Geological environment monitoring method and system

Abstract

The invention provides a geological environment monitoring method and system, and the method comprises the steps: constructing an air-space-earth-depth four-dimensional cooperative monitoring network, collecting multi-modal geological environment data, carrying out the preprocessing of the data, and obtaining multi-modal feature data and abnormal data points after dynamic load balance distribution of calculation resources; fusing the data to generate a comprehensive geological environment feature map containing space-time correlation features, abnormal hot spot distribution and a geologic body three-dimensional reconstruction model; and based on the map, performing geological risk prediction by using a geological risk prediction model to obtain a prediction probability. A monitoring network is constructed to integrate multi-modal geological data, resource allocation is optimized by combining edge calculation dynamic load balancing, geological risk prediction is realized by using a prediction model, the problems of single data, poor dynamic adaptability and weak model generalization ability of a traditional method are solved, the prediction precision is improved, the response time is shortened, and the prediction efficiency is improved. And a high-risk area is visually displayed through a three-dimensional risk thermodynamic diagram.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of environmental monitoring, in particular to a geological environment monitoring method and system. BACKGROUND

[0002] Current geological environment monitoring technology mainly relies on a single type of data source or local monitoring means, which is difficult to form full-dimensional perception ability, resulting in low prediction accuracy. For example, traditional methods focus on surface displacement monitoring (such as GNSS or synthetic aperture radar interferometry) or shallow geological signals (such as underground water level and temperature sensors), and lack effective capture means for deep geostress changes, high-precision three-dimensional structures of unmanned aerial vehicles, and multi-physical field collaborative evolution characteristics. In mountain landslide monitoring, relying only on surface displacement data can easily overlook early signals of deep fault stress accumulation; in mine area pressure monitoring, shallow sensor data cannot reflect the complex mechanical process of deep rock layer rupture. In addition, satellite remote sensing, unmanned aerial vehicle patrol, ground sensors and deep exploration systems usually operate independently, and data fusion is limited to simple superposition, and a time and space related multi-modal feature expression system has not been established. SUMMARY

[0003] The present application aims to at least solve the technical problems of low prediction accuracy in the prior art, and particularly innovatively provides a geological environment monitoring method and system.

[0004] In order to achieve the above-mentioned purpose of the present application, the present application provides a geological environment monitoring method, the method comprising: S1, constructing an air-space-ground-deep four-dimensional collaborative monitoring network to collect multi-modal geological environment data, the multi-modal geological environment data comprising surface image data, unmanned aerial vehicle laser point cloud data, ground sensor array signal and deep geostress measurement data; S2, preprocessing the multi-modal geological environment data to obtain multi-modal feature data and abnormal data points after dynamic load balancing distribution of computing resources; S3, multi-modal data fusion of the multi-modal feature data and the abnormal data points to generate a comprehensive geological environment feature map, the comprehensive geological environment feature map comprising time and space related features, abnormal hotspot distribution and a three-dimensional reconstruction model of geological bodies; S4, geological risk prediction based on the comprehensive geological environment feature map using a geological risk prediction model to obtain a geological risk prediction probability.

[0005] In another aspect, the present application also provides a geological environment monitoring system, the system being used for performing a geological environment monitoring method; the system comprising: A data acquisition module is configured to construct an aerospace-terrestrial-deep four-dimensional cooperative monitoring network to acquire multi-modal geological environment data, and the multi-modal geological environment data includes surface image data, unmanned aerial vehicle laser point cloud data, ground sensor array signals and deep geostress measurement data. A preprocessing module is connected with the acquisition module and is configured to pre-process the multi-modal geological environment data to obtain multi-modal feature data and abnormal data points after dynamic load balancing distribution of computing resources. A data fusion module is connected with the preprocessing module and is configured to perform multi-modal data fusion on the multi-modal feature data and the abnormal data points to generate a comprehensive geological environment feature map, and the comprehensive geological environment feature map includes a spatio-temporal correlation feature, an abnormal hotspot distribution and a geological body three-dimensional reconstruction model. A prediction module is connected with the data fusion module and is configured to perform geological risk prediction based on the comprehensive geological environment feature map by using a geological risk prediction model to obtain a geological risk prediction probability.

[0006] The present application has the following advantages: the present application integrates multi-modal geological data (surface image, unmanned aerial vehicle point cloud, ground sensor signal and deep geostress) by constructing an aerospace-terrestrial-deep four-dimensional cooperative monitoring network, optimizes resource distribution by combining edge computing dynamic load balancing, and realizes geological risk prediction by using a geological risk prediction model, effectively solving the problems of single data, poor dynamic adaptability and weak model generalization ability of traditional methods, significantly improving the prediction accuracy, shortening the system response time, and intuitively displaying the high-risk area through the three-dimensional risk heat map.

[0007] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0008] The above and / or additional aspects and advantages of the present application will become apparent and be readily understood from the following description, taken in conjunction with the accompanying drawings, in which: Figure 1 is a flowchart of a geological environment monitoring method of the present application. DETAILED DESCRIPTION

[0009] The embodiments of the present application will be described in detail below, and examples of the embodiments are shown in the accompanying drawings, in which the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as limiting the present application.

[0010] Example 1 As shown in Figure 1 , a geological environment monitoring method, the method comprising: S1, a space-time-ground-deep four-dimensional cooperative monitoring network is constructed to collect multi-modal geological environment data, and the multi-modal geological environment data includes surface image data, unmanned aerial vehicle laser point cloud data, ground sensor array signal and deep geostress measurement data; S2, the multi-modal geological environment data is preprocessed to obtain multi-modal feature data and abnormal data points after dynamic load balancing distribution of computing resources; S3, the multi-modal feature data and the abnormal data points are subjected to multi-modal data fusion to generate a comprehensive geological environment feature map, and the comprehensive geological environment feature map includes a space-time correlation feature, an abnormal hotspot distribution and a geological body three-dimensional reconstruction model; S4, based on the comprehensive geological environment feature map, a geological risk prediction model is used to predict the geological risk to obtain a geological risk prediction probability.

[0011] The principle of the geological environment monitoring method in this embodiment is that multi-modal data (including surface images, unmanned aerial vehicle point clouds, ground sensor signals and deep geostress measurement data) collected by the integrated space-time-ground-deep four-dimensional cooperative monitoring network is used to optimize the allocation of computing resources by using dynamic load balancing technology to ensure the efficiency of data preprocessing. Specifically, the method first performs noise reduction, calibration and standardization processing on the original data, extracts a dynamic trend feature set and abnormal data points; then, the Fourier transform is used to analyze the periodic pattern, and the space-time correlation coefficient of the surface displacement acceleration and the geomagnetic anomaly is calculated by combining the cross-correlation function to generate a space-time correlation feature of the displacement-pressure-geomagnetic cooperative change. On this basis, the abnormal data points are mapped to a space-time cube model, and the abnormal grid units are identified and clustered by using a comprehensive anomaly index and a DBSCAN density clustering algorithm to form an abnormal hotspot distribution of potential geological disaster risks. Subsequently, a multi-modal data fusion matrix is constructed, the dynamic trend features, the periodic pattern and the geological structure parameters are integrated, and a hybrid inference model (including a physical mechanism driven layer, a data driven layer and a dynamic correction layer) is input to calculate an initial risk probability. The physical mechanism driven layer evaluates the fault stability and the geostress accumulation based on the principles of geomechanics; the data driven layer uses a random forest algorithm to establish a nonlinear mapping between the features and historical disaster events; the dynamic correction layer adjusts the prediction results in real time by using a long short-term memory network, and combines the latest monitoring data (such as the abnormal cluster expansion speed, the geostress mutation amplitude and the displacement acceleration change rate) to correct the probability value, and finally generates a spatially continuous risk heat map to intuitively mark the high-risk areas. The core of this principle is to integrate the physical model and the data driven method to improve the prediction accuracy and response speed, and at the same time, to enhance the operability of the risk warning through three-dimensional visualization.

[0012] As an optional embodiment of the present application, optionally, the space-time-ground-deep four-dimensional cooperative monitoring network is constructed in step S1, and the space-time-ground-deep four-dimensional cooperative monitoring network includes: S101, deploy a satellite remote sensing system to collect ground image data, the ground image data including high-resolution optical images and synthetic aperture radar images; It needs to be explained in step S101 that in the present embodiment, the satellite remote sensing system is a high-resolution optical satellite and a synthetic aperture radar satellite. By collecting ground image data periodically or in real time, macroscopic monitoring of the geological environment is realized. High-resolution optical images can capture detailed information such as surface morphology, vegetation coverage and land use change, while synthetic aperture radar images are not limited by weather and light conditions, can penetrate clouds, and provide all-weather ground monitoring data.

[0013] S102, construct a UAV platform and carry a laser radar sensor to collect UAV laser point cloud data, and use the UAV laser point cloud data to construct a three-dimensional geological body model; It needs to be explained in step S102 that in the present embodiment, the UAV platform is equipped with a high-precision laser radar (LiDAR) sensor. By presetting a flight route (flight altitude ≤ 300 meters, lateral overlap rate ≥ 60%, heading overlap rate ≥ 80%) to scan the target area, dense point cloud data with centimeter-level spatial resolution is obtained. The point cloud data contains accurate three-dimensional coordinate information and reflection intensity value of the surface of the ground object. Then, the original point cloud is processed by point cloud processing algorithms (such as irregular triangle network TIN-based ground reconstruction algorithm or Poisson surface reconstruction algorithm) to remove noise, filter and classify the ground points and non-ground points (such as vegetation, buildings), and finally based on the ground point cloud data, a high-precision digital surface model (DSM) and a digital elevation model (DEM) are constructed, and geological exploration data (such as drilling and profile information) are fused for constrained interpolation to generate a three-dimensional geological body model reflecting the surface and shallow geological structure.

[0014] S103, install a ground sensor array to collect geological signals, the ground sensor array including displacement sensors, temperature sensors and humidity sensors; It needs to be explained in step S103 that in this embodiment, the installation of the ground sensor array includes the distribution of distributed sensor nodes at key geological monitoring points (such as fault zones, landslide front edges, and groundwater seepage areas), each node integrating a displacement sensor (using a high-precision GNSS receiver or a fiber optic grating displacement meter, with an accuracy of ±0.1 mm), a temperature sensor (using a platinum resistance PT100 or a digital temperature sensor DS18B20, with a range of -20℃ to 80℃), and a humidity sensor (using a frequency domain reflection FDR or time domain reflection TDR probe, with a range of 0% to 100% saturation). The deployment interval of the sensor array is dynamically optimized according to the complexity of the geological structure (for example, 50 meters apart in uniform areas and 10 meters densely distributed in high-risk areas), and the original signals are transmitted in real time to the edge gateway through a low-power wide-area network LPWAN (such as LoRa protocol or NB-IoT). The collection frequency is adaptively adjusted according to the state of geological activity: sampling every minute in the stable period, and increasing to continuous monitoring every second in the active period. At the same time, the original data are real-time denoised and drift corrected by combining the existing Kalman filter algorithm, to ensure the signal stability and reliability. The sensor node is also equipped with a self-powered module (solar panel + lithium battery pack) to support long-term operation in harsh outdoor environments, and a redundant design (such as dual-link communication) is used to enhance the system robustness and avoid data interruption caused by single-point failure. The collected geological signals (including displacement vector, temperature and humidity gradient, and strain change) are used to calculate derived parameters such as surface displacement acceleration and soil water content rate.

[0015] S104, collecting deep geostress measurement data by installing deep geostress measurement modules through boreholes; It needs to be explained in step S104 that in this embodiment, the deep geostress measurement module uses a piezomagnetic or strain geostress sensor, which is installed vertically through a borehole to the target rock layer (depth ≥ 50 meters), and the sensor probe is tightly coupled to the hole wall rock mass using a hydraulic anchor. The sensor monitors the principal stress components (range 0-60MPa, accuracy ±0.5%FS) in three orthogonal directions (axial, tangential, and radial) of the rock mass at a sampling frequency of ≥1Hz. After the original voltage signal is compensated by the temperature compensation circuit to eliminate the influence of thermal drift, it is converted into the effective stress value. The collected stress data are transmitted to the ground data acquisition station through an anti-interference shielded cable, and an adaptive Kalman filter is used to eliminate mechanical vibration noise, finally generating a deep geostress measurement data set containing stress amplitude, direction, and change rate. The system is provided with a dual-channel redundant verification mechanism, which automatically switches to the backup module when the main sensor fails, to ensure data continuity and reliability.

[0016] S105, preliminary fusion of the ground image data, geological body three-dimensional model, geological signal, and deep geostress measurement data using a communication network to obtain a space-air-ground-deep four-dimensional collaborative monitoring network.

[0017] It needs to be explained in step S105 that the preliminary fusion by using the communication network in the embodiment includes: deploying a distributed edge computing gateway, and integrating heterogeneous data streams from satellite remote sensing systems, unmanned aerial vehicle platforms, ground sensor arrays and deep geostress measurement modules in real time through multi-protocol interfaces (including 5G private network, satellite link and LoRaWAN). Specifically, a time stamp synchronization mechanism (accuracy ±1ms) and a spatial coordinate registration algorithm (based on WGS84 coordinate system) are used to perform space-time alignment on all input data (surface image data, geological body three-dimensional model, geological signal and deep geostress measurement data), and eliminate transmission delay and positioning error. The data fusion process uses a weighted Kalman filtering algorithm to assign dynamic weight coefficients (for example, surface image data weight 0.3, unmanned aerial vehicle point cloud data weight 0.25, ground sensor signal weight 0.2, and deep geostress data weight 0.25) to different data sources to balance the data accuracy and real-time requirements. The fusion algorithm also includes data dimension reduction processing (using principal component analysis PCA, retaining 95% variance information) and outlier rejection (based on the 3σ principle) to generate a space-time consistent four-dimensional data set. Finally, the fusion results are cached by the redundant storage module (RAID1 configuration) of the edge gateway, and pushed to the cloud analysis platform using the MQTT protocol to ensure data integrity and low-latency transmission (end-to-end delay ≤500ms). The preliminary fusion not only optimizes the input quality of subsequent processing, but also improves the network resource utilization by compressing the transmission bandwidth (reducing data volume by 50%).

[0018] As an optional embodiment of the application, the multi-modal geological environment data is optionally pre-processed in step S2 to obtain multi-modal feature data and abnormal data points after dynamic load balancing of computing resources. S201, receiving multi-modal geological environment data of the space-air-ground-deep four-dimensional cooperative monitoring network based on the edge computing node, and classifying the multi-modal geological environment data according to data types and priorities to obtain classified multi-modal geological environment data; In step S201, it is necessary to point out that in this embodiment, the edge computing node is deployed at the key position of the monitoring area (such as the edge of the fault zone or the landslide monitoring station), and receives the heterogeneous data stream from the space-air-ground-deep four-dimensional cooperative monitoring network in real time through the multi-protocol interface (including 5G private network, LoRaWAN and satellite backhaul link). Specifically, the data are classified into four categories according to the type: ground image data, unmanned aerial vehicle laser point cloud data, ground sensor signal and deep geostress measurement data, and the processing weight is dynamically allocated based on the data source priority: the data with high real-time requirement (such as displacement sensor signal and geostress mutation data) are given the highest priority (weight coefficient ≥ 0.4), the data with medium real-time requirement (such as unmanned aerial vehicle point cloud update) are given medium priority (weight coefficient 0.3), and the data with low real-time requirement (such as historical image archiving) are given basic priority (weight coefficient ≤ 0.3). The classification process adopts the existing hash filtering algorithm based on metadata label, and automatically routes to the special buffer queue of the edge computing node according to the data timestamp, spatial coordinate and sensor ID. At the same time, the dynamic load balancing module monitors the node resource state (including CPU utilization threshold ≤ 80%, memory occupation threshold ≤ 70% and network bandwidth occupation threshold ≤ 90%) in real time, dynamically allocates the computing task to the idle node through the improved weighted least connection algorithm (WLC), and combines the resource prediction model (based on ARIMA time series analysis) to pre-allocate resources, so as to ensure that the preprocessing delay is ≤ 200 ms and avoid node overload.

[0019] S202, based on the classified multi-modal geological environment data, the dynamic load balancing algorithm is used to monitor the load of the edge computing node in real time, the task allocation is dynamically adjusted according to the data priority and node load, and the multi-modal geological environment data after dynamic load balancing allocation of computing resources is obtained; It needs to be explained in step S202 that in the present embodiment, the dynamic load balancing algorithm adopts the improved weighted least connection algorithm (WLC), in which the weight coefficient of the edge computing node is dynamically calculated based on its current resource state: CPU utilization (threshold ≤80%), memory occupation (threshold ≤70%) and network bandwidth occupation (threshold ≤90%). Specifically, the resource monitoring module collects the performance indicators of each node in real time at a sampling frequency of one second, and smoothes the noise data through the exponential weighted moving average algorithm. When the node load exceeds the threshold, the task scheduler dynamically redistributes tasks to idle nodes according to data priority (high priority data such as real-time displacement signal weight coefficient ≥0.4, medium priority data such as point cloud update weight coefficient 0.3, low priority data such as historical image weight coefficient ≤0.3). The scheduling strategy combines with the resource prediction model (based on ARIMA time series analysis, prediction period 5 minutes) to pre-allocate computing resources to avoid sudden load impact. At the same time, the task queue adopts a priority preemption mechanism, allowing high-priority tasks to interrupt low-priority task processing, ensuring real-time performance of critical data and significantly improving data processing efficiency and system stability.

[0020] S203, performing feature extraction operation on the multi-modal geologic environment data after the dynamic load balancing distributes the computing resources, to obtain multi-modal feature data; It needs to be explained in step S203 that in the present embodiment, the feature extraction operation includes multi-dimensional analysis of the multi-modal geologic environment data after the dynamic load balancing distributes the computing resources. Specifically, the Fast Fourier Transform (FFT) algorithm is used to analyze the periodic pattern of time series data and extract frequency features (such as dominant frequency and spectral energy); the sliding window statistical method (window size is adaptively adjusted, range 10 seconds to 1 hour) is used to calculate the dynamic trend feature set, including mean, variance, linear regression slope and cumulative change amount; at the same time, the Isolation Forest algorithm is applied to detect abnormal data points, identify and mark the spatio-temporal abnormal coordinates based on the feature space distance (such as Euclidean distance threshold ≥3σ). The feature extraction process integrates a parallel computing framework (such as Apache Spark or CUDA), dynamically allocates thread resources (CPU core utilization rate ≤85%) according to the data volume, and ensures that the processing delay is ≤150 milliseconds. Finally, the generated multi-modal feature data includes the dynamic trend feature set, the periodic pattern feature and the abnormal data point set.

[0021] S204, performing anomaly detection on the multi-modal feature data using the Isolation Forest algorithm, marking the geomagnetic signal mutation, pore pressure surge and surface displacement acceleration anomaly scenarios, to obtain abnormal data points.

[0022] It is to be noted that in step S204, the specific process of adopting the Isolation Forest algorithm to perform anomaly detection on the multi-modal feature data in the embodiment includes: setting the number of trees to 100, the subsample size to 256, and constructing a plurality of isolation trees by randomly selecting features and split points. The algorithm determines the degree of anomaly according to the path length required for data points to be isolated - the shorter the path length, the higher the possibility of anomaly. The abnormal score threshold is set to 0.65 (range 0-1), and when the abnormal score of a data point exceeds the threshold, the data point is marked as an abnormal data point. Specifically, the system identifies and marks the following key abnormal scenarios in real time: geomagnetic signal mutation (defined as vertical component change rate ≥ 50 nT / min), pore pressure surge (defined as pressure gradient rising rate ≥ 5 kPa / s), and surface displacement acceleration anomaly (defined as horizontal displacement acceleration ≥ 0.05 m / s²). For the marked abnormal data points, the system records their occurrence timestamp, spatial position coordinates, abnormal type, abnormal score and trigger threshold, and generates a structured abnormal event record set.

[0023] As an optional embodiment of the present application, optionally, the expression of the dynamic load balancing algorithm in step S202 is: wherein, represents the weight of the task assigned to the node , and both represent the weight adjustment coefficients of the load and the priority, represents the comprehensive load of the node , represents the priority coefficient of the task , represents the capability matching coefficient of the node to the task , represents the real-time weight of the task , represents the total number of edge computing nodes, represents the comprehensive load of the node , represents the capability matching coefficient of the node to the task ; The expression of the feature extraction operation in step S203 is: wherein, denotes multimodal feature data, denotes surface image feature, denotes laser point cloud feature, denotes sensor signal feature, denotes deep data feature, denotes vector concatenation operation, and denote weight coefficients, denotes pixel gray level, denotes distance , direction under the gray co-occurrence probability, denotes pixel point number, denotes pixel normalized vegetation index, denotes octree node number, denotes whether the octree node contains valid point cloud, if =1 contains valid point cloud, if =0 does not contain valid point cloud, denotes the depth of the octree node in the octree, denotes slope weight coefficient, denotes the gradient of the terrain height in / direction, denotes taking the maximum value, denotes time length, denotes wavelet basis function coefficient under scale , translation , denotes abnormal gradient weight, denotes the mutation value of the geomagnetic signal in time , denotes the inner product of the quantum state vector after QFT transformation, denotes the operator of quantum Fourier transform, denotes density deviation weight, denotes the number of deep exploration points, denotes the density value of the th deep exploration point, denotes the regional average density; The expression of the Isolation Forest algorithm in step S204 is: in, Indicates the first 10 multimodal feature data points Indicates the characteristics of geomagnetic signals, Indicates pore pressure characteristics, Indicates the characteristics of surface displacement. Representing multimodal feature data points Path length, Represents Euler's constant. Indicates the average path length. Indicates training samples, Representing data points In the tree Path length in Representing multimodal feature data points Abnormal scores, This represents the expected value of the path length. Indicates the abnormal threshold. This represents the 95th percentile function.

[0024] As an optional embodiment of the present invention, optionally, step S3 involves multimodal data fusion of multimodal feature data and outlier data points to generate a comprehensive geological environment feature map, including: S301. Based on a unified geographic coordinate system and timestamp, spatial registration and temporal alignment are performed on surface image features, laser point cloud features, sensor signal features and deep data features in multimodal feature data to obtain a multimodal feature set with unified spatiotemporal reference. In step S301, it should be noted that the spatial registration process employs a registration algorithm based on feature point matching (such as SIFT or ORB). First, surface image feature points (scale-invariant feature transformation points ≥ 500) are extracted and matched with laser point cloud feature points (based on normal vectors and curvature features). An affine transformation matrix is ​​calculated to achieve pixel-level accuracy (error ≤ 0.5 pixels). Temporal alignment is performed using the Dynamic Time Warping (DTW) algorithm, aligning sensor signal features (sampling frequency 1Hz to 100Hz) with deep data features (sampling interval 10 seconds to 1 minute), ensuring time window synchronization (maximum delay ≤ 50 milliseconds). Specifically, the system automatically detects and compensates for clock drift from different data sources (based on NTP protocol or GPS timestamps), generating a unified spatiotemporal reference multimodal feature set, including registered feature vectors (dimension ≤ 200) and a timestamp alignment sequence.

[0025] S302, performing spatiotemporal correlation feature extraction on the multi-modal feature set unified with the spatiotemporal reference, obtaining spatiotemporal correlation features including spatiotemporal correlation of surface displacement acceleration, pore pressure mutation rate and geomagnetic anomaly by constructing a spatiotemporal cube model and superimposing dynamic trend and periodic pattern; It is necessary to explain in step S302 that the construction of the spatiotemporal cube model adopts a grid structure of 200x200x50 (spatial resolution ≤1 meter / grid, time resolution ≤10 seconds / layer). The model input is the multi-modal feature vector after registration and alignment, and the dynamic trend and periodic pattern are superimposed by the following method: Dynamic trend superposition: for each spatial grid cell, calculate the surface displacement acceleration trend (linear fitting slope ≥0.003 m / s² / minute), pore pressure mutation rate (first derivative ≥1.2 kPa / s / layer) and geomagnetic signal change gradient (Hilbert transform instantaneous frequency ≥0.8 Hz / layer) using a sliding time window (window size=30 layers); Periodic pattern extraction: apply fast Fourier transform (FFT) to continuous 50-layer data (corresponding to ≥8 minutes duration), extract pore pressure fluctuation main frequency (0.05-0.2 Hz), geomagnetic disturbance period (1-5 minutes) and surface displacement oscillation frequency (0.01-0.1 Hz), and calculate their power spectral density (≥10 dB / Hz); Spatiotemporal correlation modeling: calculate mutual information entropy (≥0.35 bits) of adjacent grid cells (3x3 neighborhood) on consecutive 5-layer time slices, quantify the spatiotemporal coupling strength of surface displacement, pore pressure and geomagnetic anomaly. The final generated spatiotemporal correlation features include: dynamic trend coefficient of each grid cell (dimension=3), periodic main frequency and amplitude (dimension=6), neighborhood mutual information matrix (dimension=9). The system automatically removes invalid grids with feature missing rate ≥15%, outputs feature matrix with dimension ≤120,000x18.

[0026] S303, spatially aggregating the geomagnetic signal mutation, pore pressure sudden increase and surface displacement acceleration anomaly scenarios in the abnormal data points, obtaining aggregated data, identifying abnormal hot spot areas based on the aggregated data using a density clustering algorithm, and obtaining abnormal hot spot distribution marked with potential geological disaster risk; The spatial aggregation of abnormal data points in step S303 uses the existing density-based clustering algorithm (DBSCAN), sets the neighborhood radius ε=15 meters (spatial aggregation radius ≤50 meters), and the minimum sample number MinPts=8. The specific implementation includes: Geomagnetic signal mutation point aggregation: take points with mutation value ≥200nT as seed points, merge adjacent points with Euclidean distance ≤ε to form initial clusters; Pore pressure sudden increase point association: Extend the initial cluster boundary by 10 meters buffer, and include data points with pressure change rate ≥5 kPa / s; Surface displacement acceleration point fusion: Spatially superimpose points with acceleration trend ≥0.005 m / s² on existing clusters (merge when overlap area ≥40%); Abnormal hot spot generation: Mark clusters that meet the density threshold (≥15 abnormal points per 100㎡) as L1 level risk (red), and upgrade clusters that contain ≥3 types of abnormal superposition to L2 level risk (purple); Boundary optimization: Use α-shape algorithm (α=0.7) to generate non-convex polygon boundary, ensuring that the terrain fitting error is ≤2 meters. The final output includes the following elements: hot spot centroid coordinates (WGS84 coordinate system, precision 0.001"), risk level (L1 / L2), influence radius (50-200 meters, step 10 meters), multi-modal anomaly index MAI, and multi-modal anomaly index MAI obtained by weighted sum of geomagnetic mutation weight, pressure change weight, and displacement acceleration weight.

[0027] S304, based on the features of laser point cloud, a three-dimensional model of the surface of the geological body is constructed, and the density anomaly and the stress distribution in the deep data features are combined to generate a three-dimensional reconstruction model of the geological body containing the rock layer inclination, fault distribution and underground water path by voxelization method; In step S304, the construction of the three-dimensional model of the surface of the geological body first utilizes the three-dimensional coordinate information in the laser point cloud features to generate a preliminary geometric model through existing triangular meshing algorithms (such as Delaunay triangulation). Subsequently, existing surface smoothing filtering algorithms (such as Laplace smoothing) are employed to reduce noise and improve model accuracy. The density anomaly information in the deep data features is converted into a continuous distribution field through interpolation methods (such as Kriging interpolation), while the stress distribution is expressed in the form of a vector field based on finite element analysis results. The voxelization method divides the three-dimensional model of the surface of the geological body into a series of small cubes (voxels), each voxel is assigned specific attributes (such as density, stress state) according to its corresponding relationship with the deep data features. Through fusion algorithms (such as maximum likelihood estimation or Markov random field), the voxelized surface model and the deep data features are integrated in three-dimensional space to generate a three-dimensional reconstruction model of the geological body containing detailed geological structure information (rock layer inclination, fault distribution, underground water path). This model not only visually displays the external morphology of the geological body, but also reveals the spatial distribution of its internal physical properties.

[0028] S305, integrate the spatio-temporal correlation features, abnormal hot spot distribution and three-dimensional reconstruction model of the geological body according to the hierarchical visualization rules to form the comprehensive geological environment feature map of the base layer, intermediate layer and top layer.

[0029] Need to be explained in detail in step S305, hierarchical visualization rule adopts three layers structure: the basic layer is original multi-modal feature data (including registered laser point cloud features and surface image features, resolution ≥ 0.5 meters / pixel), the middle layer superimposes space-time correlation features (such as dynamic trend coefficient and periodic main frequency, transparency is set to 40%-60%), the top layer fuses abnormal hot spot distribution (L1 / L2 level risk mark) and geological body three-dimensional reconstruction model (stratum inclination angle accuracy ≤1°, fault line width ≥2 pixels). The specific integration process includes: 1. Basic layer construction: take geographic information system (GIS) platform as the basement, load registered surface image (GeoTIFF format, band number=3) and laser point cloud triangle grid (vertex number ≤500,000) in alignment, adopt pseudo-color rendering (color band: elevation gradient, range-50 to +100 meters) to highlight topographic relief.

[0030] 2. Middle layer superimposition: space-time correlation feature matrix (dimension ≤120,000 × 18) is rendered by GPU acceleration through OpenGL shader, dynamic trend coefficient is indicated by arrow vector (length proportion 1:1000, color coding: red=displacement acceleration, blue=pressure mutation, green=geomagnetic gradient), periodic mode is superimposed in the form of heat map (frequency range 0.01-0.2Hz, color temperature range 3000K-6500K). The system automatically optimizes the occlusion processing (Z-buffer depth test threshold ≤0.1).

[0031] 3. Top layer fusion: abnormal hot spot distribution (polygon boundary, α-shape generation) is covered with semi-transparent filling (L1 level: red, transparency=30%; L2 level: purple, transparency=50%), at the same time, geological body three-dimensional reconstruction model (voxel size=1 cubic meter, adopts Phong illumination model, specular reflection coefficient=0.8) is embedded, and deep density anomaly (contour interval=0.1g / cm³) is shown through slicing tool (thickness ≤5 meters). Finally, comprehensive geological environment feature map (output format: WebGL interactive view or GeoTIFF raster, spatial resolution ≤0.1 meter, timestamp synchronization error ≤10 milliseconds) is generated, and can be zoomed in real time (scale 1:500 to 1:10,000) and multi-attribute inquired (click to get MAI index, risk level details).

[0032] As an optional embodiment of the application, optionally, the space-time correlation features obtained in step S302 include: S3021, constructing a spatiotemporal cube model based on the unified multi-modal feature set, each voxel in the spatiotemporal cube model represents a fixed spatial grid unit and a fixed time interval, and the spatiotemporal cube model contains time series data of ground surface displacement acceleration, pore pressure mutation rate and geomagnetic anomaly; It should be noted that in step S3021, the specific process of constructing the spatiotemporal cube model includes: first, based on the unified multi-modal feature set (dimension ≤200), the system automatically divides the spatial grid (size 200x200, spatial resolution ≤1 meter / grid) and the time layer (size 50 layers, time resolution ≤10 seconds per layer), and generates an initial cube framework. Each voxel is uniquely identified by an index (i, j, k) for spatial position (longitude deviation ≤0.0001°, latitude deviation ≤0.0001°) and time stamp (synchronization error ≤5 milliseconds), and then multi-modal feature data (including ground surface displacement acceleration, pore pressure mutation rate and geomagnetic anomaly time series) are assigned. In specific implementation, existing spatial interpolation algorithms (such as inverse distance weighted method, weight coefficient = 2.0) are used to fill in missing data (missing rate ≤5%), and data normalization processing (Z-score standardization, standard deviation threshold = 1.5) is applied to ensure consistent feature scale. The system automatically checks the integrity of the voxel (such as mirror filling for boundary voxel), and outputs a three-dimensional data array containing displacement acceleration, pressure mutation rate and geomagnetic anomaly.

[0033] S3022, applying a sliding window method to the spatiotemporal cube model to calculate a dynamic trend feature set for each voxel, the dynamic trend features including linear regression slope, acceleration change rate and mutation point detection; In step S3022, it is necessary to explain in detail that a fixed time window size of 30 layers is set (corresponding to a time span ≤ 300 seconds). For each voxel (indexed as (i,j,k)) of time series data, a dynamic trend feature set is calculated by moving the window with a step size of 1 layer. First, the least squares method is applied to the surface displacement acceleration sequence of 30 consecutive layers within the window for linear fitting, the slope value is extracted, and a significance threshold is set (absolute slope value ≥ 0.003). The system automatically records the goodness of fit (R² ≥ 0.85) to ensure the reliability of the model. Secondly, based on the displacement acceleration sequence, its first derivative is calculated (central difference method, step size = 1 layer) to obtain the rate of change. Gaussian filtering (σ = 0.5) is then used to smooth noise. The absolute value of the rate of change must be ≥0.001 to retain effective features. Next, the cumulative sum (CUSUM) algorithm is used to detect sequence mutations. Reference values ​​(historical mean) and decision intervals (±2.5 times standard deviation) are set. When the statistic exceeds the interval, mutation points are marked (location accuracy ≤ 1 layer), and the mutation magnitude (≥5% relative change) is recorded. Finally, the system outputs a dynamic trend feature set (including slope values, rate of change values, and mutation markers), which is integrated into a three-dimensional feature matrix (size 200×200×50, each voxel dimension = 3) for spatiotemporal correlation analysis.

[0034] The expression for calculating the dynamic trend feature set of each voxel is: in, This represents the slope of the linear regression. This indicates the total number of data points within the sliding window. Represents a timestamp. Indicates voxels in time The observed values, Indicates the rate of change of acceleration. Indicates voxels in time The observed values, Indicates voxels in time The observed values, Indicates time interval, This represents the mean of the observations within the sliding window. Represents the threshold coefficient. This represents the standard deviation of the observations within the sliding window; S3023, Superimposed periodic pattern analysis: Perform Fourier transform on the dynamic trend feature set to extract the periodic feature set; The time series data of each voxel (indexed as (i, j, k)) in the dynamic trend feature set needs to be described in detail in step S3023. Apply Fast Fourier Transform (FFT) to the time series data of each voxel (indexed as (i, j, k)) in the dynamic trend feature set, set the frequency range to 0.01-0.2Hz (resolution ≤0.005Hz), extract the main frequency component and its corresponding amplitude. The system automatically identifies significant periodic patterns (amplitude ≥15dB) and calculates the phase angle. Feature extraction includes: main frequency, amplitude of surface displacement acceleration sequence; main frequency, amplitude of pore pressure mutation rate sequence; main frequency, amplitude of geomagnetic anomaly sequence (dimension =1 each). Secondly, through power spectrum density analysis (window function: Hanning window, overlap rate 50%), verify the periodic stability (coefficient of variation ≤0.1), and integrate into the periodic feature set (each voxel dimension =6). The system automatically removes invalid data (signal-to-noise ratio ≤20dB), and outputs the periodic feature matrix (size 200×200×50, total dimension ≤60,000).

[0035] S3024, based on the dynamic trend feature set and the periodic feature set, calculate the spatio-temporal correlation coefficient of surface displacement acceleration and geomagnetic anomaly by cross-correlation function, generate spatio-temporal correlation features containing displacement-pressure-geomagnetic collaborative change; In step S3024, point-by-point cross-correlation analysis is performed on the surface displacement acceleration sequence and the geomagnetic anomaly sequence in the dynamic trend feature set, and the lag time range is set to (-10 seconds to +10 seconds, step 0.1 second), and the correlation coefficient (range -1 to +1, accuracy 0.01) is calculated. The system automatically identifies the maximum correlation coefficient and its corresponding lag time, and marks it as positive or negative correlation (threshold value ≥0.6 for significant correlation). Secondly, the same operation is performed on the pore pressure mutation rate sequence and the surface displacement acceleration, geomagnetic anomaly sequence respectively, and the cross-correlation coefficient and lag time are obtained. Finally, integrate the correlation coefficient matrix (size 200×200×50, each voxel dimension =3, corresponding to displacement-geomagnetic, displacement-pressure, pressure-geomagnetic correlation respectively), and apply significance test (t test, confidence level 95%), remove insignificant correlation features (p value >0.05). Output the spatio-temporal correlation feature matrix. This matrix not only reflects the spatio-temporal synchronization between multi-modal features, but also reveals the internal relationship of geological environmental changes.

[0036] The expression of cross-correlation function is: Wherein, represents the spatio-temporal correlation coefficient, represents the time delay, represents the total number of time steps, represents the surface displacement acceleration at time and spatial coordinates . a global mean value of surface displacement acceleration, a global mean value of surface displacement acceleration, a global mean value of surface displacement acceleration, a global mean value of surface displacement acceleration, a global mean value of surface displacement acceleration.

[0037] As an optional embodiment of the present application, optionally, obtaining the abnormal hotspot distribution marked with potential geological disaster risks in step S303 comprises: S3031, mapping the abnormal data points into the spatial grid cells of the space-time cube model to obtain the spatial indexed abnormal grid cells; In step S3031, it needs to be specified in detail that the system loads the space-time cube model (spatial grid size 200x200) generated in step S302 and extracts its spatial grid coordinate index (i, j). Each abnormal data point (including position coordinates (longitude, latitude, precision≤0.0001°), time stamp (synchronization error≤5 milliseconds) and abnormal type identifier (such as displacement overrun, pressure mutation, geomagnetic anomaly, etc.)) is distributed to the corresponding spatial grid cell (index (i, j)) through a spatial coordinate matching algorithm (such as the nearest neighbor method, distance threshold≤1 meter). For points that cross multiple grid boundaries or are located at the edge of the grid (accounting for≤3%), the system automatically performs spatial interpolation distribution (weight based on the reciprocal square of the distance from the grid center point, weight coefficient=2.0) or is marked as a boundary point (boundary processing strategy: evenly distributed to adjacent grids). The mapping process needs to ensure that each grid cell contains all the abnormal points within it and the threshold distance of the boundary. After the mapping is completed, the system generates an abnormal attribute set for each spatial grid cell (i, j), including: abnormal point number statistics (count≥1), main abnormal type (sorted by frequency), average abnormal intensity (such as displacement amplitude, precision 0.01; pressure change rate, precision 0.01; geomagnetic anomaly, precision 1.0) and first / last abnormal occurrence time stamp. Finally, the spatial indexed abnormal grid cell data (data structure: 200x200xN sparse matrix or key-value pair dictionary, where N is the abnormal attribute dimension of each grid cell≤5) is output, and the data integrity is automatically checked (grid coverage rate≥99%, abnormal point mapping loss rate≤0.1%).

[0038] S3032, calculating the comprehensive abnormal index for each abnormal grid cell to obtain the grid cell with the comprehensive abnormal index; In step S3032, the system first loads the spatial indexed abnormal grid cell data output in step S3031, extracts the abnormal attribute set of each grid cell (index (i, j)) including the abnormal point number statistics, main abnormal type identifier, average abnormal intensity value and first / last abnormal timestamp. Then the system automatically performs normalization processing (min-max scaling to the range [0, 1], precision 0.01) and sets the comprehensive abnormal index activation threshold ≥ 0.6 (higher than this value is marked as high risk). The process of calculating the comprehensive abnormal index needs to check the data validity (such as using adjacent cell interpolation when missing attributes, interpolation algorithm: bilinear interpolation, weight based on Euclidean distance), output grid cell data with comprehensive abnormal index (data structure: 200x200x1 matrix, each element stores the comprehensive abnormal index value, precision 0.001), and automatically record the calculation log (including weight coefficient, baseline value and abnormal type mapping table). Finally, the system ensures that the calculation error of the comprehensive abnormal index of each grid cell is ≤ ± 0.05 (verified by Monte Carlo simulation, iteration number = 1000 times), and integrates it into a three-dimensional index matrix.

[0039] The expression for calculating the comprehensive abnormal index is: wherein, represents the comprehensive abnormal index of the spatial grid cell with coordinates , , and all represent weight coefficients, represents the frequency of geomagnetic signal mutation events in the grid cell , represents the frequency of pore pressure sudden increase events in the grid cell , represents the frequency of ground surface displacement acceleration events in the grid cell ; S3033, based on the grid cell with abnormal index, using DBSCAN density clustering algorithm, setting neighborhood radius, and merging density reachable abnormal grid cells to form abnormal clusters; In step S3033, the system sets the parameters of the DBSCAN algorithm based on the three-dimensional index matrix output in step S3032, including the neighborhood radius (epsilon value, set to 0.1 according to the data distribution characteristics) and the minimum number of points (MinPts, set to 5). The DBSCAN algorithm traverses each grid cell (index (i, j)), calculates its Euclidean distance with the surrounding grid cells, and if the distance is less than epsilon, it is considered to be density reachable. The system automatically identifies the core points (i.e. grid cells with point number in the neighborhood >= MinPts) and marks them as initial cluster centers. Then, the algorithm starts from the core points and recursively merges the density-reachable grid cells to form abnormal clusters. For boundary points (i.e. grid cells with point number in the neighborhood < MinPts), if they are located within the neighborhood of the formed abnormal clusters, they are marked as cluster members; otherwise, they are marked as noise points. During the clustering process, the system automatically records the cluster label, the number of grid cells in the cluster, and the average abnormal index. Finally, the system outputs the abnormal cluster data (data structure: list or dictionary containing cluster label, cluster center coordinates, list of grid cells in the cluster, and average abnormal index), and automatically checks the rationality of the clustering results (such as cluster number, cluster size distribution, and average abnormal index whether consistent with the expected rules of geological environmental changes).

[0040] The expression of the DBSCAN density clustering algorithm is: wherein, represents a function for judging whether a point is a core point, represents the set of all points within the neighborhood with as the center and radius , represents the minimum number of neighborhood points required to form a core point, represents the density-reachable relationship between point and point , represents that point is a core point, represents the density-connected relationship between point and point , represents the connection of two conditions, represents the existence quantifier, represents the existence of an intermediate point , represents the bidirectional density connection, represents the th abnormal cluster, represents the current point, represents the existence of a point belonging to the abnormal cluster , current point located in the neighborhood of a certain core point but not being a core point itself, denotes a set of noise points, denotes the entire dataset, denotes the union of all anomaly clusters, i.e. points that are assigned to any cluster, denotes a set difference operation; S3034, filtering out isolated anomaly points in anomaly clusters, and keeping anomaly clusters containing more than or equal to two number of anomaly points as anomaly hotspot regions of potential geological disaster risk, In step S3034, the system traverses each anomaly cluster based on the anomaly cluster data output in step S3033, and counts the number of grid cells in the cluster. If the number of grid cells in the cluster is less than 2, it is considered as an isolated anomaly point, which is excluded from the anomaly cluster list. For the remaining anomaly clusters, the system automatically calculates the average anomaly index in the cluster and marks it as an anomaly hotspot region of potential geological disaster risk. In addition, the system also records the spatial range (latitude and longitude coordinate range), main anomaly type and anomaly intensity distribution of the anomaly hotspot region. Finally, the system outputs the filtered anomaly hotspot region data (data structure: a list or dictionary containing hotspot region label, region center coordinates, region range, main anomaly type, average anomaly index and cluster grid cell list), and automatically generates an anomaly hotspot distribution map (using a heat map, the color depth reflects the anomaly index).

[0041] S3035, based on the comprehensive anomaly index, the anomaly hotspot regions are classified into different levels, and the anomaly hotspot distribution marked with potential geological disaster risk is obtained.

[0042] In step S3035, it needs to be explained in detail that according to the filtered anomaly hotspot region data output in S3034, the average anomaly index of each hotspot region is extracted. According to the preset level division standard (such as low risk: 0.6≤ anomaly index <0.8; medium risk: 0.8≤ anomaly index <0.95; high risk: anomaly index ≥0.95), the system automatically assigns a risk level label to each hotspot region. For hotspot regions at the level boundary value, the system uses fuzzy comprehensive evaluation method (considering factors such as anomaly type, intensity distribution and spatial aggregation, the weight is determined based on expert scoring method) for detailed division. After the level division is completed, the system integrates the hotspot region data, including hotspot region label, level label, region center coordinates, range, main anomaly type, average anomaly index and cluster grid cell list, and automatically generates a level division report. Finally, the system outputs the anomaly hotspot distribution data marked with potential geological disaster risk (data structure: a list or dictionary containing level label), and visualizes the level distribution map (using a hierarchical color map, different colors represent different risk levels).

[0043] As an optional embodiment of the present application, optionally, the obtaining of the geological risk prediction probability in step S4 comprises: S401, based on the spatio-temporal correlation features in the comprehensive geological environment feature map, the abnormal hot spot distribution and the three-dimensional reconstruction model of the geological body, a multi-modal data fusion matrix is constructed, and the fusion matrix contains dynamic trend features, periodic patterns and geological structure parameters; It should be noted that in step S401, based on the comprehensive geological environment feature map, the abnormal hot spot distribution data and the three-dimensional reconstruction model of the geological body output in step S3, the system extracts the spatio-temporal correlation features in the feature map (such as the spatio-temporal distribution trend of abnormal points, the periodic change of geological activities, etc.), the main abnormal types, intensity and spatial aggregation features of the abnormal hot spot distribution, and the geological structure parameters (such as the rock layer dip angle, fault distribution, lithology features, etc.) in the three-dimensional reconstruction model of the geological body. Then, the system uses a multi-modal data fusion algorithm (such as a deep learning model, such as a convolutional neural network CNN or a recurrent neural network RNN, and the model training uses labeled historical geological data) to fuse the above features into a high-dimensional feature vector, forming a multi-modal data fusion matrix. During the fusion process, the system automatically performs feature selection (based on correlation analysis and importance evaluation) and feature dimension reduction (such as principal component analysis PCA) to optimize the model performance. Finally, the system outputs the multi-modal data fusion matrix.

[0044] S402, a hybrid inference model is constructed based on the fusion matrix, the hybrid inference model includes a physical mechanism driven layer, a data driven layer and a dynamic correction layer, the physical mechanism driven layer is used to calculate fault stability parameters and stress accumulation based on the principle of geomechanics, the data driven layer is used to establish a nonlinear mapping relationship between features and historical disaster events using a random forest algorithm, and the dynamic correction layer is used to adjust the prediction results in real time through a long short-term memory network, and the latest monitoring data is combined to correct the probability value; In step S402, it needs to be explained in detail that the implementation steps of the physical mechanism driven layer include: loading the fault geometric parameters (strike, dip angle, length, depth) and rock mass physical and mechanical parameters (elastic modulus, Poisson's ratio, cohesion, internal friction angle) in the three-dimensional reconstruction model of the geological body, calculating the shear stress and normal stress distribution on the fault surface using the Coulomb stress change model; based on the modified Byerlee friction law, combined with the pore pressure data, the fault sliding trend parameters (such as the sliding potential index) are calculated; the stress accumulation process under the action of regional tectonic stress field is simulated by applying the linear elastic dislocation theory, and the fault stability parameters (including the sliding trend index, stress accumulation rate) and the stress tensor of the key area (accuracy 0.1 MPa) are output, and the reasonableness of the calculation results is verified through finite element numerical simulation (element size ≤100m) (error ≤±5%).

[0045] The data-driven layer implementation steps include: inputting the feature vector (dimension ≥ 50) in the multi-modal data fusion matrix, and loading the historical disaster event database (including time, location, disaster type, intensity label); setting the random forest hyperparameters (number of decision trees = 500, maximum depth = 15, feature selection method = Gini index), performing feature importance evaluation (based on the average impurity reduction method), and selecting the top-20 key features; using stratified sampling (training set: test set = 8:2) to train the random forest classifier, and predicting the probability value of a specific geological disaster (landslide, ground subsidence, rock burst) occurring within a specific time window (such as 30 days) in the future; performing ten-fold cross-validation during model training (accuracy target ≥ 90%, AUC ≥ 0.95), and outputting the feature importance ranking report and the predicted probability intermediate value.

[0046] The dynamic correction layer implementation steps include: constructing an LSTM-based time correction model (number of layers = 2, number of hidden units = 128), inputting the initial prediction probability sequence output by the data-driven layer (time step = 10 days), and the latest multi-source monitoring data stream (geomagnetic, displacement, pore pressure, etc. abnormal indicator change rate) collected in real time in step S1; the model learns the time deviation pattern between historical prediction probability and actual disaster occurrence, and dynamically generates correction weight coefficients (range [0.8, 1.2]); the system performs online correction every 6 hours, multiplies the latest prediction probability by the correction weight coefficient, and adds an incremental adjustment term based on the real-time abnormal intensity mutation (adjustment amplitude ≤ ±0.1); the correction result is smoothed by Kalman filtering, and the dynamically updated geological risk prediction probability value (precision 0.01) is finally output, and the correction log (including correction time, correction weight, adjustment amplitude) is recorded.

[0047] S403, calculating an initial risk probability based on the hybrid inference model; In step S403, the multi-modal data fusion matrix output in step S401 is input into the hybrid inference model. The physical mechanism-driven layer calculates the fault stability parameter and the ground stress accumulation based on the input geological structure parameters and geomechanical principles. The data-driven layer uses the random forest algorithm to establish a nonlinear mapping relationship between the dynamic trend features, periodic patterns in the fusion matrix, and historical disaster event data, and outputs the initial prediction probability value. The dynamic correction layer is not working, and the initial prediction probability remains unchanged. The system integrates the output results of the physical mechanism-driven layer and the data-driven layer to form an initial risk probability distribution map (represented by a hierarchical color-coded map, with different colors representing different probability intervals).

[0048] S404, introducing real-time monitoring data to dynamically correct the initial risk probability; the real-time monitoring data includes abnormal cluster expansion speed, ground stress mutation amplitude, and displacement acceleration change rate; In step S404, the system collects monitoring data such as abnormal cluster expansion speed, ground stress mutation amplitude, and displacement acceleration change rate in real time. These data are fused with the initial risk probability distribution map output in step S403 and input into the dynamic correction layer of the hybrid inference model. The dynamic correction layer learns the time sequence deviation pattern between historical prediction probability and actual disaster occurrence through a long short-term memory network, and dynamically generates a correction weight coefficient. The system adjusts the initial risk probability in real time according to the correction weight coefficient and the abnormal intensity of real-time monitoring data, and outputs the dynamically updated geological risk prediction probability value. During the correction process, the system records correction logs, including correction time, correction weight, and adjusted probability value, for subsequent analysis and verification.

[0049] S405, map the corrected initial risk probability to the three-dimensional reconstruction model of the geological body to generate a spatially continuous risk heat map and mark high-risk areas; In step S405, the system maps the dynamically updated geological risk prediction probability value output in step S404 to the three-dimensional reconstruction model of the geological body. During the mapping process, the system considers factors such as the geometric shape of the geological body, lithology distribution, and fault structure, and uses a spatial interpolation algorithm (such as Kriging interpolation or inverse distance weighted interpolation) to calculate the risk probability of non-monitoring points, generating a spatially continuous risk heat map. The heat map uses different colors to represent different risk probability intervals, visually displaying the spatial distribution of potential geological disaster risks within the geological body. For high-risk areas, the system automatically marks them and outputs information such as the spatial range (latitude and longitude coordinate range), main geological features, risk probability value, and potential disaster type. In addition, the system also generates a three-dimensional visualization model of the high-risk area to help users intuitively understand the geological structure and disaster risk characteristics of the high-risk area. Finally, the system integrates the risk heat map, high-risk area marking information, and three-dimensional visualization model.

[0050] S406, output the geological risk prediction probability based on the risk heat map.

[0051] In step S406, the system outputs the integrated risk heat map, high-risk area label information, and three-dimensional visualization model. The risk heat map visually displays the risk probabilities of different areas within the geological body through a color gradient. Users can quickly identify the distribution of potential geological disaster risks by observing the color changes of the heat map. The label information of the high-risk area is listed in detail in text form, including the specific location, geological characteristics, risk probability value, and potential disaster type of the high-risk area, providing users with detailed risk assessment references. Meanwhile, the three-dimensional visualization model of the high-risk area helps users better understand the geological structure and disaster risk characteristics of the high-risk area through stereoscopic display. Finally, the system outputs the geological risk prediction probability data, which, combined with the risk heat map and high-risk area label information, constitutes a complete report for geological disaster risk assessment.

[0052] As an optional embodiment of the present application, the expression for calculating the initial risk probability in step S403 based on the hybrid inference model is: wherein, represents the initial risk probability, represents the time-varying physical model weight coefficient, represents the theoretical risk probability calculated by the hybrid inference model, represents the fault dip angle, represents the deep geostress value, represents the fault buffer zone distance, represents the time-varying data-driven model weight coefficient, represents the basic risk probability, represents the comprehensive anomaly index of the spatial grid element, represents the surface displacement acceleration, represents the groundwater path influence coefficient, represents the fixed weight coefficient of the geological structure correction term, represents the geological structure correction term, represents the fault density, represents the pore pressure.

[0053] Embodiment 2 A geological environment monitoring system, the system being used to perform a geological environment monitoring method; the system comprising: a data acquisition module, configured to construct an air-space-ground-deep four-dimensional cooperative monitoring network to acquire multi-modal geological environment data, the multi-modal geological environment data including surface image data, unmanned aerial vehicle laser point cloud data, ground sensor array signal, and deep geostress measurement data; The preprocessing module is connected with the acquisition module and is configured to preprocess the multi-modal geological environment data to obtain multi-modal feature data and abnormal data points after dynamic load balancing of computing resources; The data fusion module is connected with the preprocessing module and is configured to perform multi-modal data fusion on the multi-modal feature data and the abnormal data points to generate a comprehensive geological environment feature map, which includes a time-space correlation feature, an abnormal hotspot distribution, and a three-dimensional reconstruction model of a geological body. The prediction module is connected with the data fusion module and is configured to perform geological risk prediction based on the comprehensive geological environment feature map by using a geological risk prediction model to obtain a geological risk prediction probability.

[0054] It should be noted that the geological environment monitoring system in this embodiment is used to execute the geological environment monitoring method in Embodiment 1. The data acquisition module, the preprocessing module, the data fusion module, and the prediction module in the system are connected with each other through a high-speed communication bus to ensure the efficiency of real-time data transmission and processing. The data acquisition module uses the space-air-ground-deep four-dimensional cooperative monitoring network to realize all-around monitoring of the surface, the air, the underground, and the deep geological environment, thereby ensuring the comprehensiveness and accuracy of the data. The preprocessing module uses the load balancing technology to dynamically allocate computing resources according to the data volume and computing requirements, thereby improving the data processing efficiency. The data fusion module performs deep fusion on the data from different monitoring means by using a multi-modal data fusion algorithm to generate a comprehensive geological environment feature map containing rich geological information. The prediction module performs intelligent prediction by using a geological risk prediction model based on the feature map to output an accurate geological risk prediction probability. The entire system is highly integrated and has a high degree of automation, and can realize real-time monitoring and risk early warning of the geological environment.

[0055] Although the embodiments of the present application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements, and variations can be made to the embodiments without departing from the principles and purposes of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A method of monitoring a geological environment, characterized by, The method comprises: S1, constructing an aerospace-terrestrial deep four-dimensional collaborative monitoring network to collect multi-modal geological environment data, the multi-modal geological environment data comprising surface image data, unmanned aerial vehicle laser point cloud data, ground sensor array signal and deep geostress measurement data; S2, preprocessing the multi-modal geological environment data to obtain multi-modal feature data and abnormal data points after dynamic load balancing distribution of computing resources; S3, multi-modal data fusion on the multi-modal feature data and abnormal data points to generate a comprehensive geological environment feature map, the comprehensive geological environment feature map comprising time-space correlation features, abnormal hotspot distribution and a three-dimensional geological body reconstruction model; S4, geological risk prediction based on the comprehensive geological environment feature map using a geological risk prediction model to obtain a geological risk prediction probability.

2. The geologic environment monitoring method of claim 1, wherein, Constructing an aerospace-terrestrial deep four-dimensional collaborative monitoring network in step S1 comprises: S101, deploying a satellite remote sensing system to collect surface image data, the surface image data comprising high-resolution optical images and synthetic aperture radar images; S102, constructing an unmanned aerial vehicle platform and loading a laser radar sensor to collect unmanned aerial vehicle laser point cloud data, and using the unmanned aerial vehicle laser point cloud data to construct a three-dimensional geological body model; S103, installing a ground sensor array to collect geological signals, the ground sensor array comprising displacement sensors, temperature sensors and humidity sensors; S104, installing a deep geostress measurement module through a borehole to collect deep geostress measurement data; S105, using a communication network to preliminarily fuse the surface image data, the three-dimensional geological body model, the geological signals and the deep geostress measurement data to obtain an aerospace-terrestrial deep four-dimensional collaborative monitoring network.

3. The geologic environment monitoring method of claim 1, wherein, Preprocessing the multi-modal geological environment data in step S2 to obtain multi-modal feature data and abnormal data points after dynamic load balancing distribution of computing resources comprises: S201, receiving the multi-modal geological environment data of the aerospace-terrestrial deep four-dimensional collaborative monitoring network based on an edge computing node, and classifying the multi-modal geological environment data according to data types and priorities to obtain classified multi-modal geological environment data; S202, based on the classified multi-modal geological environment data, using a dynamic load balancing algorithm to monitor the load of the edge computing node in real time, and dynamically adjusting task distribution according to data priorities and node loads to obtain multi-modal geological environment data after dynamic load balancing distribution of computing resources; S203, performing feature extraction on the multi-modal geological environment data after dynamic load balancing distribution of computing resources to obtain multi-modal feature data; S204, using an isolation forest algorithm to detect abnormalities on the multi-modal feature data to mark abnormal scenarios such as geomagnetic signal mutation, pore pressure surge and surface displacement acceleration, and to obtain abnormal data points.

4. The geologic environment monitoring method of claim 3, wherein, The expression of the dynamic load balancing algorithm in step S202 is: wherein, represents the task assigned to the node , and both represent the weight adjustment coefficients of the load and the priority, represents the comprehensive load of the node , represents the priority coefficient of the task , represents the capability matching coefficient of the node to the task , represents the real-time weight of the task , represents the total number of edge computing nodes, represents the comprehensive load of the node , represents the capability matching coefficient of the node to the task ; The expression of the feature extraction operation in step S203 is: wherein, represents multi-modal feature data, represents surface image feature, represents laser point cloud feature, represents sensor signal feature, represents deep data feature, represents vector concatenation operation, and both represent weight coefficient, represents pixel gray level, represents distance , direction under gray co-occurrence probability, represents pixel point number, represents pixel normalized vegetation index, represents octree node number, represents whether octree node contains valid point cloud, if =1, it contains valid point cloud, if =0, it does not contain valid point cloud, represents depth of octree node in octree, represents slope weight coefficient, represents terrain height gradient in / direction, represents taking maximum value, represents time length, represents wavelet basis function coefficient under scale , translation , represents abnormal gradient weight, represents sudden value of geomagnetic signal in time , represents inner product of quantum state vector after QFT transformation, represents operator of quantum Fourier transform, represents density deviation weight, represents number of deep exploration points, represents density value of th deep exploration point, represents regional average density; The expression of the isolation forest algorithm in step S204 is: wherein, represents the th multimodal feature data point, represents a geomagnetic signal feature, represents a pore pressure feature, represents a surface displacement feature, represents a path length of the multimodal feature data point , represents Euler's number, represents an average path length, represents a training sample, represents a data point in a tree , represents an anomaly score of the multimodal feature data point , represents an expected value of a path length, represents an anomaly threshold value, represents a 95th percentile function.

5. The geologic environment monitoring method of claim 1, wherein, Generating a comprehensive geological environment feature map in step S3 comprises: S301, based on a unified geographic coordinate system and a timestamp, spatially registering and temporally aligning surface image features, laser point cloud features, sensor signal features and deep data features in the multi-modal feature data, to obtain a multi-modal feature set with unified space-time reference; S302, performing space-time correlation feature extraction on the multi-modal feature set with unified space-time reference, constructing a space-time cube model and superimposing dynamic trends and periodic patterns to obtain space-time correlation features including surface displacement acceleration, pore pressure mutation rate and space-time correlation of geomagnetic anomalies; S303, spatially aggregating abnormal scenarios of geomagnetic signal mutation, pore pressure surge and surface displacement acceleration in the abnormal data points to obtain aggregated data, identifying abnormal hot spot regions based on the aggregated data using a density clustering algorithm to obtain abnormal hot spot distribution marked with potential geological disaster risks; S304, constructing a geological body surface three-dimensional model based on the laser point cloud features, combining density anomalies and geostress distribution in the deep data features, and fusing the geological body surface three-dimensional model and deep data features by voxelization to generate a geological body three-dimensional reconstruction model including rock layer inclination, fault distribution and underground water path; S305, integrating the space-time correlation features, abnormal hot spot distribution and geological body three-dimensional reconstruction model according to layered visualization rules to form comprehensive geological environment feature maps of the base layer, intermediate layer and top layer.

6. The geologic environment monitoring method of claim 5, wherein, The space-time correlation features including surface displacement acceleration, pore pressure mutation rate and space-time correlation of geomagnetic anomalies obtained in step S302 include: S3021, constructing a space-time cube model based on the multi-modal feature set with unified space-time reference, each voxel in the space-time cube model representing a fixed spatial grid unit and a fixed time interval, and the space-time cube model including time series data of surface displacement acceleration, pore pressure mutation rate and geomagnetic anomalies; S3022, applying a sliding window method to the space-time cube model to calculate a dynamic trend feature set of each voxel, the dynamic trend features including linear regression slope, acceleration change rate and mutation point detection; S3023, superimposing periodic pattern analysis, performing Fourier transform on the dynamic trend feature set to extract a periodic feature set; S3024, based on the dynamic trend feature set and the periodic feature set, calculating a space-time correlation coefficient of surface displacement acceleration and geomagnetic anomalies by a cross-correlation function to generate space-time correlation features including displacement-pressure-geomagnetic collaborative changes.

7. The geologic environment monitoring method of claim 5, wherein, The abnormal hot spot distribution marked with potential geological disaster risks obtained in step S303 includes: S3031, mapping the abnormal data points to spatial grid units of the space-time cube model to obtain spatially indexed abnormal grid units; S3032, calculating a comprehensive abnormality index for each abnormal grid unit to obtain grid units with comprehensive abnormality indexes; S3033, based on the grid units with the abnormality indexes, using a DBSCAN density clustering algorithm, setting a neighborhood radius, and merging density-reachable abnormal grid units to form abnormal clusters; S3034, filtering out isolated abnormal points in the abnormal cluster, and retaining abnormal clusters containing two or more abnormal points as abnormal hot spot regions of potential geological disaster risks, S3035, grading the abnormal hot spot regions based on the comprehensive anomaly index, to obtain abnormal hot spot distribution marked with potential geological disaster risks.

8. The geologic environment monitoring method of claim 1, wherein, The obtaining of the geological risk prediction probability in step S4 includes: S401, constructing a multi-modal data fusion matrix based on the spatio-temporal correlation features in the comprehensive geological environment feature map, the abnormal hot spot distribution, and the geological body three-dimensional reconstruction model, the fusion matrix including dynamic trend features, periodic patterns, and geological structure parameters; S402, constructing a hybrid inference model based on the fusion matrix, the hybrid inference model including a physical mechanism driven layer, a data driven layer, and a dynamic correction layer, the physical mechanism driven layer being configured to calculate fault stability parameters and geostress accumulation based on geomechanical principles, the data driven layer being configured to establish a nonlinear mapping relationship between features and historical disaster events using a random forest algorithm, and the dynamic correction layer being configured to adjust the prediction results in real time through a long short-term memory network, and correct the probability value in combination with the latest monitoring data; S403, calculating an initial risk probability based on the hybrid inference model; S404, introducing real-time monitoring data to dynamically correct the initial risk probability, the real-time monitoring data including abnormal cluster expansion speed, geostress mutation amplitude, and displacement acceleration change rate; S405, mapping the corrected initial risk probability to the geological body three-dimensional reconstruction model to generate a spatially continuous risk heat map, and marking high-risk regions; S406, outputting a geological risk prediction probability based on the risk heat map.

9. The geologic environment monitoring method of claim 8, wherein, The expression for calculating the initial risk probability based on the hybrid inference model in step S403 is: wherein, represents an initial risk probability, represents a physical model weight coefficient changing over time, represents a theoretical risk probability calculated by a hybrid inference model, represents a fault dip angle, represents a deep ground stress value, represents a fault buffer distance, represents a data-driven model weight coefficient changing over time, represents a basic risk probability, represents a comprehensive anomaly index of a spatial grid cell, represents a surface displacement acceleration, represents a groundwater path influence coefficient, represents a fixed weight coefficient of a geological structure correction term, represents a geological structure correction term, represents a fault density, represents a pore pressure.

10. A geologic environment monitoring system characterized by, The system is configured to perform the geological environment monitoring method of any one of claims 1 to 9; and the system includes: a data acquisition module configured to construct an air-space-ground-deep four-dimensional cooperative monitoring network to acquire multi-modal geological environment data, the multi-modal geological environment data including surface image data, unmanned aerial vehicle laser point cloud data, ground sensor array signals, and deep geostress measurement data; a preprocessing module connected to the acquisition module and configured to preprocess the multi-modal geological environment data to obtain multi-modal feature data and abnormal data points after dynamic load balancing and distribution of computing resources; a data fusion module connected to the preprocessing module and configured to perform multi-modal data fusion on the multi-modal feature data and the abnormal data points to generate a comprehensive geological environment feature map, the comprehensive geological environment feature map including spatio-temporal correlation features, abnormal hot spot distribution, and a geological body three-dimensional reconstruction model; a prediction module connected to the data fusion module and configured to perform geological risk prediction based on the comprehensive geological environment feature map using a geological risk prediction model to obtain a geological risk prediction probability.

Images