporal matrix of the support structure, and the rock mass fragmentation tensor into the spatio-temporal attention graph neural network, and output the deformation evolution trend including diagonal displacement difference, adjacent side curvature ratio, and interior angle coefficient of variation;

[0025] Execute a multi-level early warning decision according to the component parameters of the deformation evolution trend.

[0026] In the embodiments of the present invention, a reference topology is constructed by measuring the side lengths and interior angles of quadrilateral units in real time, breaking through the discreteness limitation of traditional single-point monitoring and realizing the systematic description of the "surface deformation" of slopes. For example, when the joint plane inside the slope undergoes dislocation, the quadrilateral vertex weight calculation mechanism can automatically identify the key deformation areas near the tunnel face (such as the free face vertices triggered by the weight threshold), and track the length change rate of the corresponding sides in real time, which can improve the ability to identify the local landslide risk of the slope. A temperature-deformation coupling model is established based on the thermal expansion characteristics of materials to calculate the false deformation caused by temperature changes in real time, eliminating the interference of environmental factors on the monitoring data and improving the side length monitoring accuracy from ±2 mm to ±0.5 mm; combined with the Burgers surrounding rock rheological equation, the long-term creep effect is dynamically adjusted using the strain difference between adjacent sides to solve the defect that traditional methods do not consider the time-dependent deformation of rock masses. For example, in a mudstone-sandstone interlayer slope, the progressive influence of the cohesion reduction caused by groundwater softening during the rainy season on the slope deformation can be accurately captured. Topological deformation indexes including diagonal displacement difference (resolution 0.1 mm), adjacent side curvature ratio (accuracy 0.01), and interior angle coefficient of variation (sensitivity 0.5°) are constructed, and multi-source parameters such as the strain of the support structure (such as the change rate of anchor cable tension 1%) and the rock mass fragmentation tensor (resolution of the volume ratio of fragmented bodies 5%) are combined to form a full-element monitoring matrix. Based on the component parameters of the deformation evolution trend (such as displacement difference rate 0.5 mm / h, curvature ratio growth rate 10% / h), multi-level responses from yellow warning (local deformation) to red warning (overall instability) can be achieved, reducing the false alarm rate and missed alarm rate.

[0027] In a preferred embodiment of the present invention, the reference topology is as follows:

[0028] At selected positions in the tunnel section, the crown settlement point P1, the left convergence point P2, the right convergence point P3, and the invert uplift point P4 are arranged. The initial lengths of the sides of the quadrilateral and the interior angles between adjacent sides are measured in real time by a three-dimensional laser rangefinder, and an initial quadrilateral geometric parameter database is established;

[0029] Through the biaxial inclinometer array between P2 and P3, the change amount of the inclination angle of the side walls on both sides is collected, and the three-dimensional section scanner is synchronously triggered to perform point cloud scanning on the section where the quadrilateral is located. The scanning data is spatially registered with the laser ranging result to obtain calibration data;

[0030] According to the calibration data, the lengths of the sides and the interior angles of the initial quadrilateral are dynamically corrected to generate a reference topology including the spatial coordinates of the side lines, the interior angle values, and the section curvature.

[0031] In the embodiments of the present invention, the acquisition of the initial geometric parameters, the layout of the key points, and the collection of the basic data are realized as follows:

[0032] Install prisms or reflective targets at four characteristic positions on the vault (P1), left sidewall (P2), right sidewall (P3), and invert (P4) of the tunnel cross-section to form a quadrilateral monitoring unit (P1 - P2 - P3 - P4). These four points cover the top, both sides, and the bottom of the tunnel cross-section, constituting a closed geometric figure that can reflect the overall deformation of the cross-section.

[0033] Use a three-dimensional laser rangefinder (accuracy ±0.3 mm) to perform initial distance measurements on the 6 sides of the quadrilateral (P1P2, P1P3, P1P4, P2P3, P2P4, P3P4), and record the interior angles of adjacent sides (such as ∠P1P2P3, ∠P2P3P4, etc.) through an angle measurement module (accuracy ±5″) to establish an initial geometric parameter database containing side lengths (L1 - L6) and interior angles (θ1 - θ4).

[0034] Select a quadrilateral instead of a triangle because a quadrilateral has a higher degree of freedom in deformation sensitivity (a triangle is a rigid structure and it is difficult to capture in-plane shear deformation), and the four vertices cover the main deformation-sensitive areas of the tunnel cross-section (vault settlement, sidewall convergence, invert heave).

[0035] The four vertices correspond to the "top - side - bottom" characteristic positions of the tunnel cross-section. The constructed quadrilateral can synchronously monitor multi-mode deformations such as settlement, convergence, and heave, avoiding the one-sidedness of single-point monitoring. The laser ranging and angle measurement technologies provide initial geometric parameters with millimeter-level accuracy.

[0036] The fusion of inclination monitoring and point cloud scanning is achieved as follows:

[0037] Install biaxial inclinometers (accuracy ±0.01°) at intervals of 0.5 - 1 m along the longitudinal direction of the tunnel at the sidewall position between the left convergence point P2 and the right convergence point P3 to form an inclination monitoring array, and real-time collect the change amounts (△α, △β) of the lateral inclination angle (reflecting horizontal convergence) and the longitudinal inclination angle (reflecting twist along the tunnel axis) of the sidewall.

[0038] Point cloud scanning and spatial registration:

[0039] When the change amount of the inclination exceeds the threshold (such as △α > 0.05° or △β > 0.03°), trigger a three-dimensional cross-section scanner (resolution 0.5 mm) installed on the tunnel sidewall to collect point cloud data of the cross-section where the quadrilateral is located, and obtain thousands of dense three-dimensional coordinate points (x, y, z).

[0040] Perform spatial registration of the point cloud data and the prism coordinates of the laser ranging through the iterative closest point algorithm (ICP): Use the four vertices P1 - P4 as reference control points to adjust the point cloud coordinate system to align the scanned point cloud with the geometric framework of the laser ranging, eliminate the installation error of the device and the coordinate system deviation, and generate calibrated three-dimensional coordinate data.

[0041] The inclinometer captures minute inclination deformations of the side walls in real time (e.g., a 0.01° inclination corresponds to a 1.7 mm displacement over a 10 m distance), triggering high-precision scanning and avoiding the energy consumption and data redundancy of continuous full-time scanning; the point cloud scanning provides high-density cross-section contour data (such as the convexity and concavity of the side walls and local settlement of the vault), making up for the deficiency of laser ranging that only monitors discrete points. The integration of the two improves the geometric parameter calibration accuracy by 40% (e.g., the curve fitting error of the side wall is reduced from ±2 mm to ±1.2 mm). Triggering scanning based on the threshold of the inclination change amount enables on-demand monitoring, while ensuring data integrity, reducing the data acquisition frequency by 60% (the conventional scanning frequency is 1 time per hour, and the triggered scanning is on average 3 - 5 times per day).

[0042] Using the calibrated point cloud data, the spatial straight lines of each side of the quadrilateral are fitted by the least squares method (such as the three-dimensional straight line equation of the P1P2 side), and the actual side lengths are calculated (considering the influence of the tunnel curved surface, rather than the horizontal projected distance), correcting the initial length of the laser ranging. i ,y i ,z i ), and the actual interior angles of adjacent sides are calculated by vector cross product, eliminating the angular deviation during the initial measurement (such as the angular error caused by the instrument centering error).

[0043] Perform cross-section contour fitting (such as cubic spline curve) on the calibrated point cloud data, and calculate the local curvature radii of the vault (near P1), side walls (near P2 / P3), and invert arch (near P4) (R = 1 / k, where k is the curvature), reflecting the change in the cross-section shape (such as the increase in curvature caused by the settlement of the vault).

[0044] Integrate the corrected side lengths (L i ’), interior angles (θ i ’), three-dimensional vertex coordinates (x i ,y i ,z i ) and cross-section curvature (R1 - R3) into a reference topological database, with each parameter attached with a timestamp (accurate to seconds), supporting subsequent spatio-temporal analysis of deformation.

[0045] Upgrading from two-dimensional plane geometry (initial measurement) to three-dimensional space topology, considering the true curved surface form of the tunnel cross-section (such as the curvature characteristics of horseshoe-shaped and circular cross-sections), avoiding misjudgment of deformation caused by plane assumptions (such as the difference between the length of the inclined side in three-dimensional space and the horizontal projection can reach 3 - 5 mm / m). Introducing cross-section curvature as an additional parameter can sensitively identify local buckling deformation (such as sudden change in curvature caused by the concavity of the side wall), forming a complement with side lengths and interior angles:

[0046] The change in side length reflects "tension / compression", the change in interior angle reflects "shearing dislocation", and the change in curvature reflects "bending instability". The combination of the three can comprehensively describe the deformation mode of the cross-section (for example, the extrusion-type deformation is manifested as an increase in curvature + a decrease in interior angle, and the shear-type deformation is manifested as an increase in the length difference between opposite sides). After each calibration, a new reference topology is generated, automatically covering the old data, adapting to the stress adjustment of the surrounding rock during the tunnel construction stage (such as the convergence deformation stabilization period after the initial support), and ensuring the real-time matching of the reference parameters with the current surrounding rock state.

[0047] In a preferred embodiment of the present invention, based on the reference topology and the real-time distances between the vertices of the quadrilateral and the tunnel face, the weight of each vertex is calculated, including:

[0048] Using a three-dimensional laser rangefinder to continuously measure the three-dimensional Euclidean distances between the vertices P1, P2, P3, and P4 of the quadrilateral and the tunnel face at a sampling frequency of 5 times per second, generating the original dynamic distance sequences of each vertex; performing a sliding window average filter on the original sequences, with a window size of 10 sampling points, to generate the smoothed dynamic distance time-series data, including:

[0049] Install a three-dimensional laser ranging base station at the center position of the tunnel face (tunnel excavation face), with a built-in high-precision inertial navigation system (positioning accuracy ±2mm), emit laser beams to the four vertices (P1 - P4) of the quadrilateral, and calculate the three-dimensional Euclidean distances between the vertices and the tunnel face through the TOF (time of flight) principle.

[0050] Continuously collect data at a sampling frequency of 5Hz, generating an original distance sequence containing timestamps (for example, the distance value of point P1 is recorded every 0.2 seconds, forming the original distance sequence).

[0051] Use a 10-point sliding average window (time span of 2 seconds) to filter the original sequence: for each moment, calculate the average value from the previous 5 points to the current point, and eliminate sudden noises (such as abnormal jump values caused by equipment vibration and dust reflection), generating the smoothed distance time-series data.

[0052] The 5Hz sampling rate can capture the rapid changes in the vertex distances during the advancement of the tunnel face (such as sudden changes in distance during blasting excavation). Compared with the traditional 1Hz sampling, the response speed to dynamic deformation is increased by 4 times; the sliding window filtering effectively suppresses high-frequency noises (such as ±1mm random fluctuations caused by environmental vibration), reducing the standard deviation of the distance data from ±0.8mm to ±0.3mm. The distance data of all vertices are synchronously collected and filtered to ensure the alignment of timestamps (error < 10ms), avoiding weight calculation deviations caused by asynchronous measurements.

[0053] Based on the attenuation characteristics of rock mass stress wave propagation, establish a negative exponential mapping relationship between vertex weight and real-time distance, and fit the stress attenuation curve to determine the weight function parameters; input the smoothed dynamic distance time series data into the weight function parameters to calculate the vertex weight values in real time, including:

[0054] Through on-site tests (such as burying stress sensors under similar surrounding rock conditions), measure the attenuation law of stress wave with distance caused by face blasting or excavation, and find that the stress amplitude and distance conform to the negative exponential attenuation relationship:

[0055] , where is the initial stress at the face, is the attenuation coefficient, which is determined by fitting the on-site data through the least squares method. For example, in mudstone , in sandstone .

[0056] Based on the principle that "the closer to the face, the greater the excavation disturbance on the vertex and the higher the weight should be", define the vertex weight as being positively correlated with the stress amplitude, that is, construct a negative exponential weight function:

[0057] , where and are normalization constants, and the range is [0, 1]. For example, when then , then .

[0058] In tunnel sections with different surrounding rock grades (such as grade I to grade V), by comparing the correlation between vertex deformation and weight value (such as using the Pearson correlation coefficient), optimize the attenuation coefficient and the normalization constants (A, B). For example, set a larger k value in fractured surrounding rock (grade V) to make the weight of nearby vertices increase faster and highlight its sensitivity.

[0059] The weight function directly maps the propagation law of rock mass excavation disturbance, solves the problem that the traditional weight setting (such as empirical assignment) lacks a theoretical basis, and makes the weight value positively correlated with the actual deformation risk (for example, the weight of the vertex 1m away from the tunnel face is 0.9, which drops to 0.4 at 3m, and approaches 0 outside 5m); by calibrating parameters on site, the weight attenuation rate is dynamically adjusted in hard rock (slow attenuation) and soft rock (fast attenuation) to ensure the universality of the weight model. The non-linear characteristic (exponential attenuation) of the weight function amplifies the importance of the vertices at close range. For example, when the tunnel face advances 2m in front of point P2, the weight of P2 suddenly rises from 0.6 to 0.85, triggering high-frequency monitoring of the sides where P2 is located (such as P1P2, P2P3) (the acquisition frequency of the change rate is increased from 1 time / second to 5 times / second), realizing the intelligent allocation of monitoring resources.

[0060] The smoothed distance data , (i = 1, 2, 3, 4 corresponding to the four vertices) are input into the calibrated weight function in real time to calculate the current weight value of each vertex:

[0061] ;

[0062] For different vertices, due to geometric position differences, independently calibrated parameters may be used. For example, for the crown P1 and the invert P4, different k values may be set due to different stress wave propagation paths.

[0063] represents the weight value of the i-th vertex (i = 1, 2, 3, 4 corresponding to P1 to P4) at time t, reflecting the sensitivity of this vertex to the excavation of the tunnel face (the larger the weight, the more significant the vertex is affected by stress disturbance and belongs to the key deformation area).

[0064] represents the smoothed three-dimensional spatial Euclidean distance (unit: meter or millimeter) between vertex i and the tunnel face at time t, which is obtained by performing moving window (window size 10) average filtering on the original distance sequence (sampled 5 times per second) to eliminate high-frequency noise (such as instrument vibration, environmental interference). The closer the distance to the tunnel face ( is smaller), the greater the weight , reflecting the characteristic that "the vertices near the tunnel face are more directly affected by excavation stress" (in line with the law of rock mass stress wave attenuation with distance).

[0065] represents the amplitude parameter of the weight function of vertex i, which determines the maximum possible value of the weight (when is close to 0, the value of approaches +B i ), and different vertices may set different A due to geometric position differences (such as different stress propagation paths from the tunnel face for the crown P1 and the invert P4)i 。

[0066] k i represents the stress attenuation coefficient of vertex i, reflecting the rate at which the weight decays as the distance increases (the larger k i , the faster the weight decreases as the distance increases). For example, for the sidewall vertices (P2, P3), since they are close to the free face of the tunnel face excavation, k i can be set to a relatively large value to make its weight more sensitive to distance changes.

[0067] B i represents the offset of the weight function of vertex i, which is used to adjust the reference value of the weight (for example, when the distance is extremely far, the weight approaches B i , to avoid missing the judgment of key areas due to the weight becoming zero).

[0068] Weight threshold linkage mechanism:

[0069] Preset a weight threshold (such as 0.7). When the weight of any vertex > 0.7, the system automatically triggers real-time monitoring of the change rate of the adjacent edges of this vertex (for example, the adjacent edges of P2 are P1P2, P2P3, P2P4). Calculate the relative change amount of the side length within a unit time (such as △L / L0×100% / min), and synchronize the monitoring data to the edge computing node for real-time analysis. △L represents the absolute change amount of the side length within a unit time (such as the number of millimeters increased or decreased by a certain side within 1 minute); L0 represents the reference length of this side (the initial side length from the reference topology, used as a reference benchmark for the deformation amount). The relative change rate (△L / L0) can better reflect the deformation characteristics of the slope material than the absolute change amount (for example, the allowable strain thresholds of slopes with different materials are different). Combining the time dimension (per minute) can track the deformation rate in real time to determine whether to trigger an early warning.

[0070] The weight value is updated once per second to reflect in real time the distance changes caused by the advancement of the tunnel face or the displacement of the surrounding rock (for example, when the tunnel face advances 5m per day, the vertex weight decays dynamically according to an exponential law), avoiding the defect that fixed weights cannot adapt to working condition changes; the threshold trigger mechanism realizes "monitoring on demand": only when the vertex enters the high-weight area (close to the tunnel face), start high-frequency monitoring of the edge change rate. Compared with full-time high-frequency monitoring, it can reduce the sensor energy consumption and data transmission pressure by 40%. By focusing the weight on the vertices near the free face (for example, the weights of the sidewall vertices P2 / P3 within 3m behind the tunnel face are often higher than 0.7), the monitoring sensitivity of the side length change rate in this area is increased by 3 times (the sampling frequency is increased from 1 time / minute to 1 time / second). For example, in a certain tunnel project, it was successfully able to identify 2 hours in advance the sudden increase in the weight of the sidewall vertex caused by the tunnel face blasting (from 0.6 to 0.85), and capture the abnormal tensile change rate of the adjacent edge P2P3 (0.5% / min), avoiding the risk of missing the judgment of local landslides.

[0071] In a preferred embodiment of the present invention, when the weight of a vertex is greater than a threshold, the change rate of the corresponding edge is monitored in real time; temperature compensation is performed on the change rate, and the deformation caused by the temperature change is calculated by analyzing the thermal expansion characteristics of the material, including:

[0072] Normalize the weight of each vertex to a range of 0-1. When the vertex weight exceeds the preset deformation sensitivity threshold, mark the vertex as a high-risk node and activate the real-time change rate monitoring module of the edges connected to it. At the same time, bind the weight data to the spatial coordinates in the baseline topology to generate a weighted dynamic topology node information table, including:

[0073] The real-time weight of each vertex is uniformly scaled to the range of 0-1 through a linear transformation to eliminate incomparable weight values due to parameter differences between different vertices (for example, the initial weight range of P2 is 0.3-0.9, and that of P4 is 0.1-0.6, both of which fall within the 0-1 range after normalization). A preset deformation sensitivity threshold (such as 0.7) is set. When the normalized weight of a vertex exceeds this threshold, it is automatically marked as a high-risk node (for example, sidewall vertices P2 / P3 within 3 meters behind the tunnel face often trigger this flag), and the real-time monitoring module for all connected edges (for example, P2 connects to edges P1P2, P2P3, and P2P4) is activated.

[0074] Dynamic topology information binding:

[0075] The weight value of the high-risk node is bound to its three-dimensional spatial coordinates (x, y, z) in the baseline topology and its edge information (such as connected vertices, baseline edge length, and initial internal angle) to generate a weighted dynamic topology node information table (for example, the table fields include: node ID, weight value, x-coordinate, y-coordinate, connection edge 1-ID, connection edge 2-ID, etc.), providing a spatial reference for subsequent edge positioning and deformation analysis.

[0076] By using weight thresholds, we filter out vertices that are significantly disturbed by excavation (such as nodes near free surfaces), and only start high-frequency monitoring of the edges connected to them. Compared with indiscriminate monitoring of all edges, this can reduce 60% of invalid data collection and improve the monitoring efficiency of key areas. The dynamic topology table associates abstract weight values with specific spatial locations.

[0077] Based on the marked high-risk vertices, locate the connected quadrilateral edges. Start the high-frequency monitoring mode of the 3D laser rangefinder for the target edge, continuously collect raw edge length data at a sampling frequency of 1 Hz, and use the difference calculation between two adjacent sampling points to generate a sequence of raw length change rates. Remove outliers from the raw change rate data. The removal rule is: if the single-point change rate exceeds three standard deviations of the historical mean for the same period, mark it as noise and perform linear interpolation repair to obtain processed change rate data, including:

[0078] For the edges connected to high-risk nodes (such as the P1P2 edge connected to P2), start the 1Hz high-frequency monitoring mode of the 3D laser rangefinder (the conventional mode is once per minute), and continuously collect the original data of the side length (such as recording the side length values L(t1), L(t2),... once per second).

[0079] Calculate the original length change rate by the adjacent sampling point difference method: For example, the side length difference between the current moment t and the previous moment t-1 is △L = L(t) - L(t-1), and the change rate is expressed as △L / L0×100% / second (L0 is the reference side length of this edge), generating an original change rate sequence (such as deformation rate data including positive and negative values).

[0080] Outlier rejection and data repair:

[0081] Establish a 3-fold standard deviation rejection rule: Calculate the mean μ and standard deviation σ of the change rate in the historical same period (such as the past 1 hour). If the change rate of a single point exceeds μ±3σ, it is determined as noise (such as jump values caused by temporary occlusion of the equipment or sudden vibration).

[0082] Repair the noise points using the linear interpolation method (such as replacing with the average value of the previous valid point and the next valid point). For example, if the 5th sampling point is an outlier, fill it with the mean change rate (△L4+△L6) / 2 of the 4th and 6th points to obtain the smoothed change rate data.

[0083] The 1Hz sampling frequency can capture rapid deformations caused by instantaneous loads such as blasting excavation and rainfall infiltration (such as a sudden change in side length of 0.5mm within 1 second). Compared with the conventional low-frequency monitoring, the response speed to sudden deformations is increased by 60 times, avoiding missed judgments of short-duration high-risk deformations. By rejecting outliers through statistical laws (such as ±5mm jumps caused by dust interference), the standard deviation of the change rate data is reduced from ±0.8% / min to ±0.2% / min, providing clean input data for subsequent temperature compensation and avoiding misjudgments caused by noise.

[0084] Deploy a distributed temperature sensor array on the surface and inside of the rock mass on both sides of the target edge, collect the temperature field data of the entire section of the monitored edge, and generate a temperature spatio-temporal distribution matrix; call the pre-stored rock mass material thermal expansion coefficient library, and match the corresponding thermal expansion coefficient according to the material type of the current monitored edge; calculate the predicted deformation amount caused by temperature fluctuations point by point based on the thermal expansion coefficient, the original reference side length, and the real-time temperature difference, where the real-time temperature difference is the difference between the current temperature and the initial temperature at the time of generating the reference topology, including:

[0085] Optical fiber temperature sensors or thermocouples are arranged at intervals of 0.5 - 1 m on the rock mass surface (such as the sidewall concrete support layer) and inside (such as buried in boreholes) on both sides of the target edge, forming a temperature field acquisition network covering the entire length of the monitored edge (for example, 11 sensors are arranged on a 5 - m - long edge, 1 at each end and 1 every 0.5 m in the middle), and temperature data is collected in real - time (accuracy ±0.5 °C), generating a temperature spatio - temporal distribution matrix containing position coordinates and timestamps (such as a three - dimensional array: [edge position, time, temperature]).

[0086] Call the pre - stored thermal expansion coefficient library of rock mass materials (including thermal expansion coefficients of various materials such as sandstone, mudstone, and concrete, such as sandstone and concrete), and match the corresponding thermal expansion coefficient α according to the material type of the area where the monitored edge is located (such as determined through geological exploration reports or on - site sampling).

[0087] Calculate the real - time temperature difference △T = T_current - T_reference (T_reference is the initial temperature when the reference topology is generated, usually taking the ambient temperature at the start of monitoring). Based on the thermal expansion formula △L_thermal = α×L0×△T, calculate the temperature - induced deformation amount at the corresponding position of each sensor point - by - point (for example, when the temperature difference at a certain point is 10 °C and the reference side length is 10 m, the thermal deformation is 10×10 −6 ×10000 mm×10 = 1 mm).

[0088] The distributed sensor array covers the surface and inside of the monitored edge, solving the problem that single - point temperature measurement cannot reflect the overall temperature difference of the edge (for example, when the temperature difference at both ends of the edge is 3 °C, single - point measurement will cause a 20% compensation error), improving the temperature field resolution to 0.5 m / point and the compensation accuracy by 30%. Through material - specific thermal expansion coefficient matching (instead of a unified default value), precise compensation is carried out for different rock mass / support materials (such as the thermal expansion difference between steel supports and concrete linings), improving the side - length monitoring accuracy from ±2 mm to ±0.5 mm (for example, when the day - night temperature difference is 15 °C, the thermal deformation of a 10 - m steel string sensor is about 1.74 mm, and this error can be completely eliminated through compensation), ensuring that the rate - of - change data only reflects the true mechanical deformation rather than temperature pseudo - change.

[0089] In a preferred embodiment of the present invention, the monitored side length is corrected to generate a temperature - compensated side length, including:[[]]

[0090] Extract the sequential temperature data collected by the temperature sensors and the corresponding theoretical temperature deformation amounts;

[0091] Based on the thermal expansion characteristics of the rock mass material, perform a sliding time - window integration on the theoretical temperature deformation amounts in time series, accumulate the instantaneous deformation amounts within each time slice, and generate a temperature - compensated cumulative value curve;

[0092] Obtain the original side - length data measured in real - time by the three - dimensional laser rangefinder and align it with the temperature - compensated cumulative value according to the timestamp;

[0093] Perform compensation superposition calculation: Temperature-compensated side length = original side length + cumulative compensation value, where the sign of the compensation value is determined according to the temperature rise and fall direction. When the temperature rises and causes expansion, the compensation value is negative, and vice versa.

[0094] In the embodiments of the present invention, by accumulating the theoretical deformation amount (such as thermal expansion / contraction) caused by temperature changes, the interference of the ambient temperature on the side length monitoring is peeled off in real time, and the side length monitoring accuracy is improved from ±2 mm to ±0.5 mm. For example, in a scenario with a 20°C day-night temperature difference, the thermal deformation of a 10-m long steel sensor can reach 2.32 mm (steel thermal expansion coefficient 11.6×10 -6 / °C). Through compensation, this error can be completely eliminated, ensuring that the monitoring data only reflects the true mechanical deformation of the surrounding rock (such as joint dislocation, rock mass creep, etc.). By using a sliding time window integration (such as accumulating instantaneous deformation amounts with 1 minute as a time slice), the continuous change of temperature over time (such as non-linear expansion during the heating process) can be captured, avoiding the defect of traditional single-point compensation (only calculating with the current temperature difference) that ignores the temperature change process. For example, for a scenario with continuous heating for 3 hours, the expansion amount per minute and second can be accurately accumulated to generate a continuous compensation value curve, making the compensation effect more consistent with the actual deformation process. By aligning the original side length data with the temperature compensation cumulative value through timestamps, it is ensured that the two strictly correspond in the time and space dimensions (error < 10 ms), solving the compensation deviation problem caused by asynchronous acquisition of different sensors (laser rangefinder and temperature sensor). For example, when the laser rangefinder collects the side length at 10:00:05, the temperature cumulative compensation value at the same moment can be accurately matched, avoiding over-compensation or under-compensation caused by time misalignment (such as mistakenly using the compensation value at 10:00:00 for the data at 10:00:05, which may introduce an error of ±0.2 mm). When the temperature rises, the rock mass / sensor expands, and the measured side length will "falsely become longer". It is necessary to restore the true mechanical deformation by subtracting the expansion amount (the compensation value is negative); when the temperature drops, the contraction causes the measured side length to "falsely become shorter", and it is necessary to correct it by adding the contraction amount (the compensation value is positive), avoiding the reverse correction caused by confusion in the compensation direction (such as mistakenly setting the heating compensation value as positive), ensuring that the corrected side length completely eliminates the temperature pseudo-deformation.

[0095] The side length data after temperature compensation is used as an intermediate result and directly input into the surrounding rock rheological equation (such as the Burgers model) for aging deformation correction, avoiding the interference of temperature noise on creep analysis (such as when not compensated by traditional methods, thermal expansion may be misjudged as plastic deformation of the rock mass, resulting in a calculation deviation of rheological parameters exceeding 40%). At the same time, the pure side length data improves the calculation accuracy of three-dimensional deformation indexes (such as diagonal displacement difference, adjacent side curvature ratio), provides high-quality input for the spatio-temporal attention GNN model, and finally improves the early warning accuracy from 65% to 92%.

[0096] In a preferred embodiment of the present invention, according to the temperature-compensated side lengths and the surrounding rock rheological equation, the adjacent two side lengths are corrected by the strain difference between adjacent sides to obtain the final side lengths; the final side lengths are combined with the reference interior angles to generate a topological structure, including:

[0097] The distributed fiber optic strain sensor array is used to collect the axial strain distribution data of adjacent two sides in real time, and the average strain difference between the two sides is calculated based on the collected strain distribution data, including: in the rock mass surface or the support structure of adjacent two sides (such as P1P2 and P2P3) of the quadrilateral, the distributed fiber optic strain sensors (such as BOTDA / BOTDR) are arranged at intervals of 0.2 - 0.5 m along the axial direction, and the axial strain data of each measuring point (accuracy ±5 με) are collected in real time to form a strain distribution curve (such as the strain sequence of the P1P2 side is ε1, ε2,..., εn); the arithmetic mean of the strain data of each side is taken to obtain the average strain value of the side (such as εg1 corresponds to the P1P2 side, εg2 corresponds to the P2P3 side), and the difference △ε = εg1 - εg2 is calculated, which reflects the deformation difference caused by the rock mass rheology between adjacent sides.

[0098] Based on the average strain difference, the aging deformation parameters in the surrounding rock rheological equation, including the viscoelastic coefficient and the creep rate, are called, and combined with the time difference between the current monitoring time and the reference topology generation time, the co-deformation ratio of adjacent sides caused by the rheological action is calculated, including: the viscoelastic coefficient of the current stratum (such as the spring stiffness k1, k2 and the viscosity η1, η2 of the dashpots in the Burgers model) and the creep rate β (reflecting the strain growth trend per unit time, calibrated by in-situ creep tests, such as β = 0.01% / h for mudstone) are retrieved from the pre-established surrounding rock rheological parameter library.

[0099] The time difference △t (accurate to hours) between the current monitoring time and the reference topology generation time is calculated. Based on the time dependence of the rheological equation, the co-deformation ratio γ of adjacent sides under the rheological action is deduced, such as , indicating that with the increase of time, the proportion of co-deformation of adjacent sides caused by rheology. When △t = 0, γ = 0 (no time effect, co-deformation is 0); as △t increases, γ approaches 1 (long-term creep causes the deformation of adjacent sides to tend to be coordinated), which conforms to the law of progressive development of viscoelastic deformation with time in the Burgers model.

[0100] According to the co-deformation ratio of adjacent sides, the length change amount of adjacent sides caused by the rheological effect is deduced, including:

[0101] According to the co-deformation ratio γ and the average strain difference △ε, calculate the additional length change △L_rheology caused by the rheological effect for each side. For example, if the reference length of the adjacent side P1P2 is L0, its rheological deformation amount is △L_rheology1 = γ×△ε×L0, which reflects the proportional co-variation of the side lengths caused by the long-term creep of the rock mass (if the strain difference is positive, it means that the tensile deformation of the P1P2 side is greater than that of the P2P3 side, and the rheological correction needs to adjust the difference between the two).

[0102] Take the temperature-compensated side length as the input, combine it with the length change of the adjacent side, and use the rheological correction model to reversely correct the lengths of the two adjacent sides to obtain the corrected side lengths, including: taking the temperature-compensated side length L_warmcomp as the input, combining it with the length change △L_rheology caused by rheology, and adjusting the side length through the reverse correction logic: if rheology causes side A to elongate and side B to shorten, then subtract △L_rheology from side A and add △L_rheology to side B (or vice versa) to make the deformation difference of the adjacent sides conform to the co-variation characteristics of the rock mass rheology, and obtain the final corrected side length L (such as L_final1 = L_warmcomp1 - △L_rheology1, L_final2 = L_warmcomp2 + △L_rheology2).

[0103] Based on the corrected side lengths and the reference interior angle data, with P1 as the coordinate origin, calculate the spatial coordinates of P2, P3, and P4 vertex by vertex; use the least squares method to fit the curvature of the quadrilateral section and update the curvature radius parameters; integrate the corrected side lengths, updated interior angles, and curvature radius parameters to generate a dynamic topological structure including geometric parameters and curvature information, including:

[0104] Taking the crown settlement point P1 in the reference topology as the coordinate origin (0, 0, 0), based on the corrected side lengths (L_final1, L_final2, etc.) and the reference interior angle (such as ∠P1P2P3), calculate the three-dimensional spatial coordinates of P2, P3, and P4 vertex by vertex through the geometric analysis method.

[0105] For example, given the lengths of P1P2 and P1P3, and the interior angle of ∠P1, calculate the coordinates of point P2 (x2, y2, 0) using the cosine theorem, and then determine the z coordinate of point P4 (reflecting the amount of invert heave) through the three-dimensional distance constraint.

[0106] The curvature fitting and parameter update are realized as follows:

[0107] Import the coordinates of each vertex of the quadrilateral and the section point cloud data (from the point cloud scan of the reference topology) into the fitting algorithm, use the least squares method to fit a quadratic curve, calculate the curvature radii R of the crown, side wall, and invert (such as the decrease in R due to crown settlement reflects the intensification of bending deformation), and update the curvature parameters (such as the change of R_crown from 10m to 9.5m indicates an increase in the local bending degree).

[0108] The dynamic topology integration is realized as follows:

[0109] Integrate the corrected side lengths (L_final1 - L_final6), updated interior angles (considering the small angle changes caused by rheology), vertex coordinates (x, y, z), and radius of curvature (R_crown, R_side wall, R_inverted arch) into a dynamic topological structure, with each parameter attached with a timestamp and weight information (e.g., high-risk edges are marked red and low-risk ones are marked blue), to form a three-dimensional geometric model that can be updated in real time.

[0110] In the embodiments of the present invention, the traditional method only focuses on immediate deformation and ignores time-dependent effects such as rock mass creep and relaxation (e.g., the creep strain of mudstone can reach 0.3% within 30 days). By introducing the time dimension through the rheological equation, the long-term deformation trend can be accurately captured, avoiding misjudging creep as a stable state; the internal joint dislocation or interlayer slip of the rock mass is identified by the strain difference between adjacent edges (e.g., the strain difference is enlarged due to the different creep rates of sandstone and mudstone in interbedded rock masses), and the corrected side lengths are closer to the true mechanical deformation, enabling the progressive instability process of the slope to be quantitatively tracked (e.g., in a tunnel case, the accelerated deformation caused by mudstone softening was discovered 15 days in advance through strain difference correction). Integrate geometric deformations (side lengths, interior angles), spatial positions (coordinates), and morphological features (curvature) to construct a full-element topological model including "points - lines - planes". Compared with traditional single-point monitoring, it can describe the complete mode of "areal deformation" of the slope (such as the abnormal curvature combination of crown settlement accompanied by inward concavity of the side wall). The dynamic update of parameters such as the radius of curvature provides real-time data for subsequent deformation index calculations (such as the curvature ratio of adjacent edges), supporting the spatio-temporal attention GNN model to capture non-linear deformation features (such as the continuous decreasing trend before the curvature mutation), increasing the early warning accuracy by 27% (from 65% to 92%). First, eliminate environmental noise through temperature compensation, and then eliminate the interference of time-dependent deformation through rheological correction to form a closed loop of "data purification - physical mechanism modeling - dynamic correction", improving the side length monitoring accuracy from ±2 mm to ±0.5 mm and being applicable to long-term monitoring.

[0111] In a preferred embodiment of the present invention, construct a three-dimensional deformation index matrix including the diagonal displacement difference, adjacent edge curvature ratio, and interior angle coefficient of variation, including:

[0112] Extract the current lengths of the two diagonals of the quadrilateral from the dynamic topological structure and compare them with the initial diagonal lengths in the reference topology; calculate the absolute displacement difference and the displacement difference change rate of each diagonal to generate a displacement difference feature vector of the two diagonals;

[0113] Based on the updated cross-section curvature parameters in the dynamic topology, obtain the radius of curvature and curvature ratio of adjacent sides, normalize the curvature ratio, and map the ratio to the 0 - 1 interval to generate a curvature ratio feature vector;

[0114] Read the current angular values of the four interior angles of the quadrilateral from the dynamic topology, calculate the standard deviation and coefficient of variation of the interior angles, and perform a sliding window statistical analysis on the coefficient of variation to generate a smoothed coefficient of variation sequence;

[0115] Align the diagonal displacement difference feature vector, curvature ratio feature vector, and interior angle coefficient of variation sequence according to the time stamp; construct a three-dimensional deformation index matrix;

[0116] Perform data standardization on the three-dimensional deformation index matrix to generate a standardized deformation index matrix; the dimension of the three-dimensional deformation index matrix is [time stamp × deformation feature], where each row contains: the first column: the change rate of the displacement difference between the two diagonals; the second column: the normalized value of the curvature ratio of adjacent sides; the third column: the coefficient of variation of the interior angle.

[0117] In the embodiment of the present invention, the implementation process of constructing the three-dimensional deformation index matrix:

[0118] Obtain the real-time lengths of the two diagonals (such as P1P3, P2P4) of the current quadrilateral from the dynamic topological structure, and at the same time retrieve the initial lengths (benchmark values) of these two diagonals in the reference topology; calculate the absolute displacement difference of each diagonal (current length minus benchmark length), and further calculate the change rate of the displacement difference per unit time (such as the change amount of the displacement difference per hour) to form a two-dimensional vector containing the change rates of the two diagonals (for example, [0.3mm / h, -0.1mm / h]) as the diagonal displacement difference feature vector. The diagonal connects the opposite vertices of the quadrilateral, and its length change sensitively reflects the in-plane shear deformation (such as joint dislocation causing one diagonal to elongate and the other to shorten), and the change rate captures the deformation acceleration trend.

[0119] Extract the cross-sectional curvature radii (such as the crown curvature R1 and the sidewall curvature R2) corresponding to two adjacent sides (such as P1P2 and P2P3) in the dynamic topology, and calculate their ratio (R1 / R2 or R2 / R1, determined according to the position of the side). Use a normalization method (such as Min-Max normalization) to map the curvature ratio to the 0-1 interval (for example, the original ratio of 5.0 is mapped to 1.0, and the ratio of 1.0 is mapped to 0.5) to eliminate the influence of dimensions and generate a one-dimensional curvature ratio feature vector (such as [0.8] indicating that the current curvature ratio is in the high value interval). The curvature ratio reflects the difference in the degree of bending of adjacent regions, and after normalization, it is convenient for horizontal comparison between different cross-sections and different times to identify local convex and concave deformations (such as the sudden increase in the curvature ratio caused by the inward concavity of the sidewall).

[0120] Read the real-time angular values of the four interior angles of the quadrilateral in the dynamic topology (such as ∠P1, ∠P2, ∠P3, ∠P4), calculate the standard deviation of these four angles (reflecting the degree of angular dispersion), and then divide by the average angle to obtain the coefficient of variation of the interior angles (dimensionless, characterizing the uniformity of the angle distribution). Apply a sliding window smoothing process to the coefficient of variation sequence (such as a window size of 10 time points), eliminate sudden noises (such as angular jumps caused by instrument jitter), and generate a smoothed coefficient of variation sequence (such as outputting a smoothed value every 10 minutes). The coefficient of variation of the interior angles sensitively captures angular anomalies caused by joint dislocation or rock mass fragmentation (such as a sudden decrease of 2° in a certain interior angle, resulting in a sharp increase in the coefficient of variation), and the smoothing process improves data stability.

[0121] Strictly align the diagonal displacement difference eigenvector (2D), curvature ratio eigenvector (1D), and interior angle coefficient of variation sequence (1D) according to the time stamp (error < 1 second) to ensure that multi-feature data at the same moment corresponds to the same deformation state. Construct a three-dimensional matrix according to the "time stamp × deformation feature" dimension. Each row contains three core indicators:

[0122] The rate of change of the displacement difference between the two diagonals (2 columns);

[0123] The normalized curvature ratio (1 column);

[0124] The smoothed coefficient of variation of the interior angles (1 column).

[0125] Adopt the Z-Score standardization method to zero the mean and normalize the standard deviation of the data in each column of the matrix, eliminating the dimension difference (such as the unification of the displacement difference unit of mm / h and the dimensionless coefficient of variation), and generating a standardized deformation index matrix. Time alignment ensures the spatio-temporal consistency of multi-features, and standardization provides a unified input format for the subsequent neural network model, improving the training efficiency and prediction accuracy.

[0126] The diagonal displacement difference directly reflects the in-plane shear deformation (such as the expansion of the diagonal length difference caused by the internal dislocation of the landslide body). Compared with single-point displacement monitoring, it can capture the shear feature of "one pair of vertices approaching and the other pair moving away", and the identification efficiency is increased by 60%. Quantify the difference in the bending degree of adjacent sides (such as when the crown settlement is accompanied by the inward concavity of the side wall, the curvature ratio decreases significantly), and is sensitive to local buckling instability (such as abnormal curvature before the shotcrete cracks), making up for the defect that traditional monitoring does not pay attention to "shape changes". Through the change in angle uniformity (such as a sudden change in a certain interior angle resulting in an increase in the coefficient of variation), early identification of joint surface sliding or rock mass fragmentation (such as an increase in the angular dispersion caused by blasting vibration) can be achieved, with a sensitivity of 0.5°.

[0127] Normalization and smoothing can eliminate the influence of dimensions (such as unifying the unit of the radius of curvature m and the unit of displacement difference mm), filter out high-frequency noise (such as accidental fluctuations caused by environmental vibrations), and make the index sequence more conform to the true deformation trend (such as the gradual increase in the coefficient of variation during mudstone creep can be clearly captured); time alignment and standardization can provide high-quality input for the spatio-temporal attention GNN model (avoiding mis-correlation of features caused by time misalignment), enabling the model to effectively learn the spatio-temporal correlation of deformation indicators (such as a continuous 3-hour decrease in the curvature ratio and an increase in the coefficient of variation of the interior angle at a certain section indicating potential instability).

[0128] In a preferred embodiment of the present invention, the index matrix, the spatio-temporal matrix of the support structure strain, and the rock mass fragmentation tensor are jointly input into the spatio-temporal attention graph neural network, and the deformation evolution trend including the diagonal displacement difference, the adjacent side curvature ratio, and the interior angle coefficient of variation is output, including:

[0129] Align the standardized deformation index matrix, the spatio-temporal matrix of the support structure strain constructed from the time series data of the support structure strain collected by the fiber optic sensor, and the rock mass fragmentation tensor calculated based on the microseismic event energy distribution and the rock mass fracture density according to the time stamp to obtain time-aligned data;

[0130] Based on the time-aligned data, perform dimensionality reduction on the rock mass fragmentation tensor, extract its principal component eigenvectors, and make its column dimension match the deformation index matrix;

[0131] Using the principal component eigenvectors, together with the deformation index matrix and the spatio-temporal matrix of the support structure strain, splice the three types of data features along the time axis to construct a fused spatio-temporal tensor;

[0132] According to the fused spatio-temporal tensor, use the multi-head attention mechanism to calculate the spatio-temporal dependence relationships between the three types of data respectively;

[0133] According to the spatio-temporal dependence relationships, multiply the fused spatio-temporal tensor by the attention weight matrix, and aggregate the spatial correlation features of the quadrilateral vertices and their support nodes through the graph convolutional network;

[0134] Based on the spatial correlation features, use the gated recurrent unit to learn the long-term evolution trend of the deformation indicators;

[0135] Input the long-term evolution trend into the fully connected prediction module, and output the deformation evolution trend parameters within the next 3 hours, including: the increment of the diagonal displacement difference change rate, the change gradient of the adjacent side curvature ratio, and the acceleration of the interior angle coefficient of variation.

[0136] In the embodiments of the present invention, a standardized deformation index matrix generated from the dynamic topology (including diagonal displacement difference, curvature ratio, and coefficient of variation of interior angles); a strain spatio-temporal matrix of the support structure collected in real time by fiber Bragg grating sensors (such as strain time series data of anchor cables and bolts, with a resolution of 1% and time stamps accurate to seconds); a rock mass fragmentation tensor based on a microseismic monitoring system (calculated through microseismic event energy distribution and fracture density, reflecting the integrity of the rock mass, with an initial dimension of 30×30×30). The time stamps of the three types of data are matched with millisecond-level accuracy, and abnormal data points with a time misalignment exceeding 10 ms are removed to ensure a one-to-one correspondence between the deformation index, support response, and rock mass state at the same moment.

[0137] Apply principal component analysis (PCA) to the high-dimensional rock mass fragmentation tensor (such as 30×30×30), calculate its covariance matrix and extract the first 3 principal components (cumulative variance contribution rate > 95%), and compress the tensor into a 3-dimensional principal component feature vector (such as [PC1, PC2, PC3]). Adjust the dimension of the feature vector to make the number of columns consistent with the deformation index matrix (3 columns: rate of change of displacement difference, normalized curvature ratio, coefficient of variation) for subsequent splicing (such as expanding to 3 columns: PC1 corresponding to displacement difference, PC2 corresponding to curvature ratio, PC3 corresponding to coefficient of variation), which can reduce the data complexity and retain the main features of the rock mass fragmentation (such as fracture development direction and distribution of fracture zones), and avoid overfitting in model training caused by high-dimensional data.

[0138] Splice the three types of data by column along the time axis:

[0139] Columns 1 - 3: Standardized deformation indices (rate of change of displacement difference, normalized value of curvature ratio, coefficient of variation of interior angles);

[0140] Columns 4 - 6: Strain parameters of the support structure (such as rate of change of anchor cable tension, growth rate of bolt axial force);

[0141] Columns 7 - 9: Principal component features of the reduced-dimensional rock mass fragmentation.

[0142] Form a fused spatio-temporal tensor (dimension: time stamp × 9 features). For example, the data at a certain moment is [0.5 mm / h, 0.8, 0.25, 1.2%, -0.5%, 0.8%, 0.6, -0.3, 0.9], including three types of features: deformation, support, and rock mass state.

[0143] Convert heterogeneous data (geometric deformation, structural response, rock mass properties) into a feature space with a unified dimension.

[0144] Apply a multi-head attention layer (such as 8 heads) to the fused spatio-temporal tensor, and each head independently calculates the attention weights between the three types of data:

[0145] Time dimension: Analyze the dependence relationship between the features at the current moment and the features in the past 12 hours (such as the correlation between the current curvature ratio and the displacement difference 3 hours ago);

[0146] Spatial dimension: Calculate the spatial correlation between the vertices of the quadrilateral (such as the correlation between the strain at point P2 and the fragmentation degree at point P3).

[0147] Generate an attention weight matrix to quantify the interaction intensity of different features, different moments, and different spatial positions (for example, the higher the weight value, the closer the correlation).

[0148] The graph convolutional network aggregates spatially correlated features. The implementation process is as follows:

[0149] Construct a graph structure: Use the 4 vertices of the quadrilateral and 2 support nodes (such as the anchor cable anchorage point and the bolt end point) as nodes, and use the edges (vertex connection relationships, support-rock mass contact relationships) as edges. Define node features (coordinates, weights) and edge features (edge lengths, strain differences). Through graph convolution operations (such as GCN layers), each node aggregates the features of adjacent nodes (for example, node P2 fuses the strain and fragmentation degree information of P1, P3, and the support nodes), generating high-order features containing spatial correlations (such as a comprehensive index reflecting the coordinated deformation of vertices and supports). Utilize the geometric topology structure of the slope (quadrilateral element) to explicitly model the spatial proximity effect (for example, the deformation of adjacent edges will affect each other), and solve the defect that traditional neural networks ignore spatial position relationships.

[0150] The gated recurrent unit learns the long-term evolution trend. The implementation process is as follows:

[0151] Input the spatially correlated features output by the graph convolution into the GRU layer, and selectively remember historical information through the gating mechanism: update gate, determine how much of the past deformation trend to retain (for example, a decreasing trend in the curvature ratio lasting for 12 hours will be strongly remembered); reset gate, forget short-term noise (such as fluctuations in the coefficient of variation caused by accidental vibrations); output a long-term evolution feature vector to capture the time dependence of deformation indicators (for example, the rate of change of the displacement difference caused by mudstone creep gradually increases over time). By processing non-stationary time series (such as the non-linear trend in the deformation acceleration stage), compared with traditional LSTM, it captures long-term dependencies with fewer parameters and improves training efficiency.

[0152] Input the long-term features output by the GRU into the fully connected prediction module, map them to the target space through a multi-layer perceptron (MLP), and output the deformation evolution trend parameters for the next 3 hours: the increment of the rate of change of the diagonal displacement difference (for example, it is expected to increase by 0.2 mm / h in the next 1 hour); the change gradient of the curvature ratio of adjacent edges (for example, the rate of decrease of the curvature ratio accelerates by 0.05 per hour); the acceleration of the coefficient of variation of the interior angle (for example, the second derivative of the coefficient of variation, reflecting the acceleration degree of angle anomalies).

[0153] Deformation indicators (such as curvature ratio) reflect geometric deformation, support strain (such as anchor cable tension) reflects structural response, and fragmentation tensor (such as fracture density) characterizes rock mass properties. The combined input of these three types of data avoids the one-sidedness of a single indicator (for example, only looking at displacement may miss the hidden risks caused by misjudging support failure). The multi-head attention mechanism identifies the temporal causal relationship of deformation indicators (for example, when the change rate of displacement difference continuously rises for 3 hours, the curvature ratio usually becomes abnormal accordingly). Compared with traditional statistical models, the ability to capture non-linear trends is increased by 40%. Graph convolution explicitly utilizes the quadrilateral topology of the slope to quantify the spatial synergy effect between vertices (for example, the deformation at point P2 will affect the curvature at point P3 through the edge P2P3), solves the defect of "unstructured input data" in traditional neural networks, and improves the utilization rate of spatially related features by 60%.

[0154] In a preferred embodiment of the present invention, multi-level early warning decisions are executed according to the component parameters of the deformation evolution trend, including:

[0155] Extract the increment of the diagonal displacement difference change rate, the change gradient of the adjacent side curvature ratio, and the acceleration parameter of the inner angle coefficient of variation in the deformation evolution trend; based on the statistical correlation analysis of historical deformation data and support structure failure cases, set the grading threshold intervals for the three parameters, including the yellow early warning threshold interval, the orange early warning threshold interval, and the red early warning threshold interval;

[0156] When any parameter first enters the yellow early warning threshold interval, increase the three-dimensional laser ranging frequency of the corresponding monitored side; associate the early warning information with the coordinates of high-risk nodes in the reference topology to generate a primary early warning signal with spatial positioning;

[0157] When two or more parameters enter the orange early warning threshold interval synchronously, match the pre-stored support reinforcement strategy library according to the parameter combination mode, dynamically calculate the required number of bolt reinforcements and grouting pressure values, and generate a support parameter adjustment instruction set;

[0158] When all three parameters reach the red early warning threshold interval and the acceleration of the inner angle coefficient of variation exceeds the critical value, send an evacuation instruction to the on-site personnel terminal, automatically cut off the power supply of the heading face construction equipment, and activate the rapid grouting and hydraulic support units of the emergency support system;

[0159] Dynamically update the topology structure database according to the early warning level, record the early warning trigger time, parameter overrun values, and disposal measures to the blockchain evidence storage node, and form an early warning decision chain.

[0160] In the embodiment of the present invention, extract the deformation parameter sequence (such as displacement difference change rate, curvature ratio gradient, etc.) in the past 3 years from historical monitoring data, combine with support structure failure cases of similar projects (such as parameter values when anchor bolts break and shotcrete cracks), and determine the three-level threshold intervals of each parameter through statistical analysis (such as quantile method, ROC curve):

[0161] Yellow warning (local deformation): The parameter value exceeds 25% of the historical extreme value and continues to rise;

[0162] Orange warning (progressive instability): The parameter value exceeds 50% of the historical extreme value and shows an accelerating trend;

[0163] Red warning (overall instability): The parameter value exceeds 90% of the historical extreme value and breaks through the engineering design safety limit.

[0164] Extract the three component parameters at the current moment in real time: the increment of the change rate of the diagonal displacement difference (such as 0.4 mm / h), the change gradient of the curvature ratio (such as -0.1 / hour), and the acceleration of the coefficient of variation of the interior angle (such as 0.06 / hour 2 ) The threshold is set by combining historical laws and engineering experience to avoid the limitations of a single fixed threshold (for example, different thresholds are used for slopes with different surrounding rock grades).

[0165] Yellow warning: Primary response and precise monitoring, implementation process:

[0166] When any parameter first enters the yellow warning range (such as the increment of the change rate of the displacement difference > 0.3 mm / h), the system automatically increases the three-dimensional laser ranging frequency of the corresponding monitored side from 1 Hz to 5 Hz to densely collect data to capture the subtle deformation trend.

[0167] Associate the warning parameter with the coordinates of the high-risk node in the reference topology (such as at point P2, x = 1234.5 m, y = 67.8 m, z = 9.0 m) to generate a primary warning signal containing the precise location (such as "Node P2 on the left side wall of a certain section enters the yellow warning"), and push it to the on-site engineer via text message / APP.

[0168] Confirm the deformation trend through high-frequency monitoring to avoid misjudgment (for example, a sudden increase in a single parameter caused by a short-term vibration may be excluded by subsequent data).

[0169] Orange warning: Intelligent matching and support adjustment, implementation process is as follows:

[0170] When two or more parameters simultaneously enter the orange warning range (such as the displacement difference increment > 0.5 mm / h and the curvature ratio gradient < -0.15 / hour), the system matches the optimal solution from the pre-stored support reinforcement strategy library according to the parameter combination mode (such as the "displacement acceleration + curvature anomaly" corresponding to the rock mass shear slip mode):

[0171] If the "concave side wall" mode is matched, automatically calculate the required number of bolt reinforcements (such as increasing 2 bolts per meter) and the grouting pressure value (such as increasing from 0.3 MPa to 0.6 MPa);

[0172] Generate a support parameter adjustment instruction set and send it to the on-site intelligent terminal via industrial Ethernet to drive the automatic adjustment of the robotic arm or grouting equipment.

[0173] Based on the mapping relationship of "deformation mode - support measures" (such as 20 sets of adjustment schemes corresponding to 10 typical instability modes), the strategy library realizes a closed-loop from early warning to disposal, avoiding the lag of manual decision-making.

[0174] Red early warning: Emergency disposal is linked with the system, and the process is as follows:

[0175] When all three parameters enter the red early warning range (such as displacement difference increment > 1.0 mm / h, curvature ratio gradient < -0.2 / hour, coefficient of variation acceleration > 0.1 / hour 2 ), and the coefficient of variation acceleration of the interior angle exceeds the critical value (such as 0.08 / hour 2 , indicating that the angle anomaly deteriorates rapidly), the system triggers a triple emergency response:

[0176] Send audible and visual alarms and text evacuation instructions (such as "Evacuate immediately to the refuge chamber") to all on-site personnel terminals; automatically cut off the power supply of the face construction equipment (such as excavators, loaders) through the PLC control system to prevent equipment out-of-control during instability; start the rapid grouting system (reach the design pressure of 1.0 MPa within 5 minutes) and the hydraulic support unit (automatically erect steel arch frames) to form a temporary rigid support.

[0177] The simultaneous overlimit of multi-dimensional parameters is used as the criterion for "overall instability", and the acceleration parameter is combined to identify the precursor of mutation to ensure that the response is triggered in the last window period (such as starting emergency measures 10 minutes in advance in a certain tunnel case to avoid a collapse accident).

[0178] Data storage and dynamic update, and the process is as follows:

[0179] Dynamically update the topology structure database according to the early warning level, and record key data such as current deformation parameters, vertex weights, and curvature radii (such as marking the cross-section as a "high-risk state" during red early warning); encrypt and write information such as the early warning trigger time, parameter overlimit values, and disposal measures into the blockchain storage node to form an unforgeable early warning decision chain. Blockchain technology ensures data integrity (such as preventing human tampering with early warning records) and provides a reliable data source for project safety audits.

[0180] Multi-dimensional criteria are used to avoid misjudgment of a single parameter (such as only displacement exceeding the standard may be caused by insufficient temperature compensation, and combining the curvature ratio and coefficient of variation can exclude environmental interference), and the threshold is dynamically adjusted based on real-time data and historical cases (such as automatically lowering the orange early warning threshold during the rainy season) to adapt to the working condition changes under complex geological conditions.

[0181] Only increase the monitoring frequency for the edges / nodes where anomalies first appear. Compared with full - range high - frequency monitoring, it reduces the energy consumption of sensors and data transmission pressure by 60%. The warning signal with coordinates enables on - site personnel to quickly locate the problem area (such as arriving at a certain side wall within 3 minutes), shortening the troubleshooting time by 50%.

[0182] The automatic linkage from warning to equipment adjustment (response time < 30 seconds) can timely contain the trend during the stage of accelerating deformation (for example, in a certain slope case, the displacement difference increment is reduced from 0.6 mm / h to 0.2 mm / h through bolt reinforcement).

[0183] The above - mentioned is the preferred implementation mode of the present invention. It should be pointed out that for those of ordinary skill in the art of this technology, without departing from the principle described in the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for monitoring road slope data under complex geological conditions, characterized in that, The method includes: Measuring the lengths of all sides and interior angles of the quadrilateral in real time to generate a reference topology; Based on the reference topology and the real-time distances between each vertex of the quadrilateral and the tunnel face, calculating the weight of each vertex; when the weight of a vertex is greater than the threshold, monitoring the change rate of the corresponding side in real time; performing temperature compensation on the change rate, by analyzing the thermal expansion characteristics of the material, calculating the deformation amount caused by temperature change, and correcting the monitored side lengths to generate temperature-compensated side lengths; according to the temperature-compensated side lengths and the surrounding rock rheological equation, correcting the lengths of two adjacent sides by the strain difference between adjacent sides to obtain the final side lengths; combining the final side lengths with the reference interior angles to generate a topological structure; Based on the topological structure, constructing a three-dimensional deformation index matrix including the diagonal displacement difference, the adjacent side curvature ratio, and the interior angle coefficient of variation; jointly inputting the index matrix, the support structure strain spatio-temporal matrix, and the rock mass fragmentation tensor into a spatio-temporal attention graph neural network, and outputting the deformation evolution trend including the diagonal displacement difference, the adjacent side curvature ratio, and the interior angle coefficient of variation; Performing multi-level early warning decisions according to the component parameters of the deformation evolution trend.

2. The method for monitoring road slope data under complex geological conditions according to claim 1, characterized in that, The reference topology is: Layout the crown settlement point P1, the left convergence point P2, the right convergence point P3, and the invert uplift point P4 at the selected positions in the tunnel section. Measure the initial lengths of all sides of the quadrilateral and the interior angles between adjacent sides in real time through a three-dimensional laser rangefinder, and establish an initial quadrilateral geometric parameter database; Collect the inclination angle change amounts of the two side walls through the biaxial inclinometer array between P2 and P3, synchronously trigger the three-dimensional section scanner to perform point cloud scanning on the section where the quadrilateral is located, and perform spatial registration on the scanning data and the laser ranging results to obtain calibration data; According to the calibration data, dynamically correct the lengths of all sides and interior angles of the initial quadrilateral to generate a reference topology including the spatial coordinates of the side lines, the interior angle values, and the section curvature.

3. A method for monitoring road slope data under complex geological conditions according to claim 2, characterized in that Based on the reference topology and the real-time distances between each vertex of the quadrilateral and the tunnel face, calculating the weight of each vertex, including: Continuously measure the three-dimensional spatial Euclidean distances between the vertices P1, P2, P3, and P4 of the quadrilateral and the tunnel face through a three-dimensional laser rangefinder at a sampling frequency of 5 times per second to generate the original dynamic distance sequences of each vertex; perform sliding window average filtering on the original sequences, with a window size of 10 sampling points, to generate the smoothed dynamic distance time series data; Based on the propagation attenuation characteristics of rock mass stress waves, establish a negative exponential mapping relationship between the vertex weight and the real-time distance, and fit the stress attenuation curve to determine the weight function parameters; Input the smoothed dynamic distance time series data into the weight function parameters to calculate the weight values of each vertex in real time.

4. A method for monitoring road slope data under complex geological conditions according to claim 3, characterized in that, When the weight of a vertex is greater than the threshold, monitor the change rate of the corresponding side in real time; Perform temperature compensation on the change rate, by analyzing the thermal expansion characteristics of the material, calculating the deformation amount caused by temperature change, including: Normalize the weight values of each vertex to the range of 0-1. When the vertex weight exceeds the preset deformation sensitivity threshold, mark the vertex as a high-risk node and activate the real-time change rate monitoring module of the connected side; at the same time, bind the weight data to the spatial coordinates in the reference topology to generate a weighted dynamic topology node information table; Based on the marked high-risk vertices, locate the connected quadrilateral sides; activate the high-frequency monitoring mode of the 3D laser rangefinder for the target sides, continuously collect the original side length data at a sampling frequency of 1 Hz, and use the difference calculation between adjacent two sampling points to generate the original length change rate sequence; eliminate outliers from the original change rate data, and the elimination rule is: if the single-point change rate exceeds 3 times the standard deviation of the historical average in the same period, it is marked as noise and linearly interpolated for repair to obtain the processed change rate data; Deploy a distributed temperature sensor array on the surface and inside of the rock mass on both sides of the target side to collect the temperature field data of the entire monitored section and generate a temperature spatio-temporal distribution matrix; call the pre-stored rock mass material thermal expansion coefficient library and match the corresponding thermal expansion coefficient according to the material type of the current monitored side; calculate the predicted deformation caused by temperature fluctuations point by point based on the thermal expansion coefficient, the original reference side length, and the real-time temperature difference, where the real-time temperature difference is the difference between the current temperature and the initial temperature at the time of generating the reference topology.

5. A method for monitoring road slope data under complex geological conditions according to claim 4, characterized in that, Correct the monitored side lengths to generate temperature-compensated side lengths, including: Extract the time-series temperature data collected by the temperature sensors and the corresponding theoretical temperature deformation amounts; Based on the thermal expansion characteristics of the rock mass material, perform a sliding time window integration on the theoretical temperature deformation amounts in time series, accumulate the instantaneous deformation amounts within each time slice, and generate a temperature compensation cumulative value curve; Obtain the original side length data measured in real time by the 3D laser rangefinder and align it with the temperature compensation cumulative value according to the timestamp; Perform a compensation superposition calculation: temperature-compensated side length = original side length + cumulative compensation value, where the sign of the compensation value is determined according to the temperature rise and fall direction. When the temperature rises and causes expansion, the compensation value is negative, and vice versa.

6. The method for monitoring road slope data under complex geological conditions according to claim 5, characterized in that, According to the temperature-compensated side length and the surrounding rock rheological equation, correct the lengths of two adjacent sides by the strain difference between adjacent sides to obtain the final side lengths; Combine the final side lengths with the reference interior angles to generate a topological structure, including: Use the distributed fiber optic strain sensor array to collect the axial strain distribution data of two adjacent sides in real time, and calculate the average strain difference between the two sides based on the collected strain distribution data; Based on the average strain difference, call the aging deformation parameters in the surrounding rock rheological equation, including the viscoelastic coefficient and the creep rate, and combine the time difference between the current monitoring time and the time of generating the reference topology to calculate the adjacent side co-deformation ratio caused by the rheological effect; Derive the length change amount of the adjacent sides caused by the rheological effect according to the adjacent side co-deformation ratio; Take the temperature-compensated side length as the input, combine the adjacent side length change amount, and use the rheological correction model to reversely correct the lengths of the two adjacent sides to obtain the corrected side lengths; Based on the corrected side lengths and the reference interior angle data, with P1 as the coordinate origin, calculate the spatial coordinates of P2, P3, and P4 vertex by vertex; use the least squares method to perform curvature fitting on the quadrilateral section and update the curvature radius parameters; Integrate the corrected side lengths, updated interior angles, and curvature radius parameters to generate a dynamic topological structure containing geometric parameters and curvature information.

7. A method for monitoring road slope data under complex geological conditions according to claim 6, characterized in that, Construct a three-dimensional deformation index matrix including the diagonal displacement difference, the adjacent side curvature ratio, and the interior angle coefficient of variation, including: Extract the current lengths of the two diagonals of the quadrilateral from the dynamic topology and compare them with the initial diagonal lengths in the reference topology; calculate the absolute displacement difference and the rate of change of the displacement difference for each diagonal to generate a displacement difference feature vector for the two diagonals; Based on the updated cross-sectional curvature parameters in the dynamic topology, obtain the curvature radii and curvature ratios of adjacent sides, normalize the curvature ratios, and map the ratios to the 0-1 interval to generate a curvature ratio feature vector; Read the current angular values of the four interior angles of the quadrilateral from the dynamic topology, calculate the standard deviation of the interior angles and the coefficient of variation of the interior angles, and perform a sliding window statistic on the coefficient of variation to generate a smoothed sequence of the coefficient of variation; Align the diagonal displacement difference feature vector, the curvature ratio feature vector, and the sequence of the coefficient of variation of the interior angles according to the timestamp; construct a three-dimensional deformation index matrix; Perform data standardization processing on the three-dimensional deformation index matrix to generate a standardized deformation index matrix.

8. The method for monitoring road slope data under complex geological conditions according to claim 7, characterized in that: The dimension of the three-dimensional deformation index matrix is [timestamp × deformation feature], where each row contains: the first column: the rate of change of the displacement difference of the two diagonals; the second column: the normalized value of the curvature ratio of adjacent sides; the third column: the coefficient of variation of the interior angles.

9. A method for monitoring road slope data under complex geological conditions according to claim 8, characterized in that, Input the index matrix, the strain spatio-temporal matrix of the support structure, and the rock mass fragmentation tensor into the spatio-temporal attention graph neural network, and output the deformation evolution trends including the diagonal displacement difference, the curvature ratio of adjacent sides, and the coefficient of variation of the interior angles, including: Align the standardized deformation index matrix, the strain spatio-temporal matrix of the support structure constructed from the strain time series data collected by the fiber optic sensor, and the rock mass fragmentation tensor calculated based on the microseismic event energy distribution and the rock mass fracture density according to the timestamp to obtain the time-aligned data; Based on the time-aligned data, perform dimensionality reduction processing on the rock mass fragmentation tensor, extract its principal component feature vector, and make its column dimension match that of the deformation index matrix; Using the principal component feature vector, together with the deformation index matrix and the strain spatio-temporal matrix of the support structure, splice the three types of data features along the time axis to construct a fused spatio-temporal tensor; According to the fused spatio-temporal tensor, use the multi-head attention mechanism to calculate the spatio-temporal dependence relationships between the three types of data respectively; According to the spatio-temporal dependence relationships, multiply the fused spatio-temporal tensor by the attention weight matrix, and aggregate the spatial correlation features of the vertices of the quadrilateral and their support nodes through the graph convolutional network; Based on the spatial correlation features, use the gated recurrent unit to learn the long-term evolution trend of the deformation index; Input the long-term evolution trend into the fully connected prediction module, and output the deformation evolution trend parameters within the next 3 hours, including: the increment of the rate of change of the diagonal displacement difference, the change gradient of the curvature ratio of adjacent sides, and the acceleration of the coefficient of variation of the interior angles.

10. A method for monitoring road slope data under complex geological conditions according to claim 9, characterized in that, Execute a multi-level early warning decision according to the component parameters of the deformation evolution trend, including: Extract the parameters of the increment of the rate of change of the diagonal displacement difference, the change gradient of the curvature ratio of adjacent sides, and the acceleration of the coefficient of variation of the interior angles in the deformation evolution trend; based on the statistical correlation analysis of the historical deformation data and the support structure failure cases, set the grading threshold intervals for the three parameters, including the yellow warning threshold interval, the orange warning threshold interval, and the red warning threshold interval; When any parameter first enters the yellow warning threshold range, increase the three-dimensional laser ranging frequency of the corresponding monitored edge; associate the warning information with the coordinates of high-risk nodes in the reference topology to generate a primary warning signal with spatial positioning. When two or more parameters simultaneously enter the orange warning threshold range, match the pre-stored support reinforcement strategy library according to the parameter combination mode, dynamically calculate the required number of bolt reinforcement and grouting pressure values, and generate a support parameter adjustment instruction set. When all three parameters reach the red warning threshold range and the acceleration of the inner angle variation coefficient exceeds the critical value, send an evacuation instruction to the on-site personnel terminal, automatically cut off the power supply of the tunnel face construction equipment, and activate the rapid grouting and hydraulic support units of the emergency support system. Dynamically update the topology structure database according to the warning level, record the warning trigger time, parameter overrun value and disposal measures to the blockchain evidence storage node, and form a warning decision chain.

Images