Multi-modal agent cluster control system for dynamic evaluation of geological disasters

Abstract

The invention discloses a multi-modal intelligent agent cluster control system for dynamic evaluation of geological disasters, particularly relates to the technical field of three-dimensional geological modeling and digital twinning, and is used for solving the problem that a model expresses the state of a disaster body inaccurately due to inconsistency of multi-modal monitoring data during fusion of a digital twinning model. A data acquisition module collects multi-dimensional and multi-modal monitoring data by using a heterogeneous agent cluster, a cause judgment module judges whether the data is derived from real disaster evolution, a risk identification module identifies a high-risk key area, a task generation module generates an agent cluster collaborative observation task plan for the high-risk key area, and a task management module manages the intelligent agent cluster collaborative observation task plan for the high-risk key area. The planning optimization module re-optimizes the collaborative observation task planning, and the execution updating module controls the heterogeneous agent cluster to execute the re-optimized task planning and feeds back and updates the digital twinborn three-dimensional model in real time, thereby realizing dynamic accurate evaluation and early warning of geological disasters.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional geological modeling and digital twin technology, and in particular to a multimodal intelligent agent cluster control system for dynamic assessment of geological disasters. Background Technology

[0002] In the field of geological disaster monitoring and early warning, to overcome the limitations of traditional single-sensor networks, various mobile monitoring devices, such as drones and ground robots, can be introduced to work collaboratively with fixed sensor nodes. A three-dimensional geological model can be constructed as a platform for data integration and display. By fusing heterogeneous data from multiple sources—air, ground, and underground—the aim is to achieve a more comprehensive situational awareness of the disaster body. Digital twin technology, as a key means of connecting the physical world and virtual space, can be applied to construct three-dimensional visualization models of geological disaster bodies to reflect their dynamic changes. However, in practical implementation, this multi-platform collaborative sensing mode has not yet formed an organically unified whole.

[0003] While existing technologies can acquire multimodal monitoring data and map it into digital twin models, in practical applications, due to inherent differences in the working principles, observation perspectives, spatiotemporal scales, and accuracy of different monitoring platforms, the sensed information transmitted back often exhibits significant uncertainty or even direct contradictions during model fusion. This inconsistency stemming from the underlying data leads to inherent ambiguity in the digital twin model when expressing key states of the disaster body, such as deformation areas and stability coefficients, making it difficult to output unique and reliable assessment conclusions. This severely restricts the accuracy and reliability of risk assessment and early warning decisions based on the model. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a multimodal intelligent agent cluster control system for dynamic assessment of geological disasters.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: A multimodal intelligent agent swarm control system for dynamic assessment of geological hazards includes: The data acquisition module is used to collect multi-dimensional, multimodal monitoring data from the air, ground, and underground through a heterogeneous intelligent agent cluster that includes drones, ground robots, and underground sensors. The cause judgment module is used to input multimodal monitoring data into the three-dimensional digital twin model of geological disasters, and analyze the physical causes of data inconsistency based on the physical mechanism of geological disaster evolution to determine whether the data inconsistency originates from the actual disaster evolution. The risk identification module is used to identify high-risk key areas by collaboratively analyzing the critical paths of disaster evolution and the topological characteristics of multimodal data fields when the disaster originates from real disaster evolution. The task generation module is used to generate collaborative observation task plans for intelligent agent clusters targeting high-risk critical areas. The planning optimization module is used to predict the degree of uncertainty reduction in each region of the geological disaster digital twin 3D model after the execution of the collaborative observation task planning, and to weigh it with the task execution cost to re-optimize the collaborative observation task planning; The execution update module is used to control the heterogeneous intelligent agent cluster to execute the re-optimized collaborative observation task plan, and to feed back the new monitoring data acquired after execution to update the geological disaster digital twin 3D model in real time.

[0006] Furthermore, through a heterogeneous intelligent agent cluster comprising drones, ground robots, and underground sensors, multi-dimensional, multimodal monitoring data is collected from the air, ground, and underground, including: By using drones to fly in the airspace and collect optical images; The ground robot moves across the earth and collects laser point cloud data. Displacement data is collected by burying underground sensors inside the disaster body; Data acquisition operations from drones, ground robots, and underground sensors are synchronized via wireless communication networks based on a unified time reference to generate spatiotemporally aligned multimodal monitoring data.

[0007] Furthermore, multimodal monitoring data is input into a digital twin 3D model of geological hazards, and the physical causes of data inconsistencies are analyzed based on the physical mechanisms of geological hazard evolution to determine whether the data inconsistencies stem from the actual evolution of the hazard, including: Optical images, laser point cloud data, and displacement data are jointly registered into the unified coordinate system of the digital twin 3D model of geological hazards; The correlation between the spatial distribution of surface deformation features recorded by optical imagery and surface elevation changes generated by laser point cloud data was compared. The coupling relationship between the deep deformation patterns reflected by displacement data and the surface deformation characteristics in terms of physical mechanisms was analyzed. When the deep deformation pattern and the surface deformation characteristics conform to the mechanical mechanism of shear slip or tensile fracture of rock and soil, it is determined that the inconsistency in the data stems from the actual evolution of the disaster.

[0008] Furthermore, the three-dimensional digital twin model of geological hazards is constructed in the following ways: integrating regional geological survey data and geographic information system base maps to construct a three-dimensional geological structure framework; orthorectifying and texturing optical images from multimodal monitoring data onto the surface of the three-dimensional geological structure framework; generating a digital elevation model from laser point cloud data through interpolation algorithms and fusing it into the three-dimensional geological structure framework to characterize topographic changes; and assigning displacement data to corresponding underground monitoring points within the three-dimensional geological structure framework, thereby forming a three-dimensional digital twin model of geological hazards that integrates multi-source information from air, ground, and underground sources.

[0009] Furthermore, when the disaster originates from real-world disaster evolution, high-risk key areas are identified through collaborative analysis of the critical paths of disaster evolution and the topological characteristics of the multimodal data field, including: Spatiotemporal interpolation is performed on displacement data to generate a continuously spatiotemporally distributed displacement field; The region in the displacement field whose displacement rate exceeds a preset rate threshold is identified as the deformation initiation zone; By tracking the propagation direction and speed of isohyets in the displacement field over time, the spatiotemporal evolution sequence of deformation from the deformation initiation zone to the surrounding area is determined as the key path; Extract key hole structures characterizing potential sliding surface boundaries from the elevation change field generated from laser point cloud data; By spatially coupling the deformation acceleration segment on the critical path with the geometric discontinuity region characterized by the critical hole structure, high-risk critical areas that are simultaneously in the deformation acceleration stage and have significant geometric discontinuity characteristics are identified.

[0010] Furthermore, extracting key void structures characterizing potential sliding surface boundaries from the elevation change field generated from laser point cloud data includes: registering and differencing multi-period laser point cloud data to generate an elevation change field; using topological data analysis methods to identify persistent high Betti number regions in the elevation change field; and extracting key void structures characterizing the persistent high Betti number regions, with the key void structures corresponding to potential sliding surface boundaries with drastic elevation changes.

[0011] Furthermore, the spatial coupling analysis of the deformation acceleration segment and the geometric discontinuity region characterized by the key hole structure on the critical path includes: identifying segments on the critical path where the displacement acceleration exceeds the acceleration threshold as deformation acceleration segments; performing spatial overlay analysis of the deformation acceleration segment and the geometric discontinuity region characterized by the key hole structure; identifying spatially overlapping deformation acceleration segments and geometric discontinuity regions, and marking the overlapping deformation acceleration segments and geometric discontinuity regions as high-risk critical areas that are simultaneously in the deformation acceleration stage and have significant geometric discontinuity characteristics.

[0012] Furthermore, a collaborative observation task plan for agent clusters targeting high-risk critical areas is generated, including: Observation priorities are determined based on the spatial distribution characteristics and deformation development stages of high-risk key areas. Based on the matching relationship between observation priority and the spatial location and monitoring capabilities of UAVs, ground robots and underground sensors in the intelligent agent cluster, a collaborative observation task plan including flight trajectory, movement path and sampling point is generated. Among them, drones are responsible for large-scale optical image acquisition, ground robots are responsible for fine acquisition of laser point cloud data, and underground sensors perform continuous displacement data monitoring.

[0013] Furthermore, the degree of uncertainty reduction in each region of the geological hazard digital twin 3D model is predicted after the collaborative observation mission planning is implemented, and this is weighed against the mission execution cost to re-optimize the collaborative observation mission planning, including: Based on the current uncertainty distribution in each region of the digital twin 3D model of geological disasters, the simulation shows the effect of reducing uncertainty by adding optical images, laser point cloud data and displacement data after the collaborative observation task planning is carried out. Assess the resource consumption of the agent swarm required to execute the collaborative observation mission plan, including flight time, travel distance, and energy consumption; Based on the ratio of uncertainty reduction effect to resource consumption, adjust the task allocation and execution order of UAVs, ground robots and underground sensors in the collaborative observation mission planning.

[0014] Furthermore, the heterogeneous intelligent agent cluster is controlled to execute a re-optimized collaborative observation task plan, and the newly acquired monitoring data is fed back in real time to update the geological hazard digital twin 3D model, including: The re-optimized collaborative observation mission plan's flight trajectory, movement path, and sampling point instructions are sent to the drone, ground robot, and underground sensor respectively via wireless communication network. It receives real-time optical images collected by drones, real-time laser point cloud data collected by ground robots, and real-time displacement data collected by underground sensors. By fusing real-time optical images, real-time laser point cloud data, and real-time displacement data with monitoring data of corresponding spatiotemporal locations in the digital twin 3D model of geological hazards, the deformation field distribution and uncertainty parameters in the digital twin 3D model of geological hazards are updated.

[0015] The beneficial effects of this invention are: 1. By constructing a multimodal intelligent agent cluster control system, dynamic and accurate perception of geological hazard bodies across all time and space scales was achieved. Through collaborative observation of the air, ground, and underground by a heterogeneous intelligent agent cluster, the system acquired multidimensional monitoring data covering the surface and interior of the hazard body. Based on the physical mechanism of geological hazard evolution, the system identified the causes of data inconsistencies, effectively distinguishing between measurement errors and actual hazard evolution signals. This enabled the digital twin 3D model to accurately reflect the actual deformation state of the hazard body, significantly improving the accuracy and reliability of the model representation. At the same time, by analyzing the topological characteristics of the multimodal data field, the system can accurately identify high-risk key areas in the hazard evolution process, providing a scientific basis for subsequent targeted monitoring.

[0016] 2. Based on identified high-risk areas, it can autonomously generate optimized collaborative observation task plans and achieve optimal allocation of monitoring resources by balancing the degree of uncertainty reduction and execution costs. This not only ensures sufficient monitoring coverage of key areas but also significantly improves the working efficiency of the intelligent agent cluster. New monitoring data acquired during execution is fed back to the digital twin model in real time, forming a continuously optimized data-driven cycle. This allows the model to continuously iterate and update along with the disaster evolution process, ultimately constructing a high-fidelity digital twin that evolves synchronously with the physical world, providing solid technical support for geological disaster risk assessment and early warning decision-making. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of the multimodal intelligent agent cluster control system for dynamic assessment of geological disasters according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example: Figure 1 A schematic diagram of the multimodal intelligent agent cluster control system for dynamic assessment of geological hazards of the present invention is provided. The multimodal intelligent agent cluster control system for dynamic assessment of geological hazards includes: The data acquisition module is used to collect multi-dimensional, multimodal monitoring data from the air, ground, and underground through a heterogeneous intelligent agent cluster that includes drones, ground robots, and underground sensors. The cause judgment module is used to input multimodal monitoring data into the three-dimensional digital twin model of geological disasters, and analyze the physical causes of data inconsistency based on the physical mechanism of geological disaster evolution to determine whether the data inconsistency originates from the actual disaster evolution. The risk identification module is used to identify high-risk key areas by collaboratively analyzing the critical paths of disaster evolution and the topological characteristics of multimodal data fields when the disaster originates from real disaster evolution. The task generation module is used to generate collaborative observation task plans for intelligent agent clusters targeting high-risk critical areas. The planning optimization module is used to predict the degree of uncertainty reduction in each region of the geological disaster digital twin 3D model after the execution of the collaborative observation task planning, and to weigh it with the task execution cost to re-optimize the collaborative observation task planning; The execution update module is used to control the heterogeneous intelligent agent cluster to execute the re-optimized collaborative observation task plan, and to feed back the new monitoring data acquired after execution to update the geological disaster digital twin 3D model in real time.

[0020] By utilizing a heterogeneous intelligent agent cluster comprising drones, ground robots, and underground sensors, multi-dimensional, multimodal monitoring data from the air, ground, and underground is collected. The specific implementation is as follows: During the data acquisition phase, integrated air-ground monitoring is implemented through a heterogeneous intelligent agent cluster comprising drones, ground robots, and underground sensors. The drones utilize multi-rotor platforms equipped with visible light cameras and positioning modules, flying along a pre-defined route at a speed of 5 to 8 meters per second, for example, within an altitude range of 80 to 120 meters above the ground. The visible light camera's acquisition resolution is set to, for example, 4096 × 2160 pixels, with an image capture interval set according to the flight speed, for example, one frame every 10 meters. The ground robot uses a tracked mobile platform equipped with a 2D laser scanner and an inertial measurement unit, moving along a surface survey line at a speed of, for example, 0.5 meters per second. The laser scanner's angular resolution is set to, for example, 0.25 degrees, with a scanning frequency of, for example, 10 Hz, acquiring, for example, 1440 data points per scan cycle. The underground sensors employ microelectromechanical accelerometers, buried within boreholes at depths of, for example, 3 to 15 meters, with sensor spacing set according to geological survey results, for example, 5 to 10 meters. The sensor continuously records triaxial displacement data at a sampling frequency of, for example, 1 Hz, with a range of, for example, ±10 degrees.

[0021] To achieve spatiotemporal synchronization of multi-source data, a time reference system based on the Global Positioning System (GPS) is established. Unmanned aerial vehicles (UAVs) acquire Coordinated Universal Time (UTC) by receiving GPS signals and embed timestamps into the metadata of each image frame. Ground robots synchronize with the GPS timing module via a wireless network, marking each laser-scanned point cloud data frame with a timestamp. Underground sensors are connected via wired connections to a surface data acquisition unit, which provides a unified time reference for all sensors through the GPS timing module. Time synchronization accuracy is controlled within, for example, ±100 milliseconds, ensuring the comparability of data collected from different platforms.

[0022] During data acquisition, the UAV automatically flies according to preset waypoints, recording the latitude, longitude, and altitude information of each frame of imagery in real time. A ground robot moves along a preset survey line, recording the three-dimensional coordinates of each laser scanning point using real-time positioning and mapping technology. Underground sensors continuously monitor the internal deformation of the soil and rock mass, converting analog signals into digital displacement values ​​through analog-to-digital conversion. All acquired data is transmitted to the data center via a wireless communication network, employing a hybrid communication mode combining, for example, 4G networks and ad hoc networks. In areas with poor communication signal coverage, ad hoc network relay transmission is activated.

[0023] Optical images acquired by the UAV undergo geometric and radiometric correction to eliminate lens distortion and illumination effects. Geometric correction is achieved through camera calibration parameters, such as using a checkerboard calibration method to obtain the intrinsic parameter matrix and distortion coefficients. Radiometric correction adjusts image brightness and contrast using histogram equalization. Laser point cloud data acquired by the ground robot undergoes coordinate transformation and noise filtering to generate a centimeter-level accurate 3D surface model. Coordinate transformation converts the laser scan data from the sensor coordinate system to the global coordinate system, and noise filtering employs a statistical outlier removal algorithm, such as removing discrete noise points based on point cloud density. Displacement data acquired by underground sensors undergoes temperature compensation and drift correction to eliminate the influence of environmental factors on measurement accuracy. Temperature compensation uses a built-in temperature sensor to correct displacement readings in real time, and drift correction is achieved through periodic zero-point calibration.

[0024] The resulting multimodal monitoring data forms a structured dataset under a unified spatiotemporal benchmark, providing standardized data input for subsequent analysis. In practice, the UAV's flight altitude is dynamically adjusted based on the terrain undulation of the monitoring area. For example, in areas with significant terrain undulation, the flight altitude is increased to 120 meters to ensure image overlap, while in flat areas, the flight altitude is reduced to 80 meters to improve spatial resolution. The flight altitude setting is based on ensuring image overlap of no less than 70%, calculated through feature point matching between preceding and following images. The ground robot's movement path is planned in real-time based on the distribution of obstacles on-site, and a fusion of LiDAR and visual sensors ensures safe movement. The obstacle avoidance algorithm calculates obstacle distances based on sensor data; for example, path replanning is triggered when the obstacle distance is less than 1 meter. The installation depth of underground sensors is determined based on the location of potential slip surfaces identified through geological surveys. For example, three sensors are placed above and below the potential slip surface to form a three-dimensional monitoring network. The sensor spacing is set to ensure monitoring coverage; for example, a spacing of 5 to 10 meters is set based on the deformation propagation characteristics of the soil and rock mass to capture deformation gradients.

[0025] Multiple verification mechanisms are implemented in the data quality control process. The overlap of images acquired by UAVs is no less than 70%, and image quality is verified through feature point matching, such as using the SIFT algorithm to extract and match feature points to calculate the overlap rate. Point cloud data acquired by ground robots undergoes multiple scans for consistency verification, with a repeat scan error not exceeding 2 cm, calculated using a point cloud registration algorithm. Underground sensors undergo regular zero-point drift checks, ensuring measurement accuracy by comparing against a baseline value; for example, zero-point calibration is performed weekly and compared with the initial baseline value. All acquired data is uploaded to the cloud platform in real time. Encryption protocols are used during data transmission to ensure data security, and checksum verification ensures data integrity, such as using the CRC32 algorithm to calculate the checksum.

[0026] Multimodal monitoring data is input into a digital twin 3D model of geological hazards, and the physical causes of data inconsistencies are analyzed based on the physical mechanisms of geological hazard evolution to determine whether the data inconsistencies originate from the actual evolution of the hazard. The specific implementation is as follows: In the data integration and processing phase, multimodal monitoring data is input into the 3D digital twin model of geological hazards. First, optical images, laser point cloud data, and displacement data are registered into a unified coordinate system using coordinate transformation methods. Optical images undergo orthorectification to eliminate projection distortion; for example, geometric correction is performed on the images using a digital elevation model to ensure each pixel corresponds to a specific geographic coordinate. Laser point cloud data is aligned with the 3D model using point cloud registration algorithms; for example, iterative nearest-point algorithms are used to match point cloud data to the model surface. Displacement data is mapped to the 3D geological structure framework using sensor spatial coordinates; for example, displacement monitoring points are precisely projected onto corresponding locations on the model based on borehole location data. A seven-parameter transformation model is used during coordinate transformation, including three translation parameters, three rotation parameters, and one scale parameter. These parameters are obtained through ground control point calibration; for example, at least five control points are set up in the monitoring area for parameter calculation. Coordinate transformation accuracy is ensured through residual control to guarantee quality. For example, the coordinate transformation residual is controlled within 0.1 meters. This residual threshold is set according to the measurement accuracy requirements and is specifically determined based on the spatial scale of the monitoring area. When the coordinate transformation residual exceeds the residual threshold, the parameters are recalibrated.

[0027] The construction of a 3D digital twin model of geological hazards is achieved by integrating multi-source geological data. First, a 3D geological structure framework is established based on regional geological survey data, for example, by constructing a stratigraphic interface model using borehole core data and geological profile maps. A geographic information system (GIS) base map serves as a spatial reference benchmark, and a Gauss-Kruger projection coordinate system is used to ensure spatial consistency. Optical images from multimodal monitoring data undergo orthorectification and texture mapping, for example, using digital differential correction methods to eliminate image distortion, and the corrected images are mapped as texture maps onto the surface of the 3D geological structure framework. Laser point cloud data is used to generate a digital elevation model through interpolation algorithms, such as using Kriging interpolation to convert discrete point clouds into a continuous elevation surface, which is then integrated into the 3D geological structure framework to characterize topographic changes. Within the 3D geological structure framework, corresponding monitoring points are assigned to displacement data, for example, by establishing a 3D spatial index based on sensor burial depth and planar location, forming a 3D digital twin model of geological hazards that integrates air, ground, and subsurface multi-source information. During model building, data fusion accuracy indicators are set. For example, the elevation fusion error threshold is set to 0.05 meters. This elevation fusion error threshold is determined by the point cloud data density and terrain complexity, specifically calculated based on the point cloud sampling interval and terrain undulation. When the elevation fusion error is greater than the elevation fusion error threshold, interpolation is performed again.

[0028] In the data consistency analysis phase, the spatial distribution correlation between surface deformation features recorded by optical imagery and surface elevation changes generated by laser point cloud data is compared. Surface deformation features are extracted from optical imagery, for example, by identifying surface displacement areas using image differential interferometry and calculating the area and spatial distribution of deformation areas. Topographic change features are extracted from the elevation change field generated by laser point cloud data, for example, by calculating elevation changes through comparison of multiple point cloud data and generating elevation change contour maps. Spatial correlation analysis is used to evaluate the consistency between the two data sources, for example, by calculating the area overlap between deformation areas and elevation change areas. Areas with an overlap lower than a set overlap threshold are identified as inconsistent data areas. Spatial correlation analysis employs a gridded processing method, dividing the monitoring area into regular grids, for example, using 5m × 5m grid cells, and calculating the correlation coefficient between deformation features and elevation changes within each grid. The correlation coefficient threshold is determined through statistical significance testing. For example, when the correlation coefficient is lower than the correlation coefficient threshold, it is considered that there is significant inconsistency. This correlation coefficient threshold is obtained based on the stability analysis of historical monitoring data. Specifically, it is set by calculating the data fluctuation range under normal conditions. For example, the correlation coefficient threshold is set to 0.7. When the calculated correlation coefficient is less than the correlation coefficient threshold, a data inconsistency alarm is triggered.

[0029] This study analyzes the coupling relationship between deep deformation patterns and surface deformation characteristics reflected in displacement data, focusing on their physical mechanisms. Deep deformation features are extracted from the displacement data sequence, such as analyzing the deformation rate trend through displacement-time curves to identify accelerated deformation stages. A deformation transfer model is established based on the principles of soil and rock mechanics, for example, using elasticity theory to calculate the impact range of deep deformation on the surface. The consistency between the deep displacement vector and the surface deformation direction is compared, for example, calculating the angle between the principal directions of deep displacement and surface deformation. When the angle is less than a set angle threshold, a mechanical coupling relationship is considered to exist. The coupling relationship analysis considers the physical parameters of the soil and rock mass, such as calculating the deformation transfer coefficient based on the elastic modulus and Poisson's ratio obtained from geotechnical tests. The angle threshold is determined through the mechanical properties of the soil and rock mass; for example, an angle threshold of 15 degrees is set. This angle threshold is obtained based on experimental data of soil and rock shear strength parameters and deformation propagation laws, specifically determined through a combination of direct shear tests and numerical simulations. When the calculated angle is less than the angle threshold, a coupling relationship between deep deformation and surface deformation is confirmed.

[0030] The overlap threshold is determined jointly by photogrammetry principles and the accuracy requirements of geological disaster monitoring. This overlap threshold is set at 70%, based on the following technical considerations: First, according to the requirements for generating stereo image pairs in photogrammetry, the accuracy of 3D modeling cannot be guaranteed when the overlap is below 70%. Second, considering the minimum scale requirement for deformation identification in geological disaster monitoring, incomplete deformation feature extraction occurs when the overlap is below this threshold. The specific method for obtaining the overlap threshold is through analysis of a large amount of experimental data. For example, multiple sets of image data were collected under different terrain conditions, and the feature matching success rate and 3D coordinate calculation accuracy were statistically analyzed under different overlap levels. When the overlap is below 70%, the feature matching success rate drops significantly. In practical applications, the calculated actual overlap is compared with the overlap threshold. When the actual overlap is less than the overlap threshold, a data re-acquisition process is initiated.

[0031] When the deep deformation pattern and surface deformation characteristics conform to the mechanical mechanisms of shear slip or tensile fracturing in soil and rock masses, the inconsistency in the judgment data stems from the actual evolution of the disaster. The determination of the shear slip mechanism is based on displacement vector analysis. For example, when the deep displacement vector is consistent with the potential slip surface trend and the displacement rate continuously increases, it conforms to shear slip characteristics. The determination of the tensile fracturing mechanism is based on the deformation distribution pattern. For example, when the deep displacement exhibits tensile fracturing characteristics and continuous cracks appear on the surface, it conforms to tensile fracturing characteristics. Multiple sets of discrimination indicators are set for the mechanical mechanism determination. For example, the shear slip indicators include displacement vector angle thresholds and displacement rate thresholds; the tensile fracturing indicators include crack width thresholds and crack extension rate thresholds. Threshold comparisons are implemented through logical judgment. For example, when the measured displacement rate is greater than the displacement rate threshold and the displacement vector angle is less than the displacement vector angle threshold, a shear slip judgment is triggered; when the crack width is greater than the crack width threshold or the crack extension rate is greater than the crack extension rate threshold, a tensile fracturing judgment is triggered.

[0032] The displacement vector angle threshold is determined using shear strength parameters obtained from direct shear tests on soil and rock. In these tests, the residual strength stage after the peak strength of the soil and rock is measured by applying different normal stresses. The range of dilatation angle variation for different lithological samples is statistically analyzed, and the 75th percentile of the dilatation angle distribution is taken as the displacement vector angle threshold. This threshold setting takes into account the directional deflection characteristics caused by particle rearrangement under continuous shearing, ensuring that potential sliding surface development can be effectively identified when the measured displacement vector angle is less than this threshold.

[0033] The displacement rate threshold was determined based on statistical analysis of deformation stages in historical landslide cases. Landslide cases with complete monitoring records were collected within the region, and displacement rate data during the accelerated deformation stage were extracted. The critical value of the displacement rate threshold was determined through probability distribution analysis. This threshold setting considers the critical conditions for the transition from the creep stage to the accelerated stage of the soil and rock mass, ensuring that when the measured displacement rate exceeds this threshold, it can be accurately determined that deformation has entered the accelerated stage.

[0034] The crack width threshold was determined by combining data from tensile tests on soil and rock mass with field survey data. Tensile tests were conducted in the laboratory to measure the ultimate tensile strain of samples from different lithologies. Combined with crack development data obtained from field surveys, the width boundaries between stable and dangerous cracks were statistically analyzed. This threshold setting considers the relationship between the tensile strength of the soil and rock mass and crack propagation, ensuring that the degree of tensile fracturing can be reliably determined when the measured crack width exceeds this threshold.

[0035] The crack propagation rate threshold was determined based on statistical analysis of monitoring data from the surface fracturing process. Monitoring data of surface cracks from typical tensional fracturing cases were collected, and the relationship between crack propagation rate and fracturing development stage was analyzed. A critical value for the crack propagation rate was determined through regression analysis. This threshold setting considers the correlation between the accelerated characteristics of crack propagation and the decline in the tensile strength of the soil and rock mass, ensuring that accelerated development of tensional fracturing can be identified in a timely manner when the measured crack propagation rate exceeds this threshold.

[0036] During implementation, all analysis results are updated to the digital twin model in real time; for example, identified disaster evolution areas are marked as red alert zones. The data registration phase employs a progressive refinement method, such as performing coarse registration followed by fine registration, with registration accuracy controlled through iterative calculations. Correlation analysis incorporates a dynamic adjustment mechanism, such as adaptively adjusting the grid size based on monitoring data quality, and dynamically optimizing the grid size according to point cloud density and image resolution. Coupling relationship analysis introduces weighting coefficients, such as assigning weights to different data sources based on sensor reliability; these weighting coefficients are set based on sensor accuracy calibration results and are specifically calculated by comparing the consistency of multi-source data. Mechanical mechanism determination employs a multi-evidence fusion method; for example, disaster evolution is confirmed when both displacement vector consistency and deformation rate exceedance are simultaneously satisfied. Evidence fusion is achieved through weighted voting, with weights allocated based on evidence reliability.

[0037] When the disaster originates from a real disaster evolution, high-risk key areas are identified by collaboratively analyzing the critical paths of disaster evolution and the topological characteristics of multimodal data fields. Specifically, this is implemented as follows: In the disaster evolution analysis phase, when data inconsistencies are determined to stem from actual disaster evolution, high-risk critical areas are identified through collaborative analysis of the critical paths of disaster evolution and the topological characteristics of the multimodal data field. First, spatiotemporal interpolation is performed on displacement data to generate a continuously distributed displacement field. The spatiotemporal interpolation uses the Kriging interpolation method, which describes spatial correlation based on a variogram model and calculates the displacement value of the point to be interpolated by searching for data from neighboring monitoring points. During displacement field generation, interpolation accuracy control parameters are set, such as setting the search radius to a specific multiple of the average spacing between monitoring points, for example, three times. This multiple is determined based on spatial autocorrelation analysis, obtained by calculating the ratio of the variogram range parameter to the monitoring point spacing. The interpolated displacement field is expressed in the form of a regular grid. The grid size is determined according to the monitoring point density, for example, set to a specific proportion of the average spacing between monitoring points, such as half. This proportion is determined through interpolation error testing; when the grid size is greater than half of the average spacing between monitoring points, the interpolation error exceeds the allowable range.

[0038] Regions in the displacement field whose displacement rates exceed a preset rate threshold are identified as deformation initiation zones. The displacement rate is calculated using time-series data of the displacement field, for example, by taking the ratio of the displacement difference between two consecutive time steps to the time interval. The preset rate threshold is determined through statistical analysis of historical disaster cases. Monitoring data from historical landslide cases within the region are collected, displacement rate samples from the accelerated deformation phase are extracted, and the percentile values ​​of the displacement rate samples are calculated using statistical distribution analysis methods. The 80th percentile value is used as the preset rate threshold. This percentile value is selected based on statistical analysis of the evolution stages of historical disasters to ensure reliable identification of accelerated deformation trends when the displacement rate exceeds the preset rate threshold. When the calculated displacement rate of a region in the displacement field exceeds the preset rate threshold, that region is identified as a deformation initiation zone.

[0039] The propagation direction and velocity of contour lines in the displacement field over time are tracked to determine the spatiotemporal evolution sequence of deformation extending from the deformation initiation zone to the surrounding areas as the critical path. Contour line tracking employs a moving window method, such as setting a circular search window, to track the spatial positional changes of contour lines over continuous time steps. The propagation direction is determined by the direction of movement of the contour line centroid, and the propagation velocity is calculated as the ratio of the contour line's movement distance to the time interval. Critical path extraction is based on the continuity characteristics of contour line propagation. A second overlap threshold is set to judge path continuity. The second overlap threshold is determined through experimental analysis. The stability of contour line tracking is tested under different terrain conditions, and the path breakage probability under different overlap degrees is calculated. The minimum overlap degree with a path breakage probability less than the allowable value is selected as the second overlap threshold. When the calculated overlap of contour lines in adjacent time steps is greater than the second overlap threshold, the path is considered continuous.

[0040] Key void structures characterizing potential sliding surface boundaries are extracted from the elevation change field generated from laser point cloud data. The elevation change field is generated by registering and differencing multiple periods of laser point cloud data. Point cloud registration employs an iterative nearest-point algorithm, and differencing calculates the elevation change at the same coordinate points. Topological data analysis uses persistent cohomology theory to identify persistent high Betti number regions in the elevation change field. By calculating Betti number changes at different scales, persistent void structures at multiple scales are identified. A persistence threshold is set to screen significant void structures. The persistence threshold is determined through topological feature stability analysis. The duration distribution of each void structure in the persistent cohomology map is analyzed, and the duration corresponding to significant inflection points in the duration distribution is selected as the persistence threshold. Void structures whose calculated duration exceeds the persistence threshold are identified as key void structures. Key void structures correspond to potential sliding surface boundaries with drastic elevation changes.

[0041] Spatially coupled analysis is performed on the deformation acceleration segment and the geometric discontinuity region characterized by the key void structure on the critical path. Segments on the critical path where the displacement acceleration exceeds an acceleration threshold are identified as deformation acceleration segments. The displacement acceleration is calculated using time-series displacement rate data, for example, the ratio of the difference in displacement rate between two consecutive time steps to the time interval. The acceleration threshold is determined using soil and rock creep test data. Indoor soil and rock creep tests are conducted, and strain rate changes during the accelerated creep stage are recorded. The strain rate change rate corresponding to the acceleration creep initiation point is determined through strain rate-time curve analysis, and this rate of change is used as the acceleration threshold. When the calculated displacement acceleration exceeds the acceleration threshold, the segment is identified as a deformation acceleration segment. Spatially overlay analysis is performed on the deformation acceleration segment and the geometric discontinuity region characterized by the key void structure. Spatial overlay is performed using spatial query methods in a geographic information system, such as point-polygon inclusion analysis, to determine the spatial relationship between the deformation acceleration segment and the geometric discontinuity region. Identify spatially overlapping deformation acceleration segments and geometric discontinuities, and mark these overlapping segments as high-risk critical areas that are simultaneously in the deformation acceleration stage and have significant geometric discontinuity characteristics.

[0042] During implementation, displacement field interpolation employs cross-validation to evaluate interpolation accuracy, for example, using some monitoring points as validation points to calculate interpolation error. For deformation initiation zone identification, a minimum area constraint is set, and an area threshold is used to filter out excessively small areas. The area threshold is determined statistically from the minimum deformation zones in historical disaster cases; only areas with a calculated area greater than the area threshold are confirmed as deformation initiation zones. Critical path tracing incorporates smoothing processes, such as using moving averages to eliminate contour line jitter. For critical cavity structure extraction, a minimum aperture limit is set, and a diameter threshold is used to filter out noisy cavities. The diameter threshold is determined through testing the cavity structure's ability to characterize the sliding surface boundary; cavities with a calculated diameter less than the diameter threshold are filtered out. Spatial coupling analysis employs buffer analysis to enhance spatial relationship judgment; for example, a buffer zone is established around geometrically discontinuous areas, and spatial overlap is considered present when a deformation acceleration segment intersects with the buffer zone. All analysis results are integrated into a geological disaster digital twin 3D model to provide a basis for subsequent monitoring and early warning.

[0043] Generate a collaborative observation task plan for intelligent agent clusters targeting high-risk critical areas, specifically implemented as follows: During the task planning phase, a collaborative observation task plan for intelligent agent clusters targeting high-risk critical areas is generated. First, observation priorities are determined based on the spatial distribution characteristics and deformation development stages of these areas. Spatial distribution characteristics include region area, spatial clustering degree, and boundary complexity. Region area is obtained using the polygon area calculation formula, spatial clustering degree is calculated using the spatial autocorrelation index, and boundary complexity is calculated using the boundary fractal dimension. Deformation development stages are categorized into different levels based on displacement acceleration values ​​and deformation duration. Displacement acceleration values ​​are calculated using displacement field time-series data, and deformation duration is calculated using the time difference between the deformation initiation time and the current time. An acceleration threshold is used to determine the deformation acceleration state. This threshold is determined statistically from displacement acceleration data of historical disaster cases. Accelerated deformation stage data from multiple historical disaster cases are collected, the distribution of displacement acceleration is calculated, and a specific percentile value is used as the acceleration threshold, such as the 80th percentile. A duration threshold is used to determine the deformation persistence state. This threshold is determined based on soil creep test data. The start time of the deformation acceleration stage is recorded through indoor tests, and the average value of multiple test samples is used as the duration threshold. When the displacement acceleration value exceeds the acceleration threshold and the deformation duration exceeds the duration threshold, the area is classified as a high-development stage. Observation priority is determined using a weighted scoring method, setting spatial distribution weight and deformation development weight. The spatial distribution weight is calculated based on the product of the area and spatial clustering degree, while the deformation development weight is calculated based on the product of the displacement acceleration value and the deformation duration. The final observation priority score is the weighted sum of the spatial distribution weight and the deformation development weight. The specific values ​​of the spatial distribution weight and the deformation development weight are determined through expert experience. Multiple geological hazard experts are invited to independently score the importance of each indicator, and the arithmetic mean of all expert scores is taken as the weight value.

[0044] Based on the matching relationship between observation priorities and the spatial locations and monitoring capabilities of UAVs, ground robots, and underground sensors in the agent swarm, a collaborative observation task plan is generated, including flight trajectories, movement paths, and sampling points. Spatial locations are acquired in real time via the Global Positioning System (GPS), and monitoring capabilities include coverage, accuracy indicators, and response time. The matching relationship is established through a multi-objective optimization model, with objective functions including maximizing the overall observation priority score, minimizing the overall movement distance, and minimizing the overall response time. Model constraints include coverage of each high-risk critical area, constraints on the number of agents, and constraints on agent capabilities. Flight trajectory planning is solved using a genetic algorithm, setting the population size, number of iterations, and crossover / mutation probability. The population size is determined experimentally, for example, by setting it to a specific multiple of the number of high-risk critical areas. The number of iterations is determined based on the required accuracy, and the crossover / mutation probability is determined through parameter tuning. Movement path planning uses an ant colony algorithm, setting initial pheromone values ​​and evaporation coefficients. The initial pheromone values ​​are determined based on observation priorities, and the evaporation coefficient is determined through path planning performance experiments. The sampling point layout is based on spatial sampling theory. Sampling points are arranged in a grid in each high-risk critical area. The grid size is determined according to the area and monitoring accuracy requirements. For example, the larger the area, the larger the grid size, and the higher the monitoring accuracy requirements, the smaller the grid size.

[0045] The UAV undertakes the task of acquiring large-scale optical images. Flight trajectory planning considers flight altitude, lateral overlap, and side overlap. Flight altitude is determined based on ground sampling distance requirements; for example, a specific flight altitude is set when the ground sampling distance is required to be less than a certain value. Forward and side overlap are determined by photogrammetric requirements; for example, lateral overlap is set to a specific percentage, and side overlap is set to a specific percentage. The ground robot is responsible for the detailed acquisition of laser point cloud data. Its movement path planning considers terrain accessibility and sampling point coverage. Terrain accessibility is calculated using a digital elevation model (DEM) to determine slope value. A slope threshold is used to judge terrain accessibility and is determined through robot mobility testing. The robot's mobility performance at different slopes is tested in the laboratory, and the maximum slope at which stable movement is achieved is taken as the slope threshold. Accessibility is considered to be possible when the slope value is less than the slope threshold. Sampling point coverage is calculated using the robot's sensor field of view to ensure that each sampling point is covered. The underground sensor performs continuous displacement data monitoring. The sampling point layout considers geological structural features and the location of potential slip surfaces. Geological structural features are obtained through geological survey data, and the location of potential sliding surfaces is determined through geotechnical analysis. Sensors are deployed above and below the potential sliding surfaces to form a three-dimensional monitoring network. The task planning of all agents is coordinated through a unified timing sequence to ensure the synchronicity and integrity of data acquisition.

[0046] The degree of uncertainty reduction in each region of the geological hazard digital twin 3D model after the collaborative observation mission planning is predicted, and the collaborative observation mission planning is re-optimized by weighing the mission execution cost. The specific implementation is as follows: In the prediction and optimization phase, the degree of uncertainty reduction in each region of the geological disaster digital twin 3D model after the implementation of the collaborative observation task plan is predicted, and the cost of task execution is weighed to re-optimize the collaborative observation task plan. First, based on the current uncertainty distribution in each region of the geological disaster digital twin 3D model, the effect of adding optical images, laser point cloud data, and displacement data on uncertainty reduction after the implementation of the collaborative observation task plan is simulated. The current uncertainty distribution is obtained by calculating the statistical variance of historical monitoring data. For example, for displacement data, the time-series variance of the displacement value at each monitoring point is calculated as an uncertainty index. The simulation of uncertainty reduction effect uses a data assimilation method, updating the model state with newly added monitoring data as observation values. For example, using the Kalman filter algorithm, surface deformation data extracted from optical images, elevation change data generated from laser point cloud data, and deep displacement data collected by displacement sensors are used as observation inputs. The uncertainty reduction amount is calculated through state transition equations and observation equations. The uncertainty reduction effect is quantified by the percentage reduction in uncertainty. For example, the ratio of the uncertainty variance after assimilation to the uncertainty variance before assimilation is calculated. When this ratio is less than 1, it indicates that the uncertainty has been reduced.

[0047] The assessment evaluates the resource consumption of the agent swarm required to execute the collaborative observation mission plan, including flight time, travel distance, and energy consumption. Flight time is calculated as the ratio of the UAV's flight path length to its cruising speed. The cruising speed is determined based on the UAV's performance parameters; for example, the typical cruising speed for a multi-rotor UAV is 5 to 10 meters per second. Travel distance is calculated as the total length of the ground robot's movement path, taking into account path corrections due to terrain undulations. Energy consumption is calculated as the product of each agent's power consumption and operating time. Power consumption is determined based on the agent model's technical specifications; for example, the UAV's motor power is 200 to 500 watts, the ground robot's drive power is 50 to 100 watts, and the underground sensor circuit power consumption is 5 to 10 watts. A resource consumption threshold is set to determine mission feasibility. This threshold is determined by the overall resource constraints of the agent swarm; for example, the maximum allowable operating time is calculated based on the total battery capacity and average power consumption of all agents, and then converted into an equivalent resource consumption threshold. The mission is deemed infeasible when the predicted resource consumption exceeds the threshold.

[0048] Based on the ratio of uncertainty reduction effect to resource consumption, the task allocation and execution order of UAVs, ground robots, and underground sensors in the collaborative observation task planning are adjusted. The ratio is calculated in the form of benefit-cost ratio, with uncertainty reduction effect as the benefit value and resource consumption as the cost value. The adjustment process is based on a multi-objective optimization method, setting benefit weights and cost weights. The benefit weight is determined based on the geological disaster risk level; for example, high-risk areas are assigned higher benefit weights. The cost weight is determined based on resource scarcity; for example, when resources are scarce, higher cost weights are assigned. Task allocation adjustment is achieved by reallocating the monitoring areas of each agent; for example, areas with higher uncertainty reduction effects are assigned to agents with faster response times. Execution order adjustment is achieved through task scheduling algorithms; for example, priority scheduling algorithms are used to prioritize tasks with higher benefit-cost ratios. An adjustment threshold is set during the adjustment process to trigger replanning. The adjustment threshold is determined through statistical analysis of historical task execution effects; for example, data on the decrease in benefit-cost ratio in historical tasks is collected, and the average decrease is calculated as the adjustment threshold. Replanning is initiated when the current benefit-cost ratio decrease exceeds the adjustment threshold.

[0049] In practical implementation, the uncertainty reduction effect simulation considers data quality factors, such as the impact of optical image resolution and coverage on uncertainty reduction. This is corrected by setting data quality weights, which are determined based on sensor accuracy and acquisition conditions. Resource consumption assessment considers dynamic factors, such as the impact of wind speed on UAV flight time, and flight time calculations are corrected in real time using environmental sensor data. Task allocation adjustment employs iterative optimization methods, such as using genetic algorithms to solve for the optimal task allocation scheme, setting the fitness function as the overall benefit-cost ratio, and gradually optimizing through selection, crossover, and mutation operations. Execution order adjustment considers task dependencies; for example, some areas require surface monitoring before subsurface monitoring, and the execution order is determined using a task dependency graph. All adjustment results are fed back to the collaborative observation task planning, forming a closed-loop optimization to ensure maximum uncertainty reduction effect under resource constraints.

[0050] The heterogeneous intelligent agent cluster is controlled to execute a re-optimized collaborative observation task plan, and the newly acquired monitoring data is fed back in real time to update the three-dimensional digital twin model of geological hazards. The specific implementation is as follows: During the execution and control phase, the heterogeneous intelligent agent cluster executes the re-optimized collaborative observation task plan and updates the geological disaster digital twin 3D model in real time with the newly acquired monitoring data. First, the re-optimized collaborative observation task plan's flight trajectory, movement path, and sampling point instructions are sent to the UAV, ground robot, and underground sensors via a wireless communication network. The wireless communication network employs multi-hop ad hoc networking technology, ensuring network connectivity by setting a communication distance threshold. This threshold is determined based on the wireless signal transmission characteristics; the maximum reliable communication distance is calculated according to the wireless communication module's transmit power and receive sensitivity. A specific proportion of this maximum reliable communication distance is used as the communication distance threshold. This specific proportion is determined through on-site signal testing, such as measuring signal strength attenuation curves under different terrain conditions, and selecting the maximum distance proportion that ensures stable communication as the communication distance threshold. When the distance between intelligent agents exceeds the communication distance threshold, relay transmission is automatically initiated. Command transmission employs encryption protocols to ensure data security, such as using the AES encryption algorithm to encrypt command data. The command sending is set to confirm the mechanism. If no confirmation response is received from the agent within a set time, the command will be retransmitted. The number of retransmissions is determined by communication reliability testing, such as testing the communication success rate under typical terrain conditions, and setting the maximum number of retransmissions based on the test results.

[0051] The system receives real-time optical images collected by UAVs, real-time laser point cloud data collected by ground robots, and real-time displacement data collected by underground sensors. Data transmission employs compression algorithms to reduce bandwidth consumption; for example, JPEG compression is used for optical images, and octree compression is used for laser point cloud data. A verification mechanism is implemented for data reception, such as using cyclic redundancy check (CRC) to detect data transmission errors and requesting retransmission when an error is detected. A timeout threshold is set to determine data transmission anomalies. This threshold is determined based on network latency and data volume. The average network latency is calculated using historical transmission data, and the expected transmission time is calculated based on the current data volume. A specific multiple of the expected transmission time is used as the timeout threshold. This specific multiple is determined through statistical analysis of transmission reliability; for example, the distribution of the ratio of successful transmission time to expected transmission time in historical transmission tasks is analyzed, and the 90th percentile value is used as the specific multiple. When the data transmission time exceeds the timeout threshold, it is considered a transmission anomaly, and a backup transmission channel is activated.

[0052] Real-time optical imagery, real-time laser point cloud data, and real-time displacement data are fused with monitoring data from corresponding spatiotemporal locations in a digital twin 3D model of geological hazards to update the deformation field distribution and uncertainty parameters in the model. A weighted fusion method is used, assigning fusion weights to different data sources. These weights are determined based on data quality indicators, including resolution, accuracy, and timeliness. For example, the fusion weight for optical imagery is determined by image resolution and acquisition time, with resolution and timeliness weights set. The resolution weight is inversely proportional to the image resolution, and the timeliness weight is inversely proportional to the acquisition time difference. The final fusion weight is the product of these two weights. The fusion weight for laser point cloud data is determined by point cloud density and acquisition time, while the fusion weight for displacement data is determined by sensor accuracy and acquisition time. The weight calculation method is similar to that for optical imagery. Deformation field distribution updates employ data assimilation methods, such as using an ensemble Kalman filter algorithm, to update the model's state variables using new monitoring data as observations. Uncertainty parameter updates are achieved by calculating the variance after fusion, for example, using a Bayesian update formula to calculate the posterior variance. The model update setting threshold is used to determine whether to update the model. The update threshold is determined based on the degree of data change. The relative difference between the new monitoring data and the model prediction value is calculated. When the relative difference is greater than a set percentage, the model update is triggered. This set percentage is determined by the model accuracy requirements. For example, if the model prediction error is required to be controlled within 5%, the update threshold is set to 5%.

[0053] In practical implementation, command transmission considers network load balancing, such as dynamically selecting transmission paths based on the location and communication quality of each agent. Data reception employs a caching mechanism, such as setting up a data buffer at the receiving end to temporarily store data when the data reception rate exceeds the processing rate. Data fusion considers spatiotemporal alignment, such as using spatiotemporal interpolation methods to unify data with different spatiotemporal resolutions to the same spatiotemporal grid. Model updates adopt an incremental update approach, such as updating only the model parameters in the changed regions. The entire execution control process monitors the status of each agent in real time, such as monitoring the online status of agents through a heartbeat mechanism, setting a heartbeat timeout threshold to determine agent abnormalities. The heartbeat timeout threshold is determined based on network conditions and the criticality of the task; when the number of consecutive lost heartbeats exceeds the heartbeat timeout threshold, the agent is considered abnormal. All execution statuses and update results are recorded in a log for subsequent analysis and optimization.

[0054] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0055] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0056] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wireless or wired transmission; wired transmission methods include optical fiber, twisted pair, coaxial cable, etc.; wireless transmission includes infrared, microwave, etc. Computer-readable storage media can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0057] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0058] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0059] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0060] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0061] If a function is implemented as a software module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0062] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0063] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multimodal intelligent agent cluster control system for dynamic assessment of geological disasters, characterized in that, include: The data acquisition module is used to collect multi-dimensional, multimodal monitoring data from the air, ground, and underground through a heterogeneous intelligent agent cluster that includes drones, ground robots, and underground sensors. The cause judgment module is used to input multimodal monitoring data into the three-dimensional digital twin model of geological disasters, and analyze the physical causes of data inconsistency based on the physical mechanism of geological disaster evolution to determine whether the data inconsistency originates from the actual disaster evolution. The risk identification module is used to identify high-risk key areas by collaboratively analyzing the critical paths of disaster evolution and the topological characteristics of multimodal data fields when the disaster originates from real disaster evolution. The task generation module is used to generate collaborative observation task plans for intelligent agent clusters targeting high-risk critical areas. The planning optimization module is used to predict the degree of uncertainty reduction in each region of the geological disaster digital twin 3D model after the execution of the collaborative observation task planning, and to weigh it with the task execution cost to re-optimize the collaborative observation task planning; The execution update module is used to control the heterogeneous intelligent agent cluster to execute the re-optimized collaborative observation task plan, and to feed back the new monitoring data acquired after execution to update the geological disaster digital twin 3D model in real time.

2. The multimodal intelligent agent cluster control system for dynamic assessment of geological disasters according to claim 1, characterized in that, By utilizing a heterogeneous intelligent agent cluster comprising drones, ground robots, and underground sensors, multi-dimensional, multimodal monitoring data is collected from the air, ground, and underground, including: By using drones to fly in the airspace and collect optical images; The ground robot moves across the earth and collects laser point cloud data. Displacement data is collected by burying underground sensors inside the disaster body; Data acquisition operations from drones, ground robots, and underground sensors are synchronized via wireless communication networks based on a unified time reference to generate spatiotemporally aligned multimodal monitoring data.

3. The multimodal intelligent agent cluster control system for dynamic assessment of geological disasters according to claim 1, characterized in that, Multimodal monitoring data is input into a digital twin 3D model of geological hazards, and the physical causes of data inconsistencies are analyzed based on the physical mechanisms of geological hazard evolution. This analysis determines whether the data inconsistencies stem from the actual evolution of the hazard, including: Optical images, laser point cloud data, and displacement data are jointly registered into the unified coordinate system of the digital twin 3D model of geological hazards; The correlation between the spatial distribution of surface deformation features recorded by optical imagery and surface elevation changes generated by laser point cloud data was compared. The coupling relationship between the deep deformation patterns reflected by displacement data and the surface deformation characteristics in terms of physical mechanisms was analyzed. When the deep deformation pattern and the surface deformation characteristics conform to the mechanical mechanism of shear slip or tensile fracture of rock and soil, it is determined that the inconsistency in the data stems from the actual evolution of the disaster.

4. The multimodal intelligent agent cluster control system for dynamic assessment of geological disasters according to claim 3, characterized in that, The three-dimensional digital twin model of geological hazards is constructed in the following ways: integrating regional geological survey data and geographic information system base maps to construct a three-dimensional geological structure framework; orthorectifying and texturing optical images from multimodal monitoring data onto the surface of the three-dimensional geological structure framework; generating a digital elevation model from laser point cloud data through interpolation algorithms and fusing it into the three-dimensional geological structure framework to characterize topographic changes; and assigning displacement data to corresponding underground monitoring points within the three-dimensional geological structure framework, thereby forming a three-dimensional digital twin model of geological hazards that integrates multi-source information from air, ground, and underground sources.

5. The multimodal intelligent agent cluster control system for dynamic assessment of geological disasters according to claim 1, characterized in that, When the disaster originates from real-world disaster evolution, high-risk key areas are identified through collaborative analysis of the critical paths of disaster evolution and the topological characteristics of multimodal data fields, including: Spatiotemporal interpolation is performed on displacement data to generate a continuously spatiotemporally distributed displacement field; The region in the displacement field whose displacement rate exceeds a preset rate threshold is identified as the deformation initiation zone; By tracking the propagation direction and speed of isohyets in the displacement field over time, the spatiotemporal evolution sequence of deformation from the deformation initiation zone to the surrounding area is determined as the key path; Extract key hole structures characterizing potential sliding surface boundaries from the elevation change field generated from laser point cloud data; By spatially coupling the deformation acceleration segment on the critical path with the geometric discontinuity region characterized by the critical hole structure, high-risk critical areas that are simultaneously in the deformation acceleration stage and have significant geometric discontinuity characteristics are identified.

6. The multimodal intelligent agent cluster control system for dynamic assessment of geological disasters according to claim 5, characterized in that, Extracting key void structures representing potential sliding surface boundaries from the elevation change field generated from laser point cloud data includes: registering and differencing multi-period laser point cloud data to generate an elevation change field; using topological data analysis methods to identify persistent high Betti number regions in the elevation change field; and extracting key void structures represented by persistent high Betti number regions, which correspond to potential sliding surface boundaries with drastic elevation changes.

7. The multimodal intelligent agent cluster control system for dynamic assessment of geological disasters according to claim 5, characterized in that, The spatial coupling analysis of the deformation acceleration segment and the geometric discontinuity region characterized by the key hole structure on the critical path includes: identifying segments on the critical path where the displacement acceleration exceeds the acceleration threshold as deformation acceleration segments; performing spatial overlay analysis of the deformation acceleration segment and the geometric discontinuity region characterized by the key hole structure; identifying spatially overlapping deformation acceleration segments and geometric discontinuity regions, and marking the overlapping deformation acceleration segments and geometric discontinuity regions as high-risk key areas that are simultaneously in the deformation acceleration stage and have significant geometric discontinuity characteristics.

8. The multimodal intelligent agent cluster control system for dynamic assessment of geological disasters according to claim 1, characterized in that, Generate a collaborative observation task plan for agent clusters targeting high-risk critical areas, including: Observation priorities are determined based on the spatial distribution characteristics and deformation development stages of high-risk key areas. Based on the matching relationship between observation priority and the spatial location and monitoring capabilities of UAVs, ground robots and underground sensors in the intelligent agent cluster, a collaborative observation task plan including flight trajectory, movement path and sampling point is generated. Among them, drones are responsible for large-scale optical image acquisition, ground robots are responsible for fine acquisition of laser point cloud data, and underground sensors perform continuous displacement data monitoring.

9. The multimodal intelligent agent cluster control system for dynamic assessment of geological disasters according to claim 1, characterized in that, The degree of uncertainty reduction in each region of the geological hazard digital twin 3D model after the implementation of the collaborative observation mission plan is predicted, and the cost of mission execution is weighed to re-optimize the collaborative observation mission plan, including: Based on the current uncertainty distribution in each region of the digital twin 3D model of geological disasters, the simulation shows the effect of reducing uncertainty by adding optical images, laser point cloud data and displacement data after the collaborative observation task planning is carried out. Assess the resource consumption of the agent swarm required to execute the collaborative observation mission plan, including flight time, travel distance, and energy consumption; Based on the ratio of uncertainty reduction effect to resource consumption, adjust the task allocation and execution order of UAVs, ground robots and underground sensors in the collaborative observation mission planning.

10. The multimodal intelligent agent cluster control system for dynamic assessment of geological disasters according to claim 1, characterized in that, The system controls a cluster of heterogeneous intelligent agents to execute a re-optimized collaborative observation task plan, and feeds back the newly acquired monitoring data in real time to update the three-dimensional digital twin model of geological hazards, including: The re-optimized collaborative observation mission plan's flight trajectory, movement path, and sampling point instructions are sent to the drone, ground robot, and underground sensor respectively via wireless communication network. It receives real-time optical images collected by drones, real-time laser point cloud data collected by ground robots, and real-time displacement data collected by underground sensors. By fusing real-time optical images, real-time laser point cloud data, and real-time displacement data with monitoring data of corresponding spatiotemporal locations in the digital twin 3D model of geological hazards, the deformation field distribution and uncertainty parameters in the digital twin 3D model of geological hazards are updated.

Images