Multi-parameter fusion dangerous rock mass instability precursor identification method

Abstract

The invention discloses a multi-parameter fusion dangerous rock mass instability precursor identification method, which comprises the following steps: S1, generating a thermodynamic diagram of dangerous rock mass crack distribution, dividing each sub-region and setting each risk level; constructing a monitoring network; s2, performing time alignment and space alignment on the monitoring data; s3, the GNN model calculates and outputs each instability probability of each monitoring node; and S4, generating a global risk thermodynamic diagram according to each instability probability, and performing graded early warning based on a preset probability threshold. According to the method, the unmanned aerial vehicle thermodynamic diagram is used for guiding sensor optimization deployment, multi-dimensional data such as deformation, osmotic pressure, COs and infrasonic waves are fused, the model for identifying the instability of the dangerous rock mass is constructed by using the GNN model, accurate calculation and graded early warning of the instability probability are achieved, and the identification accuracy and early warning real-time performance are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of rock mass engineering technology, specifically to a method for identifying early signs of unstable rock masses through multi-parameter fusion. Background Technology

[0002] Dangerous rock masses refer to rock formations located on steep slopes, cliffs, canyons, or artificially excavated slopes that have already become unstable or are potentially prone to collapse or landslides due to factors such as geological structure, weathering, earthquakes, rainfall, freeze-thaw cycles, or human engineering activities. These rock masses pose a high risk of geological disasters; once unstable, they could cause serious consequences such as casualties, property damage, and traffic disruption. Therefore, identifying dangerous rock masses and predicting their instability risk has become a crucial aspect of disaster prevention and mitigation work in mountainous areas.

[0003] Current monitoring of unstable rock masses primarily relies on single or limited types of sensors, such as GNSS positioning equipment for displacement monitoring, 3D laser scanners for surface deformation analysis, and acoustic emission sensors for rock mass fracture monitoring. These technologies typically operate independently and lack multi-parameter collaborative analysis capabilities. Traditional methods only monitor deformation parameters, neglecting key precursor parameters such as seepage pressure, internal rock mass gas concentration, and temperature changes. Existing patents mainly rely on deformation threshold warnings, failing to integrate multimodal data, resulting in incomplete identification of precursor information. Furthermore, existing systems often use fixed threshold alarms, and the warning models lack adaptive capabilities, leading to high false alarm and false negative rates.

[0004] The existing technology has the following limitations: 1) Insufficient multimodal fusion and one-sided data, mainly relying on single-type parameters, which cannot reflect the multi-dimensional precursor characteristics of unstable rock masses. Although some solutions have attempted to deploy multiple sensors, they lack spatiotemporal alignment algorithms, resulting in data from different sampling frequencies not being effectively correlated; 2) Low deployment efficiency and easy to miss detections. Using manual deployment of sensors in the corresponding areas relies too much on personal deployment experience, which is time-consuming and labor-intensive. The deployed sensors are also difficult to correlate with the distribution of rock mass cracks, resulting in insufficient monitoring density in key areas and a high rate of missed detections; 3) Timeliness and cost issues. Traditional monitoring equipment has a short battery life (usually <6 months), and high-precision sensors are expensive, making large-scale deployment difficult. Summary of the Invention

[0005] To address the aforementioned problems, the purpose of this invention is to propose a multi-parameter fusion method for identifying precursors of unstable rock masses. This method guides the optimized deployment of sensors using UAV thermal maps, integrates multi-dimensional data such as deformation, seepage pressure, CO2, and infrasound, and utilizes a GNN model to construct a model for identifying unstable rock masses. This enables accurate calculation of instability probability and graded early warning, significantly improving the accuracy of identification and the real-time nature of early warning.

[0006] This was achieved through the following technical solutions: A multi-parameter fusion method for identifying precursors of unstable rock masses includes the following steps: S1. Conduct drone aerial photography and 3D reconstruction of the target area of ​​the unstable rock mass to generate a heat map of the distribution of cracks in the unstable rock mass. Divide the area into sub-regions based on the heat map and set corresponding risk levels. Based on each sub-region and each risk level, set each monitoring node to construct a monitoring network. S2. Collect monitoring data from the monitoring network, perform time and spatial alignment on the monitoring data, and form fused data with a unified spatiotemporal reference; S3. Input the fused data into the GNN model, construct a graph structure with each monitoring node as a vertex and the spatial distance and parameter correlation between nodes as edges, and calculate and output the instability probability of each monitoring node. S4. Generate a global risk heat map based on each instability probability, and provide graded early warnings based on preset probability thresholds.

[0007] Preferably, each risk level in step S1 includes low risk and high risk, and each monitoring node is equipped with a corresponding monitoring combination, and each monitoring combination includes a corresponding sensor; each monitoring node set in the sub-regions corresponding to low risk and high risk are respectively a basic node and a composite node, and the deployment density of the monitoring combination in the composite node is N times the deployment density of the monitoring combination in the basic node, where N≥3.

[0008] Preferably, when setting each monitoring node in step S1, each geological type of each sub-region is first determined, and each corresponding monitoring combination is selected according to each geological type. The geological types include loose deposits, bedrock weathering layers, and intact bedrock. The sensors in the monitoring combination corresponding to loose deposits include crack gauges, GNSS, and rain gauges. The sensors in the monitoring combination corresponding to bedrock weathering layers include at least strain gauges and GNSS. The sensors in the monitoring combination corresponding to intact bedrock include at least GNSS.

[0009] Preferably, the sensors in the monitoring assembly corresponding to the bedrock weathering layer also include an infrasound detector, and the sensors in the monitoring assembly corresponding to the intact bedrock also include a CO2 detector or a water pressure detector.

[0010] Preferably, during time alignment in step S2, the data collected by any sensor within time [t0, t1] is denoted as y(t). A sliding window interpolation method is used, with a target frequency of 1 minute and a 5-minute sliding window for cubic spline interpolation of y(t). The interpolation function used is S(t) = a i t 3 +b i t 2 +c i t+d i, where S(t) i )=y(t i And the derivative S'(t) i (The numbers are continuous, and i is a positive integer.)

[0011] Preferably, during spatial alignment in step S2, the raw coordinate data collected by any sensor are converted into corresponding planar coordinate data, and then the planar coordinate data are corrected for errors using the trilateration method based on UWB calibration to obtain the corrected spatial coordinates.

[0012] Preferably, after inputting the fused data into the GNN model in step S3, the fused data is divided into temporal features, environmental features, and spatial features with each sensor as a vertex. The lag time corresponding to the environmental features is calculated by CCF standardization, the spatial features are weighted by the GNN adjacency matrix, and the trend term, seasonal term, and residual term are decomposed from the temporal features by STL.

[0013] Preferably, the calculation formula used in CCF standardization is: τ represents the lag days, taken as... The τ at its maximum is taken as the corresponding lag time.

[0014] Preferably, when weighting spatial features using the GNN adjacency matrix, the adjacency weights are set to... , α represents the weighting percentage of spatial distance. d ij Let be the Euclidean distance between nodes i and j, where i and j are both positive integers. d max This represents the maximum node spacing in the subregions where nodes i and j are located. r ij Let Pearson correlation coefficient be the displacement sequence of nodes i and j, and >0.3.

[0015] Preferably, after weighting the spatial features using the GNN adjacency matrix, the formula is first followed. Perform spatial feature extraction. For the normalized adjacency matrix, D For degree matrix, Learnable weights; then, an update gate is set. Reset door Candidate hidden state Hidden state Used for capturing temporal features, x t The time feature at time t; z t =1 indicates that past information is completely preserved, z t =0 indicates that past information is completely discarded (1-z) tThis indicates the proportion of past information retained; then the attention weights are calculated. Obtain weighted features As the corresponding probability of instability and Let represent the i-th and j-th features, m be the total number of features, w be the attention vector, and W be the feature transformation matrix.

[0016] The beneficial effects of this invention compared to the prior art are: The technical solution of this invention guides the optimized deployment of heterogeneous sensors such as deformation, osmotic pressure, CO2, and infrasound through UAV thermal mapping. It collects multi-dimensional data and uses a spatiotemporal alignment algorithm to fuse multi-source heterogeneous data. The data is then input into a model with a graph neural network as its core. By analyzing the spatial and parameter correlations between sensor nodes, the instability probability is calculated in real time, thereby enabling accurate graded early warning. Attached Figure Description

[0017] Figure 1 This is a flowchart of a multi-parameter fusion method for identifying precursors to rock mass instability. Detailed Implementation

[0018] The following will be combined with the present invention. Figure 1 The technical solutions in the embodiments of the present invention will be described in detail below.

[0019] like Figure 1 As shown, a flowchart of a multi-parameter fusion method for identifying precursors of unstable rock masses is presented. The method guides the optimized deployment of multiple sensor combinations through UAV thermal maps, integrates multi-dimensional data such as deformation, seepage pressure, CO2, and infrasound, and uses a GNN model to construct a model for identifying unstable rock masses. This enables accurate calculation of instability probability and graded early warning, significantly improving the identification accuracy and real-time warning performance.

[0020] The method specifically includes the following steps: S1. Conduct drone aerial photography and 3D reconstruction (e.g., SFM 3D reconstruction) of the target area of ​​the unstable rock mass to generate a heat map of the rock mass crack distribution. Divide the area into sub-regions based on the heat map and set corresponding risk levels. Then, based on each sub-region and each risk level, set up each monitoring node to construct a monitoring network.

[0021] Each risk level includes low risk and high risk (based on the density of cracks in the heat map), and each monitoring node is equipped with a corresponding monitoring combination, which includes each corresponding sensor.

[0022] When setting up each monitoring node, it is necessary to first determine each geological type in each sub-region, and then select the corresponding monitoring combination for each geological type. Referring to Table 1 below, geological types include loose deposits, weathered bedrock, and intact bedrock. The determination is based on the baseline density ρ of crack distribution. The sensors in the monitoring combination corresponding to loose deposits include crack gauges, GNSS, and rain gauges. The sensors in the monitoring combination corresponding to weathered bedrock include strain gauges, GNSS (and optional infrasound detectors). The sensors in the monitoring combination corresponding to intact bedrock include GNSS (and optional CO2 detectors or water pressure detectors).

[0023] Table 1:

[0024] The data parameters collected by various sensors are shown in Table 2 below: Table 2:

[0025] Each monitoring node set up in the corresponding sub-regions of low risk and high risk is a basic node and a composite node, respectively. The deployment density of the monitoring combination in the composite node is N times that in the basic node, where N≥3. The deployment density corresponds to the number of sensors used, and the distance between two adjacent sensors in any sub-region is ≤5m to reduce data errors.

[0026] S2. Collect monitoring data from the monitoring network, including heat maps and point maps of each sensor (coordinates, ID, and sensor type parameters); perform time and spatial alignment on the monitoring data to form fused data with a unified spatiotemporal reference.

[0027] For time alignment, a sliding window interpolation method is used. Data collected by any sensor within the time interval [t0, t1] is denoted as y(t). A target frequency of 1 minute is set, and cubic spline interpolation is performed on y(t) using a 5-minute sliding window. The interpolation function used is S(t) = a i t 3 +b i t 2 +c i t+d i Among them, S(t) needs to be satisfied. i )=y(t i And the derivative S'(t) i (The sequence is continuous, and i is a positive integer.) Then, the data after interpolation can be aligned to verify whether the time series is consistent. If it is consistent, spatial alignment can continue.

[0028] Spatial alignment is based on GPS-UWB fusion positioning for coordinate calibration. It converts raw coordinate data collected by any sensor into corresponding planar coordinate data, and then, based on UWB calibration, uses trilateration to correct errors in the planar coordinate data, obtaining the corrected spatial coordinates. For example, converting the raw GPS coordinates (B, L, H) of any sensor into planar coordinates (X, Y): , , where a is the semi-major axis of the Earth's reference ellipsoid, e 2 These are the ellipsoidal parameters.

[0029] Then, using three UWB base stations deployed in the target area, each sensor is treated as a node, and the distance between each node and the base station is measured. The GPS error is corrected using trilateration, and the corrected coordinates (X', Y', Z') are calculated to reduce spatial errors (e.g., within 5cm).

[0030] S3. Input the fused data into the GNN model, construct a graph structure with each monitoring node as a vertex and the spatial distance and parameter correlation between nodes as edges, and calculate and output the instability probability of each monitoring node.

[0031] After inputting the fused data into the GNN model, as shown in Table 3 below, the fused data is divided into temporal features, environmental features, and spatial features, with each sensor as a vertex (node). The lag time corresponding to the environmental features is calculated by CCF standardization, the spatial features are weighted by the GNN adjacency matrix, and the trend term, seasonal term, and residual term are decomposed from the temporal features by STL (R of the residual term is the random error / abnormal fluctuation after removing the trend and period).

[0032] Table 3:

[0033] The calculation formula used in CCF standardization is: τ represents the lag days (within 0-15 days), taking... The τ at the maximum value is taken as the corresponding lag time, and the CCF-standardized R reflects the lag correlation strength between environmental factors and displacement. This represents the time series mean of displacement data. This represents the time series mean of environmental factors.

[0034] Weighting spatial features using a GNN adjacency matrix is ​​intended to balance spatial distance and data relevance. The adjacency weights can be set as follows: , α is the spatial distance weighting percentage (which can be set to 0.6). d ijLet be the Euclidean distance between nodes i and j, where i and j are both positive integers. d max This represents the maximum node spacing in the subregions where nodes i and j are located. r ij Let Pearson correlation coefficient be the displacement sequence of nodes i and j, and >0.3 (avoid invalid edges).

[0035] After weighting the spatial features using the GNN adjacency matrix, first follow the formula... Perform spatial feature extraction. For the normalized adjacency matrix, A For nodes, ReLU For activation function, The output feature matrix of layer l+1, D For degree matrix, The learnable weights of the l-th layer, This is the bias vector for the l-th layer (training parameters used to correct feature offsets). Then from... This involves capturing the dynamic changes of time series data to address the long-short dependency issue, including setting update gates. Reset door Candidate hidden state (Final) Hidden State Used for capturing time features, where t is time t, ⊙ represents the element-wise dot product, and x t The time feature at time t; z t =1 indicates that past information is completely preserved, z t =0 indicates that past information is completely discarded (1-z) t This represents the proportion of past information retained. Then, the attention weights are calculated. Obtain weighted features As the corresponding probability of instability; and Representing the i-th and j-th features, T denoted as the total number of time steps, m as the total number of features, w as the attention vector, and W as the feature transformation matrix.

[0036] S4. Generate a global risk heatmap based on each instability probability, and perform graded early warnings based on preset probability thresholds. Output both the global risk heatmap and the graded early warning results as the final output. The output also includes the ID of each node, the displacement rate of the unstable rock mass in each sub-region, and the suggested measures corresponding to the graded early warning. When generating the global risk heatmap, different colors can be used to display different risk values. For example, blue, yellow, and red warnings can be used according to the instability probability from smallest to largest.

[0037] Furthermore, for unstable rock masses in any sub-region, 1000 bootstrap samplings can be performed to calculate a 95% confidence interval. At the same time, the warning levels are classified according to Table 4 below, and corresponding suggestions are given to deal with various different situations: Table 4:

[0038] In summary, this invention guides the optimized deployment of heterogeneous sensors such as deformation, osmotic pressure, CO2, and infrasound through UAV thermal mapping, collects multi-dimensional data, and uses a spatiotemporal alignment algorithm to fuse multi-source heterogeneous data. The data is then input into a model based on a graph neural network. By analyzing the spatial and parameter correlations of sensor nodes over time, the instability probability is calculated in real time, thereby enabling accurate graded early warning. This invention represents a significant advancement.

[0039] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A method for identifying precursors of unstable rock mass through multi-parameter fusion, characterized in that, Includes the following steps: S1. Conduct drone aerial photography and 3D reconstruction of the target area of ​​the unstable rock mass to generate a heat map of the distribution of cracks in the unstable rock mass. Divide the area into sub-regions based on the heat map and set corresponding risk levels. Based on each sub-region and each risk level, set each monitoring node to construct a monitoring network. S2. Collect monitoring data from the monitoring network, perform time and spatial alignment on the monitoring data, and form fused data with a unified spatiotemporal reference; S3. Input the fused data into the GNN model, construct a graph structure with each monitoring node as a vertex and the spatial distance and parameter correlation between nodes as edges, and calculate and output the instability probability of each monitoring node. S4. Generate a global risk heat map based on each instability probability, and provide graded early warnings based on preset probability thresholds.

2. The method for identifying precursors of unstable rock mass based on multi-parameter fusion according to claim 1, characterized in that, Each risk level in step S1 includes low risk and high risk. Each monitoring node is equipped with a corresponding monitoring combination, and each monitoring combination includes a corresponding sensor. The monitoring nodes set in the sub-regions corresponding to low risk and high risk are respectively basic nodes and composite nodes. The deployment density of the monitoring combination in the composite node is N times that of the deployment density of the monitoring combination in the basic node, where N≥3.

3. The method for identifying precursors of unstable rock mass based on multi-parameter fusion according to claim 2, characterized in that, When setting up each monitoring node in step S1, first determine each geological type of each sub-region, and select the corresponding monitoring combination for each geological type. Geological types include loose deposits, weathered bedrock, and intact bedrock. The sensors in the monitoring combination corresponding to loose deposits include crack gauges, GNSS, and rain gauges. The sensors in the monitoring combination corresponding to weathered bedrock include at least strain gauges and GNSS. The sensors in the monitoring combination corresponding to intact bedrock include at least GNSS.

4. The method for identifying precursors of unstable rock mass based on multi-parameter fusion according to claim 3, characterized in that, The sensors in the monitoring system corresponding to the weathered bedrock layer also include an infrasound detector, while the sensors in the monitoring system corresponding to intact bedrock also include a CO2 detector or a water pressure detector.

5. The method for identifying precursors of unstable rock mass based on multi-parameter fusion according to claim 2, characterized in that, In step S2, during time alignment, the data collected by any sensor within the time interval [t0, t1] is denoted as y(t). A sliding window interpolation method is used, with a target frequency of 1 minute and a 5-minute sliding window for cubic spline interpolation of y(t). The interpolation function used is S(t) = a i t 3 +b i t 2 +c i t+d i , where S(t) i )=y(t i And the derivative S'(t) i (The numbers are continuous, and i is a positive integer.) 6. The method for identifying precursors of unstable rock mass based on multi-parameter fusion according to claim 2, characterized in that, In step S2, during spatial alignment, the raw coordinate data collected by any sensor is converted into corresponding planar coordinate data. Then, based on UWB calibration, the planar coordinate data is corrected for errors using the trilateration method to obtain the corrected spatial coordinates.

7. The method for identifying precursors of unstable rock mass based on multi-parameter fusion according to claim 2, characterized in that, Step S3 involves inputting the fused data into the GNN model, using each sensor as a vertex, and dividing the fused data into temporal features, environmental features, and spatial features. The lag time corresponding to the environmental features is calculated using CCF standardization, the spatial features are weighted using the GNN adjacency matrix, and the trend term, seasonal term, and residual term are decomposed from the temporal features using STL.

8. The method for identifying precursors of unstable rock mass based on multi-parameter fusion according to claim 7, characterized in that, The calculation formula used in CCF standardization is: τ represents the lag days, taken as... The τ at its maximum is taken as the corresponding lag time. This represents the time series mean of displacement data. This represents the time series mean of environmental factors.

9. The method for identifying precursors of unstable rock mass based on multi-parameter fusion according to claim 7, characterized in that, When weighting spatial features using the GNN adjacency matrix, the adjacency weights are set to... , α represents the weighting percentage of spatial distance. d ij Let be the Euclidean distance between nodes i and j, where i and j are both positive integers. d max This represents the maximum node spacing in the subregions where nodes i and j are located. r ij Let Pearson correlation coefficient be the displacement sequence of nodes i and j, and >0.

3.

10. The method for identifying precursors of unstable rock mass based on multi-parameter fusion according to claim 9, characterized in that, After weighting the spatial features using the GNN adjacency matrix, first follow the formula... Perform spatial feature extraction. For the normalized adjacency matrix, A For nodes, ReLU For activation function, The output feature matrix of layer l+1; then the update gate is set. Reset door Candidate hidden state Hidden state Used for capturing temporal features, x t The time feature at time t; z t =1 indicates that past information is completely preserved, z t =0 indicates that past information is completely discarded (1-z) t This indicates the proportion of past information retained; then the attention weights are calculated. Obtain weighted features As the corresponding probability of instability and Let represent the i-th and j-th features, m be the total number of features, w be the attention vector, and W be the feature transformation matrix.

Images