An air-ground collaborative risk prediction and dynamic electronic fence construction method of a fusion generative large model

Abstract

This invention discloses a method for air-ground collaborative risk prediction and dynamic electronic fence construction based on a generative large-scale model, comprising: obtaining a standardized air-ground collaborative multimodal spatiotemporal dataset; generating a structured state representation for risk inference; outputting risk semantic interpretation, candidate risk links, and risk factor embedding representations; constructing a spatiotemporal risk field and performing forward prediction based on the risk factor embedding representation and the air-ground collaborative state map, obtaining a risk probability heatmap and uncertainty map within the future prediction time window; generating a three-dimensional dynamic electronic fence set, and assigning a confidence weight and an effective timestamp to each fence boundary; and performing event-triggered iterative updates to the air-ground collaborative state map, risk probability heatmap, and three-dimensional dynamic electronic fence set based on real-time new data. In dynamic scene simulation, this invention reduces the frequency of false fence triggering compared to periodic recalculation methods, while maintaining a significant decrease in risk omission rate.

Description

Technical Field

[0001] This invention relates to the field of electronic fence technology, and in particular to a method for air-ground collaborative risk prediction and dynamic electronic fence construction based on a fusion generative large model. Background Technology

[0002] With the gradual opening of low-altitude airspace, the large-scale application of drones, and the widespread deployment of ground-based intelligent agents, air-ground collaborative operations have become an important technological form for smart city management, emergency rescue, security patrols, and complex scenario control. In multi-entity collaborative operation environments, there are spatial, communication, and task coupling relationships between aerial and ground platforms, resulting in highly dynamic and time-varying system operation. Simultaneously, scenarios often simultaneously contain dynamic targets, temporary obstacles, sources of emergencies, and unstable communication links, making risk factors multi-source, multi-scale, and highly uncertain.

[0003] In existing technologies, most research on the safety of air-ground collaborative operations adopts rule-based static safety zone delineation methods, which restrict the platform's activity range by setting up no-fly zones, no-entry zones, or fixed electronic fences. However, such technologies usually rely on manually set geometric boundaries or simple threshold judgments, making it difficult to dynamically update the boundary shape based on real-time sensing data. When risk factors change rapidly or emergencies occur, the fence range cannot be adaptively adjusted, which can easily lead to problems such as overly conservative approaches that restrict missions or insufficient boundaries that expose risks.

[0004] In terms of risk identification, existing technologies mostly rely on single-modal sensing data or simple fusion strategies to independently analyze the state information of air and ground platforms, lacking a unified collaborative state modeling framework. In complex scenarios with multiple subjects and multiple nodes, existing methods often fail to establish a structured relationship model that includes air platforms, ground platforms, infrastructure, and risk sources, making it difficult to characterize the relative motion relationships, communication and cooperation relationships, and spatial proximity relationships between nodes. As a result, risk judgment remains at the local level and cannot form system-level risk reasoning results.

[0005] With the development of generative artificial intelligence technology, although some studies have attempted to use large models for scene understanding or event analysis, existing applications are mostly concentrated at the semantic interpretation or text generation level, failing to deeply couple semantic reasoning results with structured spatiotemporal data. In the risk prediction process, the lack of an alignment mechanism between semantic output and numerical model makes it difficult to directly use the generated results for risk quantification modeling. At the same time, existing technologies lack effective constraints on the credibility of large model outputs and have not established a verification mechanism combined with perceptual confidence and uncertainty propagation, which easily introduces semantic drift or inconsistency in reasoning.

[0006] In the risk prediction stage, traditional methods often use single trajectory extrapolation or probability models based on historical statistics to make short-term predictions of future states, but they lack the ability to model the comprehensive propagation of multi-source uncertainties. When uncertainty increases significantly, existing technologies often fail to establish a mechanism that links with the dynamic adjustment of security boundaries, resulting in a lack of coordinated updating between prediction results and fencing strategies. Summary of the Invention

[0007] One objective of this invention is to propose a method for air-ground collaborative risk prediction and dynamic electronic fence construction that integrates generative large models. In dynamic scene simulation, this invention reduces the frequency of fence false triggering compared to periodic recalculation, while maintaining a significant decrease in risk omission rate.

[0008] A method for air-ground collaborative risk prediction and dynamic electronic fence construction based on a fusion generative large model according to an embodiment of the present invention includes: Raw data of the air-ground collaborative scenario is acquired by sensors on air and ground platforms. Time synchronization, coordinate unification and quality screening are performed on the raw data to obtain a standardized air-ground collaborative multimodal spatiotemporal dataset. The standardized air-ground collaborative multimodal spatiotemporal dataset is mapped onto the air-ground collaborative state graph to construct a heterogeneous state graph containing air platform nodes, ground platform nodes, infrastructure nodes, target object nodes and risk source nodes. Edge connections are constructed based on relative motion relationships, spatial proximity relationships and communication cooperation relationships to generate a structured state representation for risk reasoning. The structured state representation and task constraint information are combined to form a multimodal cue sequence, which is then input into a generative large model to output risk semantic interpretation, candidate risk links, and risk factor embedding representation. Based on risk factor embedded representation and air-ground collaborative state map, a spatiotemporal risk field is constructed and forward prediction is performed to obtain a risk probability heat map and uncertainty map within the future prediction time window. Based on the risk probability heatmap, uncertainty map and task constraint information, the fence optimization objectives and constraints are constructed. Constraint optimization is performed on the fence boundaries to generate a three-dimensional dynamic electronic fence set. Each fence boundary is assigned a confidence weight and an effective timestamp. Based on real-time new data, the air-ground collaborative status map, risk probability heat map, and three-dimensional dynamic electronic fence set are updated in an event-triggered iterative manner to form a versioned fence strategy and distribute it to the control execution terminals of the air platform and the ground platform to restrict the platform from entering high-risk areas and automatically adjust the fence range when the risk changes.

[0009] Optionally, the process of performing time synchronization, coordinate unification, and quality screening on the original data includes: Perform sensor intrinsic and extrinsic parameter calibration and distortion correction on visible light images, infrared images, laser point clouds and inertial navigation data acquired by the aerial platform, and perform measurement consistency verification on millimeter-wave radar measurements, vehicle positioning information and road semantic information acquired by the ground platform. Based on a unified time reference, timestamp alignment is performed on data from each modality, and interpolation compensation and frame loss repair strategies are used to align asynchronous data to the same sampling period. Based on a unified geographic coordinate system, a coordinate transformation chain is established between aerial platform coordinates, ground platform coordinates and map coordinates. Coordinate reprojection is performed on each frame of data to obtain the target position, speed and heading under unified coordinates. Perform outlier removal and confidence level weighting on the reprojected data; Outlier removal includes outlier removal based on Mahalanobis distance and velocity mutation suppression based on kinematic constraints. Confidence weighting includes generating modal confidence weights based on sensor field of view, occlusion rate, and signal-to-noise ratio. The multimodal data, after being weighted by confidence, is aggregated into spatiotemporal data blocks according to time series. The platform state vector, target object state vector, and risk source attribute vector are recorded simultaneously in the spatiotemporal data blocks to obtain a standardized air-ground collaborative multimodal spatiotemporal dataset.

[0010] Optionally, the construction of the air-ground cooperative state diagram includes: Define a node feature vector for each type of node; Define an edge feature vector for each type of edge; Heterogeneous message passing updates are performed based on node feature vectors and edge feature vectors, outputting a structured state representation.

[0011] Optionally, the construction of the multimodal cue sequence includes: The air-ground collaborative state diagram is encoded into a serializable structured description, and the task constraint information is encoded into a set of rule fragments; The structured description is enhanced based on the risk knowledge base to obtain candidate risk templates and treatment templates that match the current risk type. The candidate risk templates and treatment templates are then inserted into the prompt sequence to constrain the semantic range of the generative large model output. The prediction time window, spatial resolution, output format, and confidence expression method are explicitly labeled in the prompt sequence, so that the risk semantic interpretation output by the generative large model can be aligned with the numerical prediction results of the subsequent risk prediction steps. When the output of the generative large model contains statements that conflict with the task constraints, a conflict resolution process is triggered to form a set of candidate risk links that meet the constraints.

[0012] Optionally, the generation of the risk factor embedding representation includes: Multimodal alignment encoding is performed on the standardized air-ground collaborative multimodal spatiotemporal dataset. Through cross-modal attention, image semantics, point cloud geometry, radar velocity measurement and map prior are aligned to the same feature space to obtain a unified spatiotemporal feature tensor. Introducing contrast consistency constraints during the training or adaptation phase ensures that the risk factor embedding representations obtained from the same risk event under different observation angles, heights, or lighting conditions remain close in the feature space, while also ensuring that different risk types remain separable in the feature space, thereby improving the robustness of risk semantic reasoning to complex environmental disturbances.

[0013] Optionally, the step of constructing a spatiotemporal risk field and performing forward-looking prediction based on risk factor embedding representation and air-ground collaborative state map includes: An air reachable domain is constructed based on the flight envelope constraints of the air platform, and a ground reachable domain is constructed based on the reachable domain constraints of the ground platform. Based on the ground reachability domain, multiple hypothesis trajectory extrapolation is performed on the target object and risk source within the prediction time window. The state extrapolation based on Kalman filtering and multimodal uncertainty propagation based on particle sampling are used to obtain the occupancy probability distribution at future time. Based on the occupancy probability distribution, the candidate risk links output by the generative large model are mapped to numerical risk terms, and adaptive weights are assigned to each risk term according to the risk factor embedding representation. The risk probability heatmap is obtained by weighting and combining each risk item, and the uncertainty map is output simultaneously.

[0014] Optionally, the generation of the three-dimensional dynamic electronic fence set includes: Thresholding is performed on the risk probability heatmap to obtain a set of high-risk voxels, and three-dimensional connected component analysis is performed on the high-risk voxel set to obtain a set of high-risk connected regions. For each high-risk connected region, perform boundary extraction and geometric simplification, and obtain deployable fence polylines or fence polygon patches through Douglas-Puk simplification and curvature constraint smoothing. Based on the different safety margin requirements of the aerial platform and the ground platform, a safety expansion calculation is performed on the initial fence boundary, and a three-dimensional expansion with height layer is performed on the aerial fence, so that the expansion amount is positively correlated with the value of the uncertainty diagram; The fence optimization objectives and constraints simultaneously incorporate task completion constraints and traffic continuity constraints.

[0015] Optionally, the distribution and control of the fence policy includes: The three-dimensional dynamic electronic fence set is encoded into an executable rule set, and the executable rule set is mapped to the flight path / no-fly zone instructions that can be recognized by the air platform flight control interface and the speed / path constraint instructions that can be recognized by the ground platform controller. A dual-channel verification mechanism is introduced during the distribution process. The first channel is the geometric legality verification of the fence, which at least verifies the fence self-intersection, fence holes, and conflicts between the fence and the map restricted area. The second channel is the policy consistency verification, which at least verifies the consistency between the fence rules and the task constraint information and priority passage rules. When any verification fails, a rollback strategy is triggered, which keeps the previous version of the fence unchanged, reduces the platform's running speed, increases the safety distance threshold between the platform and the target object, and requests the output of the generative large model to suggest alternative fences. The system monitors the platform's location and its relationship with the fence in real time at the execution end. When it detects that a boundary crossing condition is about to be triggered, it outputs progressive control commands in advance, which are handled in stages from prompting, deceleration, detour, to forced return / stopping.

[0016] The beneficial effects of this invention are: This invention constructs a dual-domain coupled algorithm framework of structured air-ground collaborative state diagram, generative large model semantic reasoning, and numerical risk field mapping. It realizes an alignable mapping mechanism between semantic reasoning results and spatiotemporal probability models, breaks through the problem of the separation between semantic analysis and numerical prediction in existing technologies, and realizes an interpretable mapping of causal semantic links and probability heat distribution. This enables risk prediction to not only have numerical probability results, but also have a clear semantic causal source.

[0017] This invention introduces a multi-hypothesis trajectory extrapolation and uncertainty propagation coupling mechanism in the risk prediction stage, and embeds the risk factors output by the generative large model into the representation to adaptively adjust the weights of each risk item, realizing dynamic weighted fusion of the risk field. By mapping collision risk items, boundary crossing risk items, disconnection risk items, and group aggregation risk items into independent numerical risk channels, and performing weight adjustment on each risk channel based on the risk semantic embedding vector, the risk heatmap can reflect the real risk priority in the task context. An uncertainty map is constructed and linked with the risk heatmap. When the uncertainty exceeds the threshold, spatial safety expansion is automatically performed, forming a closed-loop mechanism that couples prediction, confidence, and boundary. In dynamic scene simulation, the frequency of fence false triggering is reduced compared with the periodic recalculation method, while maintaining a significant decrease in risk omission rate. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for air-ground collaborative risk prediction and dynamic electronic fence construction based on a fusion generative large model proposed in this invention. Detailed Implementation

[0019] Example 1: Reference Figure 1A method for air-ground collaborative risk prediction and dynamic electronic fence construction that integrates generative large-scale models includes: Raw data of the air-ground collaborative scenario is acquired by sensors on air and ground platforms. Time synchronization, coordinate unification and quality screening are performed on the raw data to obtain a standardized air-ground collaborative multimodal spatiotemporal dataset. The standardized air-ground collaborative multimodal spatiotemporal dataset is mapped onto the air-ground collaborative state graph to construct a heterogeneous state graph containing air platform nodes, ground platform nodes, infrastructure nodes, target object nodes and risk source nodes. Edge connections are constructed based on relative motion relationships, spatial proximity relationships and communication cooperation relationships to generate a structured state representation for risk reasoning. The structured state representation and task constraint information are combined to form a multimodal cue sequence, which is then input into a generative large model to output risk semantic interpretation, candidate risk links, and risk factor embedding representation. Based on risk factor embedded representation and air-ground collaborative state map, a spatiotemporal risk field is constructed and forward prediction is performed to obtain a risk probability heat map and uncertainty map within the future prediction time window. Based on the risk probability heatmap, uncertainty map and task constraint information, the fence optimization objectives and constraints are constructed. Constraint optimization is performed on the fence boundaries to generate a three-dimensional dynamic electronic fence set. Each fence boundary is assigned a confidence weight and an effective timestamp. Based on real-time new data, the air-ground collaborative status map, risk probability heat map, and 3D dynamic electronic fence set are updated in an event-triggered iterative manner to form a versioned fence strategy and distribute it to the control execution terminals of the air platform and the ground platform to restrict the platform from entering high-risk areas and automatically adjust the fence range when the risk changes.

[0020] In this embodiment, the processing of original data for time synchronization, coordinate unification, and quality screening includes: Perform sensor intrinsic and extrinsic parameter calibration and distortion correction on visible light images, infrared images, laser point clouds and inertial navigation data acquired by the aerial platform, and perform measurement consistency verification on millimeter-wave radar measurements, vehicle positioning information and road semantic information acquired by the ground platform. Based on a unified time reference, timestamp alignment is performed on data from each modality, and interpolation compensation and frame loss repair strategies are used to align asynchronous data to the same sampling period. Based on a unified geographic coordinate system, a coordinate transformation chain is established between aerial platform coordinates, ground platform coordinates and map coordinates. Coordinate reprojection is performed on each frame of data to obtain the target position, speed and heading under unified coordinates. Perform outlier removal and confidence level weighting on the reprojected data; Outlier removal includes outlier removal based on Mahalanobis distance and velocity mutation suppression based on kinematic constraints. Confidence weighting includes generating modal confidence weights based on sensor field of view, occlusion rate, and signal-to-noise ratio. The multimodal data, after being weighted by confidence, is aggregated into spatiotemporal data blocks according to time series. The platform state vector, target object state vector, and risk source attribute vector are recorded simultaneously in the spatiotemporal data blocks to obtain a standardized air-ground collaborative multimodal spatiotemporal dataset.

[0021] In this embodiment, the construction of the air-ground cooperative state diagram includes: Define a node feature vector for each type of node; The node feature vectors include air platform node feature vectors, ground platform node feature vectors, and risk source node feature vectors. The air platform node feature vectors include position, altitude, speed, remaining flight time, available communication bandwidth, and payload field of view parameters. The ground platform node feature vectors include position, speed, steering angle, braking capacity, road constraint parameters, and local perception coverage. The risk source node feature vectors include risk type, risk intensity, radius of influence, and triggering conditions. Define an edge feature vector for each type of edge; The edge feature vectors include relative motion edges and communication cooperation edges; relative motion edges include relative velocity, predicted meeting time and minimum distance, spatial proximity edges include Euclidean distance, reachability connectivity and occlusion coefficient, and communication cooperation edges include link quality, latency and packet loss rate; Heterogeneous message passing updates are performed based on node feature vectors and edge feature vectors, outputting a structured state representation.

[0022] The credibility propagation constraint is introduced during the update process, which attenuates the influence weight of low-confidence nodes caused by high occlusion or low signal-to-noise ratio in the state graph. This allows the input state representation of the generative large model to simultaneously contain both factual state and credibility information, reducing the risk of misjudgment caused by false detection of a single modality.

[0023] In this embodiment, the construction of the multimodal cue sequence includes: The air-ground collaborative state diagram is encoded into a serializable structured description, and the task constraint information is encoded into a set of rule fragments; The structured description is enhanced based on the risk knowledge base to obtain candidate risk templates and treatment templates that match the current risk type. The candidate risk templates and treatment templates are then inserted into the prompt sequence to constrain the semantic range of the generative large model output. The prediction time window, spatial resolution, output format, and confidence expression method are explicitly labeled in the prompt sequence, so that the risk semantic interpretation output by the generative large model can be aligned with the numerical prediction results of the subsequent risk prediction steps. When the output of the generative large model contains statements that conflict with the task constraints, a conflict resolution process is triggered to form a set of candidate risk links that meet the constraints.

[0024] The conflict resolution process includes backfilling the conflict item into the prompt sequence and requiring the generative big model to provide alternative risk links and alternative solutions.

[0025] In this embodiment, the generation of the risk factor embedding representation includes: Multimodal alignment encoding is performed on the standardized air-ground collaborative multimodal spatiotemporal dataset. Through cross-modal attention, image semantics, point cloud geometry, radar velocity measurement and map prior are aligned to the same feature space to obtain a unified spatiotemporal feature tensor. The structured state representation of the unified spatiotemporal feature tensor and the air-ground co-location state diagram are input together into the multimodal interface of the generative large model or its external alignment encoder, and the output is the risk factor embedding representation.

[0026] Risk factor embedding representation includes risk source intensity embedding, trigger condition embedding, vulnerability embedding of affected objects, and environmental constraint embedding.

[0027] Introducing contrast consistency constraints during the training or adaptation phase ensures that the risk factor embedding representations obtained from the same risk event under different observation angles, heights, or lighting conditions remain close in the feature space, while also ensuring that different risk types remain separable in the feature space, thereby improving the robustness of risk semantic reasoning to complex environmental disturbances.

[0028] In this embodiment, a spatiotemporal risk field is constructed and forward-looking prediction is performed based on the risk factor embedding representation and the air-ground collaborative state map, including: An air reachable domain is constructed based on the flight envelope constraints of the air platform, and a ground reachable domain is constructed based on the reachable domain constraints of the ground platform. Flight envelope constraints include maximum climb rate, maximum pitch angle, minimum turning radius, and altitude limit; reachability constraints include road topology connectivity, maximum braking distance, and speed limit rules.

[0029] Based on the ground reachability domain, multiple hypothesis trajectory extrapolation is performed on the target object and risk source within the prediction time window. The state extrapolation based on Kalman filtering and multimodal uncertainty propagation based on particle sampling are used to obtain the occupancy probability distribution at future time. Based on the occupancy probability distribution, the candidate risk links output by the generative large model are mapped to numerical risk terms, and adaptive weights are assigned to each risk term according to the risk factor embedding representation. The risk probability heatmap is obtained by weighting and combining each risk item, and the uncertainty map is output simultaneously.

[0030] The uncertainty map is formed by the superposition of trajectory extrapolation uncertainty, perceived confidence uncertainty and semantic reasoning uncertainty. When the uncertainty exceeds a preset threshold, a conservative strategy is triggered to make the risk heat map undergo safe expansion processing in space.

[0031] In this embodiment, the generation of the three-dimensional dynamic electronic fence set includes: Thresholding is performed on the risk probability heatmap to obtain a set of high-risk voxels, and three-dimensional connected component analysis is performed on the high-risk voxel set to obtain a set of high-risk connected regions. For each high-risk connected region, perform boundary extraction and geometric simplification, and obtain deployable fence polylines or fence polygon patches through Douglas-Puk simplification and curvature constraint smoothing. Based on the different safety margin requirements of the aerial platform and the ground platform, a safety expansion calculation is performed on the initial fence boundary, and a three-dimensional expansion with height layer is performed on the aerial fence, so that the expansion amount is positively correlated with the value of the uncertainty diagram; The fence optimization objectives and constraints simultaneously incorporate task completion constraints and traffic continuity constraints.

[0032] The continuity constraint ensures that the topological connectivity of available passageways is not disrupted before and after fence adjustment; confidence weights and effective timestamps are assigned to the fence boundaries obtained by optimization, and the fence policy version number is output, so that the dynamic fence has traceable version management capabilities when it is continuously updated.

[0033] In this embodiment, the issuance and control of the fence policy include: The three-dimensional dynamic electronic fence set is encoded into an executable rule set, and the executable rule set is mapped to the flight path / no-fly zone instructions that can be recognized by the air platform flight control interface and the speed / path constraint instructions that can be recognized by the ground platform controller. The set of rules includes rules for prohibiting entry into a region, rules for limiting speed in a region, rules for limiting height in a region, rules for triggering alarms upon entry, and rules for handling boundary violations.

[0034] A dual-channel verification mechanism is introduced during the distribution process. The first channel is the geometric legality verification of the fence, which at least verifies the fence self-intersection, fence holes, and conflicts between the fence and the map restricted area. The second channel is the policy consistency verification, which at least verifies the consistency between the fence rules and the task constraint information and priority passage rules. When any verification fails, a rollback strategy is triggered, which keeps the previous version of the fence unchanged, reduces the platform's running speed, increases the safety distance threshold between the platform and the target object, and requests the output of the generative large model to suggest alternative fences. The system monitors the platform's location and its relationship with the fence in real time at the execution end. When it detects that a boundary crossing condition is about to be triggered, it outputs progressive control commands in advance, which are handled in stages from prompting, deceleration, detour, to forced return / stopping.

[0035] Example 2: A suspected incident of an abnormal stop by a hazardous chemical transport vehicle occurred along Binhai Avenue in a coastal new area. The area is located at the intersection of a main urban road and a road connecting to the port area. There are elevated bridge structures, underground tunnel entrances, and residential areas in the surrounding area. The open airspace altitude is below 120 meters. At the same time, a medium-sized multi-rotor UAV was dispatched to perform a high-altitude inspection mission, flying at an altitude of about 60 meters and a cruising speed of 12 m / s. A ground emergency response vehicle drove into the scene from east to west with an initial speed of 36 km / h. There were multiple social vehicles and pedestrians on the road.

[0036] At the time of the incident, the sea breeze gusts reached 8 m / s, and local building obstruction caused GNSS positioning fluctuations of 2-3 meters. The ground millimeter-wave radar had multipath reflection errors. The communication network was a 5G public network with an average latency of 28 ms, but there were instantaneous jitters of up to 75 ms under the overpass. Traditional systems in this type of scenario usually only use fixed no-fly zones and fixed ground no-entry zones. The electronic fence boundary is a preset rectangular area with an update cycle of once every 2 seconds.

[0037] In this embodiment 2, the method of the present invention is used to predict risks and generate dynamic electronic fences, and to verify its prediction accuracy, fence response speed and false alarm control capability in complex environments.

[0038] Following the incident, the drone transmitted visible light video streams and infrared images with a resolution of 1920×1080 and a frame rate of 30fps; simultaneously, the lidar point cloud density was approximately 1.2 million points per frame; the ground vehicle transmitted data showing 17 targets detected by millimeter-wave radar, with a vehicle positioning error standard deviation of 1.8 meters. The system first performed time synchronization processing on the aerial and ground data, unifying it to a 50ms time step; calibrated the drone's attitude and IMU data, completing coordinate unification; and performed Mahalanobis distance outlier removal on the millimeter-wave radar, eliminating 4 false detection targets, accounting for 23.5% of the original target count.

[0039] After data standardization, the system constructs an air-ground collaborative state diagram. The state diagram includes one air platform node, one ground platform node, 14 social vehicle nodes, five pedestrian nodes, and one risk source node (an abnormally parked hazardous materials vehicle). The system calculates a total of 48 relative motion edges, 2 communication cooperation edges, and 34 spatial proximity edges. Through an edge type gating mechanism, the weight of low-confidence spatial edges caused by GNSS fluctuations is automatically reduced, with an attenuation ratio of approximately 0.32.

[0040] The structured state diagram is encoded into a cue sequence and input into a generative large model. The model version is an internally adapted multimodal 7B parameter model, with an input length of approximately 3600 tokens. The generative large model outputs the following risk semantic interpretation: An abnormal vehicle poses a potential risk of combustion and explosion. If surrounding vehicles continue to approach and are not evacuated in time, a high-risk situation of group aggregation will form within the next 90-150 seconds. At the same time, if the aerial platform continues to hover at low altitude, it may be at risk of falling due to a sudden explosion blast wave. The model generates 3 candidate risk links and outputs corresponding confidence scores: 0.81, 0.74, and 0.62.

[0041] The system maps risk links to numerical risk items, with the initial weights of collision risk (0.35), boundary crossing risk (0.21), group aggregation risk (0.29), and loss of connection risk (0.15). After adaptive adjustment by embedding risk factors, the weights become 0.28, 0.19, 0.41, and 0.12, respectively, with group aggregation risk being elevated to the dominant risk item.

[0042] In the risk prediction phase, the system employs multi-hypothesis trajectory extrapolation. Kalman extrapolation is performed on 14 vehicles, and particle sampling is used to generate 50 trajectory hypotheses for each vehicle, resulting in a total of 700 trajectory samples. The prediction time window is 180 seconds into the future, with a time resolution of 0.5 seconds. The risk probability heatmap has a resolution of 0.5-meter voxel grids, with a total of approximately 160,000 voxels. The system simultaneously generates an uncertainty map, with an average trajectory extrapolation variance of 3.2 m² and a perceived confidence decay coefficient of 0.87.

[0043] At the 70th second of prediction, the risk heatmap showed a peak risk probability of 0.76 and an uncertainty value of 0.14 within an 18-meter radius of the abnormal vehicle; at the 110th second, the peak risk increased to 0.83. The system automatically triggered the generation of a 3D electronic fence. First, the risk heatmap was segmented with a threshold of 0.6, resulting in a set of approximately 3800 high-risk voxels; 3D connected component analysis yielded two connected regions; an initial fence patch with 12 vertices was generated using the α-Shape algorithm; this was then reduced to 8 vertices through curvature smoothing; and Minkowski dilation was performed based on the uncertainty graph, with a dilation distance of 2.3 meters.

[0044] The generated 3D dynamic electronic fence has a volume of approximately 2350 cubic meters and a ground projection area of ​​approximately 420 square meters, with an effective timestamp of 15:32:45. The fence strategy is sent to the control terminals of the drone and ground vehicles after dual-channel verification. The drone automatically increases its altitude to 85 meters and adjusts its flight path, while the ground emergency vehicle slows down to 15 km / h and changes to a detour route.

[0045] During execution, the system monitors risk changes in real time. At 15:33:20, a civilian vehicle suddenly drove against traffic and approached the risk area. Within 120ms, the system detected that the risk peak drift exceeded the threshold of 0.12, triggering an event update. The new fence reconstruction time was 310ms, which is approximately 84.5% shorter than the traditional 2-second cycle update method.

[0046] The entire incident was brought under control at 15:36:10, with no explosion occurring. The drone and ground vehicles did not enter the high-risk area.

[0047] To verify the effectiveness, the method of this invention was compared with the traditional fixed-rule fencing method. The test data came from real air-ground collaborative samples collected from January 20xx to June 20xx, totaling 18,600 training samples, including 4,200 high-risk scenario samples, 6,100 medium-risk samples, and 8,300 low-risk samples; and 2,400 validation samples.

[0048] Regarding the accuracy of risk prediction, within a 120-second prediction window, the average convergence time error of the method of this invention is 2.4 seconds, while that of the traditional method is 3.1 seconds; the risk peak position error is 2.1 meters, while that of the traditional method is 3.7 meters; and the overall risk prediction F1 value is 0.87, while that of the traditional method is 0.73.

[0049] In terms of fence generation efficiency, the average fence generation time of the method of this invention is 0.48 seconds, while that of the traditional periodic update method is 2.1 seconds; the fence false trigger rate is 3.6%, while that of the traditional method is 7.9%; and the fence area redundancy rate is 14.8%, while that of the traditional fixed fence is 29.4%.

[0050] In the scenario simulation test, a total of 50 high-dynamic environment simulation scenarios were constructed, and GNSS error disturbances and communication jitter were added. The method of this invention still maintains the consistency of the fence policy version under the condition of communication jitter, and there is no phenomenon of old policy covering new policy. The traditional method shows boundary lag or coverage conflict in 12% of the test scenarios.

[0051] In a real-world comparative test, a field demonstration was conducted in an industrial park. The test lasted for 2 hours and triggered 9 risk events. The average response time for fence updates using the method of this invention was 280ms, while the average response time for the traditional system was 1.96 seconds. The number of times drones entered the risk area was 0, while the number of times the traditional system entered the risk area was 2. The average additional mileage for ground vehicles to detour increased by 8.2%, while the number of times the traditional system entered the risk area was 17.6%.

[0052] In the semantic reasoning and numerical model alignment test, 120 sets of sudden event samples were selected for verification. The risk link generated by the method of this invention has a matching rate of 91.3% with the actual event type, while the traditional rule-based matching method has a matching rate of 76.5%.

[0053] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for air-ground collaborative risk prediction and dynamic electronic fence construction integrating a generative large-scale model, characterized in that, include: Raw data of the air-ground collaborative scenario is acquired by sensors on air and ground platforms. Time synchronization, coordinate unification and quality screening are performed on the raw data to obtain a standardized air-ground collaborative multimodal spatiotemporal dataset. The standardized air-ground collaborative multimodal spatiotemporal dataset is mapped onto the air-ground collaborative state graph to construct a heterogeneous state graph containing air platform nodes, ground platform nodes, infrastructure nodes, target object nodes and risk source nodes. Edge connections are constructed based on relative motion relationships, spatial proximity relationships and communication cooperation relationships to generate a structured state representation for risk reasoning. The structured state representation and task constraint information are combined to form a multimodal cue sequence, which is then input into a generative large model to output risk semantic interpretation, candidate risk links, and risk factor embedding representation. Based on risk factor embedded representation and air-ground collaborative state map, a spatiotemporal risk field is constructed and forward prediction is performed to obtain a risk probability heat map and uncertainty map within the future prediction time window. Based on the risk probability heatmap, uncertainty map and task constraint information, the fence optimization objectives and constraints are constructed. Constraint optimization is performed on the fence boundaries to generate a three-dimensional dynamic electronic fence set. Each fence boundary is assigned a confidence weight and an effective timestamp. Based on real-time new data, the air-ground collaborative status map, risk probability heat map, and three-dimensional dynamic electronic fence set are updated in an event-triggered iterative manner to form a versioned fence strategy and distribute it to the control execution terminals of the air platform and the ground platform to restrict the platform from entering high-risk areas and automatically adjust the fence range when the risk changes.

2. The method for air-ground collaborative risk prediction and dynamic electronic fence construction based on a fusion generative large model as described in claim 1, characterized in that, The process of performing time synchronization, coordinate unification, and quality screening on the raw data includes: Perform sensor intrinsic and extrinsic parameter calibration and distortion correction on visible light images, infrared images, laser point clouds and inertial navigation data acquired by the aerial platform, and perform measurement consistency verification on millimeter-wave radar measurements, vehicle positioning information and road semantic information acquired by the ground platform. Based on a unified time reference, timestamp alignment is performed on data from each modality, and interpolation compensation and frame loss repair strategies are used to align asynchronous data to the same sampling period. Based on a unified geographic coordinate system, a coordinate transformation chain is established between aerial platform coordinates, ground platform coordinates and map coordinates. Coordinate reprojection is performed on each frame of data to obtain the target position, speed and heading under unified coordinates. Perform outlier removal and confidence level weighting on the reprojected data; Outlier removal includes outlier removal based on Mahalanobis distance and velocity mutation suppression based on kinematic constraints. Confidence weighting includes generating modal confidence weights based on sensor field of view, occlusion rate, and signal-to-noise ratio. The multimodal data, after being weighted by confidence, is aggregated into spatiotemporal data blocks according to time series. The platform state vector, target object state vector, and risk source attribute vector are recorded simultaneously in the spatiotemporal data blocks to obtain a standardized air-ground collaborative multimodal spatiotemporal dataset.

3. The method for air-ground collaborative risk prediction and dynamic electronic fence construction based on a fusion generative large model as described in claim 1, characterized in that, The construction of the air-ground coordinated state diagram includes: Define a node feature vector for each type of node; Define an edge feature vector for each type of edge; Heterogeneous message passing updates are performed based on node feature vectors and edge feature vectors, outputting a structured state representation.

4. The method for air-ground collaborative risk prediction and dynamic electronic fence construction based on a fusion generative large model as described in claim 1, characterized in that, The construction of the multimodal cue sequence includes: The air-ground collaborative state diagram is encoded into a serializable structured description, and the task constraint information is encoded into a set of rule fragments; The structured description is enhanced based on the risk knowledge base to obtain candidate risk templates and treatment templates that match the current risk type. The candidate risk templates and treatment templates are then inserted into the prompt sequence to constrain the semantic range of the generative large model output. The prediction time window, spatial resolution, output format, and confidence expression method are explicitly labeled in the prompt sequence, so that the risk semantic interpretation output by the generative large model can be aligned with the numerical prediction results of the subsequent risk prediction steps. When the output of the generative large model contains statements that conflict with the task constraints, a conflict resolution process is triggered to form a set of candidate risk links that meet the constraints.

5. The method for air-ground collaborative risk prediction and dynamic electronic fence construction based on a fusion generative large model as described in claim 1, characterized in that, The generation of the risk factor embedded representation includes: Multimodal alignment encoding is performed on the standardized air-ground collaborative multimodal spatiotemporal dataset. Through cross-modal attention, image semantics, point cloud geometry, radar velocity measurement and map prior are aligned to the same feature space to obtain a unified spatiotemporal feature tensor. Introducing contrast consistency constraints during the training or adaptation phase ensures that the risk factor embedding representations obtained from the same risk event under different observation angles, heights, or lighting conditions remain close in the feature space, while also ensuring that different risk types remain separable in the feature space, thereby improving the robustness of risk semantic reasoning to complex environmental disturbances.

6. The method for air-ground collaborative risk prediction and dynamic electronic fence construction based on a fusion generative large model as described in claim 1, characterized in that, The construction of a spatiotemporal risk field and forward-looking prediction based on risk factor embedding representation and air-ground collaborative state map includes: An air reachable domain is constructed based on the flight envelope constraints of the air platform, and a ground reachable domain is constructed based on the reachable domain constraints of the ground platform. Based on the ground reachability domain, multiple hypothesis trajectory extrapolation is performed on the target object and risk source within the prediction time window. The state extrapolation based on Kalman filtering and multimodal uncertainty propagation based on particle sampling are used to obtain the occupancy probability distribution at future time. Based on the occupancy probability distribution, the candidate risk links output by the generative large model are mapped to numerical risk terms, and adaptive weights are assigned to each risk term according to the risk factor embedding representation. The risk probability heatmap is obtained by weighting and combining each risk item, and the uncertainty map is output simultaneously.

7. The method for air-ground collaborative risk prediction and dynamic electronic fence construction based on a fusion generative large model as described in claim 1, characterized in that, The generation of the three-dimensional dynamic electronic fence set includes: Thresholding is performed on the risk probability heatmap to obtain a set of high-risk voxels, and three-dimensional connected component analysis is performed on the high-risk voxel set to obtain a set of high-risk connected regions. For each high-risk connected region, perform boundary extraction and geometric simplification, and obtain deployable fence polylines or fence polygon patches through Douglas-Puk simplification and curvature constraint smoothing. Based on the different safety margin requirements of the aerial platform and the ground platform, a safety expansion calculation is performed on the initial fence boundary, and a three-dimensional expansion with height layer is performed on the aerial fence, so that the expansion amount is positively correlated with the value of the uncertainty diagram; The fence optimization objectives and constraints simultaneously incorporate task completion constraints and traffic continuity constraints.

8. The method for air-ground collaborative risk prediction and dynamic electronic fence construction based on a fusion generative large model as described in claim 1, characterized in that, The issuance and control of the fence policy include: The three-dimensional dynamic electronic fence set is encoded into an executable rule set, and the executable rule set is mapped to the flight path / no-fly zone instructions that can be recognized by the air platform flight control interface and the speed / path constraint instructions that can be recognized by the ground platform controller. A dual-channel verification mechanism is introduced during the distribution process. The first channel is the geometric legality verification of the fence, which at least verifies the fence self-intersection, fence holes, and conflicts between the fence and the map restricted area. The second channel is the policy consistency verification, which at least verifies the consistency between the fence rules and the task constraint information and priority passage rules. When any verification fails, a rollback strategy is triggered, which keeps the previous version of the fence unchanged, reduces the platform's running speed, increases the safety distance threshold between the platform and the target object, and requests the output of the generative large model to suggest alternative fences. The system monitors the platform's location and its relationship with the fence in real time at the execution end. When it detects that a boundary crossing condition is about to be triggered, it outputs progressive control commands in advance, which are handled in stages from prompting, deceleration, detour, to forced return / stopping.

Images