Unmanned aerial vehicle cluster reversible task reconstruction method and system based on risk prediction

Abstract

The invention discloses an unmanned aerial vehicle cluster reversible task reconstruction method and system based on risk prediction, and relates to the field of unmanned aerial vehicle cluster cooperative control and intelligent task scheduling. According to the method, a task graph is constructed based on job task constraints, cross-modal coding of causal consistency constraints is performed on environmental perception data and task semantic data, and stable risk characterization is obtained; predicting dynamic risk distribution in combination with unmanned aerial vehicle state data and cluster link data, and generating a risk-driven reconstruction boundary; therefore, task division, compensation node inheritance, update task dependency edge reconstruction and update subtask chain inheritance distribution are completed. According to the method, the task continuity, the reconstruction accuracy, the result inheritance effectiveness and the collaborative robustness in a complex dynamic environment can be improved.

Description

Technical Field

[0001] This invention relates to the field of collaborative control and intelligent task scheduling technology for unmanned aerial vehicle (UAV) swarms, and in particular to a method and system for reversible task reconfiguration of UAV swarms based on risk prediction. Background Technology

[0002] With the development of unmanned aerial vehicle (UAV) platforms, airborne sensing, swarm communication, and intelligent decision-making technologies, UAV swarms have been widely applied in scenarios such as emergency reconnaissance, inspection and monitoring, target search, material delivery, disaster assessment, and operations in complex areas. In these scenarios, multiple UAVs typically need to collaboratively execute complex tasks with preconditions, spatial transfer constraints, and resource coupling constraints around the same operational objective. Therefore, maintaining continuous mission execution under conditions of dynamic environmental changes, link fluctuations, single-unit failures, loss of connection, and rapid spread of local risks has become a key technical problem in UAV swarm applications.

[0003] In existing technologies, most drone swarm task adjustment schemes focus on static task allocation, local path replanning, task transfer after failure, or risk assessment based on perception information. These schemes can usually correct the executing entity or execution path after a local anomaly occurs, but most still remain at the technical level of "risk assessment + rescheduling" or "fault detection + reassignment". They lack mechanisms for boundary-based, partitioned, and inherited reconstruction around the task graph structure, making it difficult to balance the protection of completed task results, the continuous continuation of incomplete tasks, and the overall stability of the task graph structure in complex dynamic scenarios.

[0004] Furthermore, in complex operational areas, risk sources are typically multi-source, heterogeneous, and coupled, including not only flight environment risks but also communication risks, energy risks, and mission failure risks. Existing technologies for utilizing environmental perception data and mission semantic data often rely on simple fusion or correlation modeling, which struggles to effectively suppress the interference of spurious semantic correlations and transient disturbances on risk assessment. This results in insufficient stability of risk prediction results when the scenario changes, thus affecting the accuracy of subsequent mission reconfiguration decisions.

[0005] Furthermore, when a drone is unable to continue its current task segment due to malfunction, loss of connection, or exceeding risk limits, existing technologies typically reassign the unfinished task directly to other drones. This simple redistribution method often fails to retain the original task segment's mission objectives, remaining time constraints, resource requirements, execution context state, and result boundary information. This can easily lead to task duplication, task omissions, boundary misalignment, or data context interruptions, thereby reducing task continuity and overall collaborative efficiency.

[0006] Meanwhile, existing technologies typically lack risk-driven reconfiguration boundary generation mechanisms based on the direction of risk propagation, the increasing trend of risk propagation, and the situation where task-dependent edges cross high-risk areas. They cannot clearly define which task nodes should be locked and protected, which task nodes are allowed to migrate and reorganize, or which areas are suitable as access locations for compensation nodes. Consequently, when risks spread rapidly or vulnerable areas of the link change dynamically, existing solutions struggle to promptly suppress the propagation of risks to areas where tasks have already been completed, and also fail to achieve reversible task reconfiguration oriented towards the task graph structure.

[0007] Therefore, there is an urgent need to propose a reversible task reconfiguration method and system for UAV swarms based on risk prediction, in order to solve the problems of insufficient stable risk characterization capability, unclear risk-driven reconfiguration boundary, weak ability to continue unfinished tasks after failure, and insufficient adaptability of task inheritance and allocation in the existing technology. Summary of the Invention

[0008] The purpose of this invention is to provide a reversible task reconstruction method and system for UAV swarms based on risk prediction, in order to solve the problems of insufficient stability of risk representation, unclear task graph reconstruction boundaries, difficulty in continuing unfinished tasks after failure or loss of connection, easy destruction of completed task results, and inability to take into account inheritance and continuity adaptability in task redistribution under dynamic and complex environments in the prior art. This invention improves the task continuity, task reconstruction accuracy, result inheritance effectiveness, and collaborative execution robustness of UAV swarms under dynamically changing risk conditions.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A method for reversible task reconfiguration of drone swarms based on risk prediction, comprising: Construct a task graph based on the preconditions, spatial transition constraints, and resource coupling constraints of the task. The environmental perception data and task semantic data are then subjected to cross-modal encoding with causal consistency constraints to obtain the task graph. Stability risk representation of task nodes ; Based on the aforementioned stability risk characterization UAV status data and cluster link data are used to predict the comprehensive risk value of each location in the future time domain within the operational area at each time, forming a dynamic risk distribution. The propagation direction vector of key risk factors is determined, and a risk-driven reconstructed boundary is generated based on the risk propagation enhancement trend, the location of risk gradient abrupt changes, and the task dependency edges crossing high-risk areas. ; Based on the aforementioned risk-driven boundary reconstruction The task graph It is divided into a locked node area, a migrated node area, and a compensation access area. The completed results of task nodes in the locked node area are retained and valid. Task nodes in the migrated node area are allowed to be decoupled and recombined. The compensation access area is used to insert compensation nodes. Generate compensation nodes for unfinished task segments corresponding to faults, loss of connection, or risks exceeding limits. The compensation node Inherit the original task segment's task objective, remaining time limit constraints, resource requirement constraints, execution context state, and result boundary information; Cutting through the risk-driven reconstruction boundary Furthermore, the task dependency edges connecting high-risk nodes and low-risk nodes are preserved, and the task dependency edges within the locked node area are retained, thus enabling the compensation nodes to... Establish update task dependency edges with adjacent locked nodes or migrateable nodes to generate update subtask chains; Based on the UAV's remaining energy, payload capacity, spatial reachability, link reachability, and task inheritance mismatch cost, the updated subtask chain is assigned to the target UAV for execution. The task inheritance mismatch cost represents the target UAV's accessibility to the compensation node. The degree of mismatch in payload continuity, trajectory continuity, data context continuity, and link continuity.

[0011] To achieve the above objectives, the present invention also adopts the following technical solution:

[0012] A risk-prediction-based reversible mission reconfiguration system for drone swarms, characterized by comprising: The task graph construction and stability risk characterization module is used to construct task graphs based on the preconditions, spatial transition constraints, and resource coupling constraints of the tasks. The environmental perception data and task semantic data are then subjected to cross-modal encoding with causal consistency constraints to obtain the task graph. Stability risk representation of task nodes ; The boundary generation module is used to predict the comprehensive risk value of each location in the future time domain of the operation area at each time, based on the stability risk characterization, UAV status data, and cluster link data. Based on the aforementioned comprehensive risk value Risk propagation enhancement trends, risk gradient mutation locations, and task-dependent edges traversing high-risk areas generate risk-driven boundary reconstruction. ; The zoning module is used to reconstruct boundaries based on the aforementioned risk-driven approach. The task graph It is divided into a locked node area, a migrated node area, and a compensation access area; The compensation inheritance module is used to generate compensation nodes for unfinished task segments corresponding to faults, loss of connection, or risk exceeding limits. and make the compensation node Inherit the original task segment's task objective, remaining time limit constraints, resource requirement constraints, execution context state, and result boundary information; The task chain reconstruction module is used to cut off the path across the risk-driven reconstruction boundary. Furthermore, the task dependency edges connecting high-risk nodes and low-risk nodes are preserved, and the task dependency edges within the locked node area are retained, thus enabling the compensation nodes to... Establish update task dependency edges with adjacent locked nodes or migrateable nodes to generate update subtask chains; The inheritance allocation module is used to allocate the updated subtask chain to the target UAV for execution based on the UAV's remaining energy, payload capacity, spatial reachability, link reachability, and task inheritance mismatch cost. The task inheritance mismatch cost represents the target UAV's accessibility to the compensation node. The degree of mismatch in payload continuity, trajectory continuity, data context continuity, and link continuity.

[0013] Compared with the prior art, the present invention has at least the following beneficial effects:

[0014] 1. Improve the stability of risk representation. This invention obtains a stable risk representation by performing cross-modal coding with causal consistency constraints on environmental perception data and task semantic data. It can suppress the impact of spurious semantics and transient disturbances on risk assessment, thereby improving the stability and reliability of dynamic risk distribution prediction.

[0015] 2. Implementing risk-driven refactoring for task graph structures. This invention does not merely make local adjustments to paths or execution order, but generates risk-driven refactoring boundaries based on the risk propagation enhancement trend, the location of risk gradient abrupt changes, and task dependency edges traversing high-risk regions. And further develop the mission map The task is divided into a locked node area, a migrated node area, and a compensation access area, thereby providing clear boundary and structural basis for task reconfiguration and improving the pertinence and controllability of task reconfiguration.

[0016] 3. Ensure the validity of the inheritance of completed task results. This invention protects completed results by setting a locked node area, preventing rollback, destruction, or re-execution of completed tasks in the event of changes in risk, failure, or loss of connection, thereby improving the consistency of task execution results and overall work efficiency.

[0017] 4. Enhanced continuity of unfinished tasks. This invention generates compensation nodes for unfinished task segments corresponding to faults, disconnections, or risks exceeding limits. and make the compensation node Inheriting the original task segment's task objectives, remaining time constraints, resource requirements, execution context state, and result boundary information, the target UAV can continuously execute around the original task segment, reducing task breakpoints, task omissions, boundary misalignments, and data context interruptions.

[0018] 5. Improve the logical consistency of task chain reconstruction. This invention achieves this by severing the boundary of risk-driven reconstruction. For high- and low-risk cross-regional task dependencies, retain task dependencies within the locked node region and compensate the nodes. By establishing update task dependency edges with adjacent locked nodes or migrateable nodes, and generating update subtask chains, task reconstruction is upgraded from traditional task reallocation to a structured reconstruction process with boundary constraints, inheritance constraints, and access relationship constraints.

[0019] 6. Improve the adaptability and collaborative robustness of inheritance allocation. This invention, during the process of updating sub-task chain allocation, simultaneously considers the UAV's remaining energy, payload capacity, spatial reachability, link reachability, and the cost of task inheritance mismatch, enabling a comprehensive evaluation of the target UAV's impact on compensation nodes. It has the ability to perform payload continuity, trajectory continuity, data context continuity, and link continuity, thereby improving the adaptability of compensation continuity and the robustness of UAV swarm collaborative execution.

[0020] 7. Improve task completion rate in complex dynamic environments. This invention, through a technical closed loop of "stable risk characterization—risk-driven boundary reconstruction—compensation node inheritance—updating sub-task chain reconstruction—inheritance allocation execution," enables UAV swarms to achieve reversible task reconstruction and continuous execution in complex dynamic environments, thus effectively improving task completion rate and system operational stability. Attached Figure Description

[0021] Figure 1 This is a flowchart of the reversible mission reconstruction method for UAV swarms based on risk prediction provided by the present invention. Detailed Implementation

[0022] It should be noted that the following specific embodiments are used to further illustrate the present invention, but should not be construed as limiting the scope of protection of the present invention. Equivalent substitutions, simple modifications, or improvements made by those skilled in the art to the technical solutions of the present invention without departing from the concept of the present invention should all fall within the scope of protection of the present invention.

[0023] This specific implementation uses a deployment of "ground collaborative control platform + edge collaborative nodes + multiple UAVs" as an example to illustrate the method and system for reversible task reconstruction of UAV swarms based on risk prediction. In this implementation, the ground collaborative control platform is used to perform all or part of the calculations in task graph construction, stable risk characterization generation, risk-driven reconstruction boundary generation, task zoning, compensation inheritance, task chain reconstruction, and inheritance allocation; the edge collaborative nodes are used to perform local rapid risk updates, local compensation access judgment, and link continuity assessment; the UAVs are used to collect environmental perception data, report UAV status data, and execute the updated sub-task chains after allocation. The entire method revolves around a technical closed loop of "stable risk characterization—risk-driven reconstruction boundary—compensation node inheritance—updated sub-task chain reconstruction—inheritance allocation execution," where the output of the previous step serves as the input of the next step.

[0024] In this embodiment, the operational tasks performed by the UAV swarm can be set as one or more of the following: emergency reconnaissance, inspection and monitoring, target search, link relay, disaster assessment, and material delivery. The environmental perception data may include one or more of the following: visible light images, infrared images, laser point clouds, terrain grids, electromagnetic environment information, meteorological information, and obstacle distribution information; the task semantic data may include one or more of the following: task objective description, operation area description, operation time limit requirements, priority description, safety constraint description, and resource requirement constraint description; the UAV status data may include one or more of the following: current position, speed, heading, altitude, remaining energy, effective payload capacity, sensor status, and control mode status; the swarm link data may include one or more of the following: link latency, bandwidth, signal strength, packet loss rate, reachability matrix, and link topology.

[0025] Figure 1 This is a flowchart of the reversible task reconfiguration method for UAV swarms based on risk prediction provided by the present invention. The following is in conjunction with the attached diagram. Figure 1 Steps S1 to S7 are described in detail.

[0026] S1. Acquire multi-source data and construct a basic dataset and task graph for the job scenario.

[0027] This step provides a unified data input foundation for subsequent stable risk characterization generation, risk-driven reconstruction boundary generation, compensation node inheritance, and inheritance allocation.

[0028] Specifically, firstly, environmental perception data of the target work area is acquired and recorded as follows: ; Obtain task semantic data and denote it as Acquire drone status data and record it as... ; Obtain cluster link data and record it as Subsequently, regarding the aforementioned , , and Perform unified timestamp alignment, spatial coordinate unification, noise filtering, outlier removal, and missing value completion to form a basic dataset for the work scenario. .

[0029] , In the formula, This serves as the basic dataset for the job scenario. For environmental sensing data; For task semantic data; For drone status data; For cluster link data; For the task diagram; A set of task nodes; This is the set of task-dependent edges.

[0030] Among them, the task graph From task semantic data It is constructed in conjunction with the task rule base. Specifically, the construction method involves parsing the task objectives, sub-objectives, spatial regions, execution order, resource requirements, and constraints in the task semantic data into several task nodes, and establishing task dependency edges between these task nodes based on the following three types of relationships: First, preorder constraints are used to characterize the sequential relationship of "execution first, execution later"; Secondly, spatial transfer constraints are used to characterize the spatial continuity relationship between adjacent work action units; Third, resource coupling constraints are used to characterize the dependence of a task node on specific payload, communication capabilities, endurance, or relay capabilities.

[0031] In one implementation, each task node All are bound to node attribute packages The node attribute package It should include at least the task objective, the work area, the completion criteria, the resource requirements, the time limit, and the execution status. The execution status is used in subsequent steps to determine whether the task node belongs to the locked node area or the movable node area.

[0032] This step forms This will be directly passed to steps S2 and S3: where, and Used to generate stable risk characterization ; and This will be related to the stable risk characterization. Together, they are used to generate dynamic risk distributions, propagation direction vectors of key risk factors, and risk-driven reconstructed boundaries. Task Map The task zoning for step S4, the compensation node inheritance for step S5, and the task chain reconstruction for step S6 are all performed.

[0033] This step involves constructing a unified basic dataset for job scenarios. and task map This avoids deviations in subsequent risk assessment and distortions in task reconstruction caused by inconsistencies in time, coordinates, or semantic mapping of multi-source data, thus providing a stable data input foundation for the entire technology loop. Its beneficial effect is to improve the availability and consistency of subsequent risk assessment and task reconstruction.

[0034] S2. Perform cross-modal coding with causal consistency constraints on environmental perception data and task semantic data to obtain a stable risk characterization.

[0035] This step is used to process environmental sensing data. and task semantic data Stability information related to the task risk generation mechanism is extracted to generate a stability risk representation at the task node level. This is to suppress the impact of spurious semantics and transient perturbations on subsequent risk prediction.

[0036] Specifically, regarding the task graph Each task node in Based on its corresponding work area and execution time window, from environmental perception data Extract environmental sub-data related to this task node. From task semantic data Extract semantic sub-data related to the task node. Subsequently, the environmental sub-data The input environment coding network obtains the environment feature vector. semantic sub-data Inputting into a semantic coding network yields semantic feature vectors. Then, the joint feature vector is obtained through a cross-modal alignment network. Based on this, the joint feature vector Perform causal decomposition to obtain a stable risk feature vector. Instable perturbation eigenvectors and only the stable risk feature vector Mapped to task nodes Stability risk characterization .

[0037] , In the formula, For task nodes The joint eigenvectors; For cross-modal alignment mapping; For task nodes The corresponding environmental feature vector; For task nodes The corresponding semantic feature vector; For stable risk feature vectors; This represents the eigenvector of an unstable perturbation. For task nodes Stability risk characterization; For stable risk mapping functions.

[0038] In a preferred embodiment, the environment coding network may consist of a visual coding sub-network, a scene coding sub-network, and a fusion layer thereof. The visual coding sub-network is used to extract visually relevant information such as images, infrared images, and obstacle contours; the scene coding sub-network is used to extract scene-related information such as terrain constraints, no-fly zone distribution, electromagnetic environment distribution, and wind field distribution. The semantic coding network is used to extract task semantic information such as task objectives, time limits, task boundaries, priorities, and resource constraints. The cross-modal alignment network is used to integrate environmental feature vectors. With semantic feature vector Mapped to a unified representation space. The causal decomposition network is used to extract features from the joint feature vector according to the intervention consistency rule. The system screens out feature components that are stable and related to the formation mechanism of mission risks, and suppresses unstable feature components caused by changes in illumination, local noise, short-term link fluctuations, and occasional background interference.

[0039] The cross-modal coding model with causal consistency constraints is trained using an "offline pre-training + causal consistency training + online small-step fine-tuning" approach. The training method is as follows: First, construct a historical sample set. Each historical sample includes historical environmental perception data, historical task semantic data, historical risk labels, and historical task continuity labels. Secondly, joint training is performed using original samples and intervention samples. The intervention samples are generated by perturbing non-essential factors, including illumination, partial occlusion, background texture, local electromagnetic noise, and short-term obstacle changes, while the task objective and the real risk causes remain unchanged. Finally, during the online execution phase, the latest collected job data is used as incremental samples to update the model parameters in small steps, thereby improving the model's adaptability to the current scenario.

[0040] In one implementation, the model training loss function can be expressed as: , In the formula, The total loss of the cross-modal coding model with causal consistency constraints; For cross-modal alignment loss; Loss of causal consistency; For risk monitoring losses; , and These are the weighting coefficients for the corresponding loss terms.

[0041] Wherein, the cross-modal alignment loss Used to constrain the consistency between the environmental representation and semantic representation of the same task node in a unified space; the causal consistency loss Used to constrain the stability risk feature vector before and after intervention under the same task objective. Maintaining consistency or near consistency; the aforementioned risk monitoring loss Used to constrain the stable risk characteristic vector Generated stable risk characterization Consistent with the actual risk label.

[0042] The output of this step is the task graph. The set of stable risk representations corresponding to each task node in the process The set of stable risk characteristics This will be compared with the drone status data in step S1. and cluster link data Enter the risk prediction and boundary generation process in step S3.

[0043] This step extracts semantic components that are stably related to the formation mechanism of task risks from environmental perception data and task semantic data through cross-modal coding with causal consistency constraints. This can reduce the impact of false related information and accidental disturbances on risk judgment. Its beneficial effect is to improve the stability, accuracy and cross-scenario robustness of subsequent dynamic risk distribution prediction.

[0044] S3. Based on stable risk characterization, UAV status data, and cluster link data, predict dynamic risk distribution and generate risk-driven reconfiguration boundaries.

[0045] This step is used to determine the stable risk characterization set output from step S2. Combined with drone status data and cluster link data It predicts the dynamic risk distribution and the propagation direction vector of key risk factors in the future time domain, and generates risk-driven reconstructed boundaries accordingly. .

[0046] Specifically, firstly, based on the task graph constructed in step S1 Characterize the stability risk of each task node. The node-level risk prediction input is formed by binding the corresponding work area location, work time window, work type, UAV status information, and link status information. Subsequently, a spatiotemporal risk prediction sequence is constructed based on the node-level risk prediction input and input into the risk prediction model to obtain the location of each node within the future prediction time domain of the work area. At each moment Overall risk value The overall risk value is composed of all of the above. This constitutes a dynamic risk distribution and further determines the propagation direction vector of key risk factors. .

[0047] In one implementation, the comprehensive risk value Determined according to the following formula: , In the formula, For position At any moment The overall risk value; Flight risk value; This represents the communication risk value. Energy risk value; This represents the risk value for task failure. To characterize stability risk The semantic risk value obtained from the mapping; , , , and These are the weighting coefficients for the corresponding risk items.

[0048] Among them, the flight risk value The communication risk value can be determined based on obstacle density, no-fly zone coverage, terrain oppression, and wind disturbance intensity; The energy risk value can be determined based on link latency, bandwidth attenuation, signal strength degradation, and topology breakage probability; The risk value of mission failure can be determined based on path length, hovering time, climb cost, and return energy margin; The semantic risk value can be determined based on the remaining time limit, the probability of successful execution, and the probability of task interruption; Then, the stability risk is characterized by the task nodes associated with that location. Obtained through risk mapping.

[0049] To reflect the diffusion trend of risk in the time and space dimensions, this implementation further calculates the risk propagation enhancement trend. and the propagation direction vector of key risk factors .

[0050] , In the formula, For position At any moment The risk transmission trend value is increasing; The rate of change of the overall risk value over time; The spatial gradient of the comprehensive risk value; The magnitude of the spatial gradient of the comprehensive risk value; This is the gradient adjustment coefficient.

[0051] , In the formula, For position At any moment The propagation direction vector of key risk factors; To prevent positive decimals with a denominator of zero.

[0052] In one implementation, the risk gradient abrupt change location is defined as satisfying The set of locations, where, This is the gradient mutation threshold. Further, the locations satisfying the following conditions are defined as risk-driven reconstruction boundaries. Candidate generation location: I. The risk transmission enhancement trend value at the corresponding location meets the requirements. ,in, The first is the trend threshold; the second is that this location is a location of a sudden change in the risk gradient. III. Task Map There is a task-dependent edge that traverses a high-risk area and connects high-risk nodes with low-risk nodes, passing through this location.

[0053] Subsequently, along the propagation vector of key risk factors The outer regions with good link reachability connect the aforementioned candidate generation locations, forming a risk-driven reconstruction boundary. The term "outer side" here refers to the direction offset from the high-risk core area towards the low-risk side, used to reduce the probability of secondary risk propagation after compensation access.

[0054] In one implementation, the risk prediction model may employ a spatiotemporal graph prediction network, whose input includes a stable risk representation within a sliding time window. UAV status data Cluster link data and environmental perception data The time-varying information is included; its output includes the comprehensive risk value distribution for multiple future predicted times. Increased risk transmission trend And risk-driven reconstruction of boundary candidate regions. The training method of the risk prediction model is as follows: using historical sliding time window samples as input, and using risk labels, trend labels and boundary labels corresponding to future time domains as supervision signals for training, and using the latest feedback data for rolling updates during the online execution phase.

[0055] In one implementation, the training loss function of the risk prediction model can be expressed as: , In the formula, This represents the total loss of the risk prediction model; To predict losses based on the comprehensive risk value; Enhanced trend forecasting for losses to facilitate risk propagation; Predict loss for candidate boundary locations; , and These are the weighting coefficients for the corresponding loss terms.

[0056] The output of this step includes a dynamic risk distribution. Key risk factor propagation direction vector Risk-driven boundary reconstruction The risk-driven boundary reconstruction This will be used as the direct input for step S4, which involves dividing the task map.

[0057] This step goes beyond simply identifying "where the risks are," and goes on to provide boundary information on "how the risks propagate and where the task map is segmented." Its beneficial effect is that it directly transforms the risk prediction results into a basis for restructuring the task map structure, thereby improving the foresight, relevance, and controllability of the task restructuring.

[0058] S4. Based on the risk-driven boundary reconstruction, the task graph is divided into a locked node area, a migrated node area, and a compensation access area.

[0059] This step is used to visualize the task graph. It is transformed into a structured task space with the ability to "protect completed results, allow partial migration and reorganization, and reserve compensation access locations", providing a foundation for subsequent compensation inheritance and task chain reconstruction.

[0060] Specifically, the risk-driven boundary is first reconstructed based on the result obtained in step S3. For the task graph Each task node The spatial location, execution status, and result dependencies are used for joint judgment. The judgment rules include: I. If task node If a task node is completed, or if it is not fully finished but its output has been consumed by downstream tasks and is not suitable for rollback, then the task node should be terminated. Included in the locked node area; II. If task node If a task node is not yet completed, and its executing entity, execution order, or execution path allows for adjustments based on the current risk distribution, then this task node will be... Included in the migrated node region; III. Reconstructing Boundaries Driven by Risk The location on the outer edge, with good link reachability, and capable of establishing new dependencies with the locked node area or the migrateable node area, is identified as the compensation access area.

[0061] In one implementation, task nodes within the locked node area are not allowed to be deleted, rolled back, or redone; only their existing results and their dependencies on subsequent tasks are allowed to be retained. Task nodes within the migrated node area are allowed to be decoupled, rearranged, and reconnected, provided that remaining time constraints, resource requirement constraints, and link continuity constraints are met. The compensation access area is used to accommodate subsequent compensation nodes generated in step S5. .

[0062] Furthermore, regarding the task graph For task dependency edges in the data, perform the following processing: For traversing the risk-driven reconstruction boundary Furthermore, the task dependency edges connecting high-risk nodes and low-risk nodes are marked as dependency edges to be cut off. For task-dependent edges within the locked node region, they are directly retained; For nodes located within the migrated node region and not crossing the risk-driven reconstruction boundary Task-dependent edges are retained as candidate edges for refactoring.

[0063] The output of this step includes: a locked node area, a migrateable node area, a compensation access area, and a set of dependent edges to be cut. This output will be directly passed to steps S5 and S6, where the compensation access area is used for compensation nodes in step S5. Access location selection; the set of dependent edges to be cut, the locked node area, and the migrated node area are used to update the task dependent edge set in step S6. The construction.

[0064] In one implementation, when the risk propagation trend intensifies satisfy At that time, risk-driven reconstruction boundaries are preferentially generated along the outer side of the propagation direction of key risk factors. And will be related to the risk-driven reconstructed boundary. Intersecting or with the risk-driven reconstructed boundary The distance is not greater than the preset distance threshold. Migratable nodes are marked as priority splitting nodes. Edge cutting, reconnection, or compensatory access processing is preferentially performed on these priority splitting nodes to suppress the propagation of risk to the locked node area. In the formula, This is a preset distance threshold.

[0065] This step involves risk-driven boundary reconstruction. For the task graph By partitioning the data, subsequent task reconstruction can have clear division and structural boundaries. Its beneficial effect is to avoid damaging the results of completed tasks and improve the precision of subsequent compensation access and task chain reconstruction.

[0066] S5. Generate compensation nodes for unfinished task segments corresponding to faults, loss of connection, or risk exceeding limits, and construct inheritance information.

[0067] This step is used to convert an abnormally interrupted, incomplete task segment into a compensation node that can be continuously executed by other drones. and for this compensation node Bind complete inheritance information to ensure the continuity and boundary consistency of subsequent tasks.

[0068] Specifically, continuously monitor drone status data. Cluster link data and dynamic risk distribution The generation of a compensation node is triggered when any of the following conditions are met: 1. A drone malfunctions while performing a certain task segment; II. The drone lost contact while performing a certain mission segment; III. The location of the drone performing a certain task segment meets the following conditions. ,in, This is the threshold for risk exceeding the limit.

[0069] Once the triggering conditions are met, from the task graph Extract the corresponding unfinished task segment and denote it as the th segment. There are several incomplete task segments. Then, the task objective, remaining time constraints, resource requirement constraints, execution logs, cached data, communication session information, control mode information, and completed job boundary information corresponding to each incomplete task segment are read. An inheritance information package for this incomplete task segment is constructed, and the j-th compensation node is generated based on this inheritance information package. .

[0070] , In the formula, For the first One compensation node; Generate a function for the compensation node; For the first The task objectives of the unfinished task segment; For the first The remaining time constraints for each unfinished task segment; For the first Resource requirement constraints for each unfinished task segment; For the first The execution context state of an incomplete task segment; For the first The result boundary information for each incomplete task segment.

[0071] Wherein, the execution context state It includes at least two of the following states: stage data state, perception cache state, communication session state, and control mode state. Specifically, the stage data state is used to characterize which operation stage the original task segment has been executed to; the perception cache state is used to characterize perception data that has been collected but not uploaded or processed; the communication session state is used to characterize the link session or data session that still needs to be maintained; and the control mode state is used to characterize which mode the UAV was in before the interruption, whether it was autonomous cruise, hovering observation, target tracking, or relay dwelling.

[0072] The result boundary information It includes at least two of the following: the scope of completed tasks, the starting point of incomplete tasks, and the information of the continuation execution interface. Specifically, the scope of completed tasks is used to define the parts of the job that do not need to be repeated; the starting point of incomplete tasks is used to define the compensation node. The start position of the continuation execution; the continuation execution interface information is used to define the compensation node. The interface boundary that connects with the preceding locked node or subsequent task node.

[0073] In one implementation, the compensation node It is also possible to further bind compensation type tags to indicate which type of compensation node it belongs to, such as reconnaissance compensation node, relay compensation node, delivery compensation node, or identification compensation node, thereby providing a more granular task adaptation basis for the inheritance allocation in step S7.

[0074] The output of this step is a set of compensation nodes. and its inherited information package. The set of compensation nodes. The compensated access area output from step S4 is input into step S6 to reconstruct and update the task dependency edge set. And update the sub-task chain.

[0075] This step involves building compensation nodes for unfinished task segments that were abnormally interrupted. It binds complete inheritance information, which solves the problems of task context loss, boundary misalignment, duplicate execution and omission execution in the traditional simple redistribution method. Its beneficial effect is to improve the continuity, correctness and consistency of task succession.

[0076] S6. Based on the risk-driven boundary reconstruction, task zoning results, and compensation node inheritance information, reconstruct and update task dependency edges, and generate update subtask chains.

[0077] This step is used to rebuild the dependencies of the task graph based on the protection of locked node areas, reconstruction of movable node areas, and access of compensated access areas, forming an update sub-task chain that satisfies the constraints.

[0078] Specifically, firstly, based on the set of dependent edges to be cut output in step S4, the task graph is... Crossing risks drives boundary reconstruction Furthermore, task-dependent edges connecting high-risk nodes and low-risk nodes are cut off; task-dependent edges within the locked node region are directly retained, forming a set of retained edges. For task nodes within the migrated node area, they are reordered and reconnected based on remaining task objectives, resource requirements, spatial continuity, and link reachability to form a set of reconstructed edges. The compensation nodes generated in step S5 Insert a compensation access region and establish new dependency edges between it and adjacent locked nodes or adjacent migrateable nodes to form a set of compensation access edges. .

[0079] , In the formula, To update the set of task dependency edges; The set of task-dependent edges reserved within the locked node region; This is the set of task-dependent edges that can be reconstructed within the migrated node region; This is a set of task-dependent edges added after a node is connected to compensate for the new task.

[0080] In one implementation, it is based on updating the set of task dependency edges. Build and update the task graph ,in, It consists of task nodes within the locked node area, task nodes within the migrated node area, and a set of compensation nodes. Subsequently, the task graph is updated. Perform topological sorting and constraint verification to generate the following set of updated subtask chains: In the formula, To update the set of subtask chains, For the first Update sub-task chain, To update the number of subtask chains. Constraint checks include at least: 1. Reachability check, used to determine whether sequential execution is possible between preceding and following task nodes under the current spatial and path conditions; 2. Time limit consistency check, used to determine whether the update subtask chain meets the remaining time limit constraint; III. Resource consistency verification is used to determine whether the executing entity meets the corresponding compensation node requirements. Resource demand constraints; IV. Link continuity verification is used to determine whether the updated link topology meets the communication continuity requirements.

[0081] Only task chains that pass the above constraint checks are retained as update sub-task chains. In other words, step S6 is not an arbitrary reconnection of edges, but a structured reconstruction under the combined effect of boundary constraints, inheritance constraints, and execution constraints.

[0082] The output of this step is an update to the task dependency edge set. and update the set of subtask chains The updated subtask chain set It will be used as the target task object for inheritance allocation in step S7.

[0083] This step involves risk-driven boundary reconstruction. Under the constraints of locked node region, migrateable node region and compensation node The dependencies between tasks are restructured, which upgrades task refactoring from simple task redistribution to reversible refactoring at the task graph level. Its beneficial effects are to reduce the probability of task chain breakage, improve the logical consistency of task continuation and the overall success rate of operation.

[0084] S7. Perform inheritance allocation based on UAV capabilities, link reachability, and task inheritance mismatch cost.

[0085] This step is used to select the most suitable target drone from the candidate drones to continue the update sub-task chain, thereby ensuring the compensation node. And the actual execution effect of updating the sub-task chain.

[0086] Specifically, let the current set of candidate drones available for reallocation be . For updating the set of subtask chains Each update sub-task chain in the process calculates the performance of each candidate drone. For the compensation nodes in the updated sub-task chain Task inheritance mismatch cost In the formula, For candidate drones, For the first Candidate drones, The number of candidate drones, To update the set of subtask chains, For the first One compensation node, For the first Candidate UAVs for the compensation node The cost of task inheritance mismatch.

[0087] , In the formula, For the first The cost of mission succession mismatch for candidate drones; For the first Payload matching cost for candidate drones; For the first The cost of trajectory continuity for candidate drones; For the first The cost of missing data context for candidate drones; For the first The cost of link continuity loss for candidate drones; , , and These are the weight coefficients for the corresponding cost terms.

[0088] Wherein, the load matching cost Used to characterize candidate drones Effective payload capacity and compensation nodes Resource demand constraints The differences between them; the cost of trajectory continuation Used to characterize candidate drones Cut into the compensation node from its current position. The cost of trajectory adjustment required for the starting point of an incomplete task; the cost of missing data context. Used to characterize candidate drones Missing execution context state The continuity cost incurred during critical states; the cost of link continuity loss. Used to characterize candidate drones The adverse effects on link topology stability and link quality after continued execution.

[0089] In one implementation, it is also for each candidate drone Calculate the overall allocation cost : , In the formula, For the first The overall allocation cost of candidate drones; For the cost of time; For the cost of energy; For the cost of risk; For link cost; For the first The cost of mission succession mismatch for candidate drones; , , , and These are the weight coefficients for the corresponding cost terms.

[0090] Wherein, the time cost Used to characterize candidate drones The total time required to initiate and complete the corresponding update sub-task chain; the energy cost Used to characterize candidate drones The total energy consumption required to execute the updated sub-task chain; the risk cost. Used to characterize candidate drones The cumulative risk of traversing risk areas along the planned execution path; the link cost. Used to characterize candidate drones The impact on the overall link quality after executing the updated sub-task chain.

[0091] Furthermore, calculate candidate drones Sensitivity to key risk factors The sensitivity mentioned Based on candidate drones The overlap between the planned access trajectory and high-response areas of key risk factors, the cumulative exposure time, and the overlap between vulnerable areas of the links are weighted and determined. Only when... and At that time, the candidate drone Only then are they included in the set of allocatable objects, among which, The mismatch cost threshold, This is the sensitivity threshold. If multiple candidate drones meet the criteria, then a comprehensive cost allocation is chosen. The smallest candidate drone is selected as the target drone.

[0092] In one implementation, step S7 can be performed using a minimum-cost maximum flow algorithm or a constrained bipartite graph matching algorithm to simultaneously satisfy the collaborative allocation requirements of multiple update sub-task chains. After the solution is completed, the mapping result of each update sub-task chain and its corresponding target UAV is written back to the update task graph. The command is then sent to the corresponding drone for execution.

[0093] The input for this step includes the set of updated subtask chains output from step S6. The compensation node output in step S5 and its inherited information, UAV status data in step S1 And the risk information in step S3. Its output is the final inheritance allocation result of "Update Sub-task Chain - Target UAV".

[0094] This step addresses the mismatch cost of task inheritance. The process of updating the sub-task chain allocation is introduced, so that the allocation decision is no longer based solely on distance, time, or energy, but also takes into account compensation nodes. The adaptability of payload continuity, trajectory continuity, data context continuity, and link continuity has the beneficial effect of improving the stability of task continuity, the accuracy of compensation execution, and the robustness of cluster collaborative execution.

[0095] In a preferred embodiment, steps S2 to S7 follow a fixed rolling cycle. Repeated execution. That is, at the end of each rolling cycle, environmental perception data is collected again. UAV status data and cluster link data Update the stable risk characterization Dynamic risk distribution Key risk factor propagation direction vector Risk-driven boundary reconstruction The process involves re-executing task zoning, generating compensation nodes, updating task dependency edge reconstruction, and inheritance allocation. Therefore, this invention does not generate a task reconstruction scheme only once in the initial stage, but can continuously and adaptively reconstruct it as environmental risks, link topology, and UAV status change.

[0096] The reversible mission reconstruction system for UAV swarms based on risk prediction provided by this invention includes: a mission graph construction and stable risk characterization module, a boundary generation module, a zoning module, a compensation inheritance module, a mission chain reconstruction module, and an inheritance allocation module.

[0097] The task graph construction and stability risk characterization module is used to acquire environmental perception data, task semantic data, UAV status data, and cluster link data. It performs timestamp alignment, spatial coordinate unification, anomaly handling, and constraint resolution on the multi-source data to construct the task graph. Furthermore, cross-modal coding with causal consistency constraints is performed on environmental perception data and task semantic data to generate stable risk representations corresponding to task nodes. .

[0098] The boundary generation module is used to predict the comprehensive risk value of each location in the future time domain of the operation area at each time, based on the stability risk characterization, UAV status data, and cluster link data. The comprehensive risk value To form a dynamic risk distribution and determine the propagation direction vector of key risk factors. Further based on the aforementioned comprehensive risk value Risk propagation enhancement trends, risk gradient mutation locations, and task-dependent edges traversing high-risk areas generate risk-driven boundary reconstruction. .

[0099] The zoning module is used to reconstruct boundaries based on the aforementioned risk-driven approach. The task graph is structurally partitioned into locked node areas, migrateable node areas, and compensated access areas, and the boundaries crossing the risk-driven reconfiguration are marked. And the dependent edges that need to be cut off connect high-risk nodes and low-risk nodes.

[0100] The compensation inheritance module is used to extract the corresponding unfinished task segments from the task graph and generate compensation nodes when a fault, loss of connection, or risk exceeding the limit event is detected. and for the compensation node Inherited information is bound; the inherited information includes the original task segment's task objective, remaining time limit constraints, resource requirement constraints, execution context state, and result boundary information.

[0101] The task chain reconstruction module is used to cut off the boundary across the risk-driven reconstruction. For high- and low-risk cross-regional task dependency edges, retain task dependency edges within the locked node area, rebuild task dependency relationships within the migrated node area, and compensate nodes. Access the compensation access zone to generate an updated task dependency edge set. and update the set of subtask chains .

[0102] The inheritance allocation module is used for updating the set of subtask chains. Each update sub-task chain in the process, combined with the candidate UAV's remaining energy, payload capacity, spatial reachability, link reachability, and task inheritance mismatch cost, performs inheritance allocation, outputting a mapping relationship of "update sub-task chain - target UAV". The task inheritance mismatch cost is used to characterize the candidate UAV's relationship with the compensation node. The degree of mismatch in payload continuity, trajectory continuity, data context continuity, and link continuity.

[0103] In one implementation, the above modules can be deployed entirely on the ground-based collaborative control platform, or distributed across the ground-based collaborative control platform, edge collaborative nodes, and UAV-borne computing units. For computationally intensive cross-modal coding and risk prediction processes, deployment on the ground-based collaborative control platform is preferred; for boundary local updates, compensation access judgments, and link continuity judgments with high latency requirements, deployment on edge collaborative nodes or UAV-borne computing units is preferred.

[0104] The present invention has been described in detail above with reference to specific embodiments. It should be understood that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. For those skilled in the art, various equivalent substitutions, simple transformations, or improvements can be made to the model structure, data source, boundary generation strategy, compensation node inheritance rules, task chain reconstruction rules, inheritance allocation solution method, and module deployment method of the present invention without departing from the spirit and substance of the present invention, and all such equivalent substitutions, simple transformations, or improvements should fall within the scope of protection of the present invention.

Claims

1. A method for reversible task reconfiguration of unmanned aerial vehicle (UAV) swarms based on risk prediction, characterized in that, include: Construct a task graph based on the preconditions, spatial transition constraints, and resource coupling constraints of the task. , The environmental perception data and task semantic data are then subjected to cross-modal encoding with causal consistency constraints to obtain the task graph. Stability risk representation of task nodes ; Based on the aforementioned stability risk characterization UAV status data and cluster link data are used to predict the comprehensive risk value of each location in the future time domain within the operational area at each time, forming a dynamic risk distribution. The propagation direction vector of key risk factors is determined, and a risk-driven reconstructed boundary is generated based on the risk propagation enhancement trend, the location of risk gradient abrupt changes, and the task dependency edges crossing high-risk areas. ; Based on the aforementioned risk-driven boundary reconstruction The task graph It is divided into a locked node area, a migrated node area, and a compensation access area. The completed results of task nodes in the locked node area are retained and valid. Task nodes in the migrated node area are allowed to be decoupled and recombined. The compensation access area is used to insert compensation nodes. Generate compensation nodes for unfinished task segments corresponding to faults, loss of connection, or risks exceeding limits. The compensation node Inherit the original task segment's task objective, remaining time limit constraints, resource requirement constraints, execution context state, and result boundary information; Cutting through the risk-driven reconstruction boundary Furthermore, the task dependency edges connecting high-risk nodes and low-risk nodes are preserved, and the task dependency edges within the locked node area are retained, thus enabling the compensation nodes to... Establish update task dependency edges with adjacent locked nodes or migrateable nodes to generate update subtask chains; Based on the UAV's remaining energy, payload capacity, spatial reachability, link reachability, and task inheritance mismatch cost, the updated subtask chain is assigned to the target UAV for execution. The task inheritance mismatch cost represents the target UAV's accessibility to the compensation node. The degree of mismatch in payload continuity, trajectory continuity, data context continuity, and link continuity.

2. The method for reversible task reconfiguration of UAV swarms based on risk prediction according to claim 1, characterized in that, The cross-modal encoding of the causal consistency constraint includes: performing cross-modal alignment on multiple perturbation samples under the same task objective to obtain a joint feature vector. ; for the joint feature vector By performing causal decomposition, a stable risk feature vector is obtained. unstable perturbation eigenvectors Only the stable risk feature vector Mapped to the stable risk representation , In the formula, For joint feature vectors, To stabilize the risk feature vector, This represents the eigenvector of an unstable perturbation. For stable risk mapping functions.

3. The method for reversible task reconfiguration of UAV swarms based on risk prediction according to claim 1, characterized in that, Location in the work area At any moment Overall risk value Determined according to the following formula: In the formula, For flight risk values, For communication risk values, Energy risk value, This represents the risk value of mission failure. To characterize the stability risk The semantic risk value obtained from the mapping, , , , and These are the weighting coefficients for the corresponding risk items.

4. The method for reversible task reconfiguration of UAV swarms based on risk prediction according to claim 3, characterized in that, Increased risk transmission trend Determined according to the following formula, and based on the aforementioned risk propagation enhancement trend. Generate the risk-driven reconfiguration boundary : In the formula, For position At any moment The risk transmission enhancement trend value, The rate of change of the overall risk value over time. The spatial gradient of the comprehensive risk value. To represent the magnitude of the spatial gradient of the comprehensive risk value, For gradient adjustment coefficient; when Exceeding the preset threshold Furthermore, when a task dependency edge crosses a high-risk area, the position corresponding to that task dependency edge is determined as the risk-driven reconstruction boundary. Candidate generation positions.

5. The method for reversible task reconfiguration of UAV swarms based on risk prediction according to claim 1, characterized in that, The update task depends on the set of edges Determined according to the following formula: In the formula, To update the task dependency edge set, The set of task-dependent edges retained within the locked node region. This is the set of task-dependent edges that can be reconstructed within the migrated node region. This is the set of task dependency edges added after the compensation node is connected; wherein, the The update task in the middle depends on the edge connecting the compensation node. With adjacent locked nodes or migrateable nodes.

6. The method for reversible task reconfiguration of UAV swarms based on risk prediction according to claim 1, characterized in that, The compensation node The inherited execution context state includes at least two of the following: the stage data state of the original task segment, the perception cache state, the communication session state, and the control mode state; the result boundary information includes at least two of the following: the scope of completed tasks, the starting point of uncompleted tasks, and the continuation execution interface information.

7. The method for reversible task reconfiguration of UAV swarms based on risk prediction according to claim 1, characterized in that, The cost of task inheritance mismatch Determined according to the following formula: ， In the formula, The cost of task inheritance mismatch For load matching cost, The cost of continuing the trajectory The cost of missing data context, As a cost to link continuity loss, , , and These are the weighting coefficients for the corresponding cost terms.

8. The method for reversible task reconfiguration of UAV swarms based on risk prediction according to claim 7, characterized in that, The first The cost of task inheritance mismatch between the candidate drone and its compensation node C As a candidate selection criterion and component of the comprehensive allocation cost for updating subtask chain allocation; only those that meet the criteria will be considered. and Candidate drones are included in the set of allocable objects, and the first drone included in the set of allocable objects is... The comprehensive allocation cost for each candidate drone is determined according to the following formula. : In the formula, For the first The overall allocation cost corresponding to each candidate drone. For the cost of time, For the cost of energy, For the cost of risk, For link cost, For the first The cost of mission inheritance mismatch for each candidate drone. , , , and These are the weighting coefficients for the corresponding cost terms. The mismatch cost threshold, For the first The sensitivity of candidate drones to key risk factors. This is the sensitivity threshold.

9. The method for reversible task reconfiguration of UAV swarms based on risk prediction according to claim 4, characterized in that, when At that time, the risk-driven reconstruction boundary is preferentially generated along the outer side of the propagation direction of the key risk factors. And prioritize the segmentation of the risk-driven reconstruction boundary. Intersecting or with the risk-driven reconstructed boundary The distance is not greater than the preset distance threshold. Migrative nodes are used to suppress the spread of risk to locked node areas.

10. A reversible mission reconfiguration system for unmanned aerial vehicle (UAV) swarms based on risk prediction, characterized in that, include: The task graph construction and stability risk characterization module is used to construct task graphs based on the preconditions, spatial transition constraints, and resource coupling constraints of the tasks. The environmental perception data and task semantic data are then subjected to cross-modal encoding with causal consistency constraints to obtain the task graph. Stability risk representation of task nodes ; The boundary generation module is used to predict the comprehensive risk value of each location in the future time domain of the operation area at each time, based on the stability risk characterization, UAV status data, and cluster link data. Based on the aforementioned comprehensive risk value Risk propagation enhancement trends, risk gradient mutation locations, and task-dependent edges traversing high-risk areas generate risk-driven boundary reconstruction. ; The zoning module is used to reconstruct boundaries based on the aforementioned risk-driven approach. The task graph It is divided into a locked node area, a migrated node area, and a compensation access area; The compensation inheritance module is used to generate compensation nodes for unfinished task segments corresponding to faults, loss of connection, or risk exceeding limits. and make the compensation node Inherit the original task segment's task objective, remaining time limit constraints, resource requirement constraints, execution context state, and result boundary information; The task chain reconstruction module is used to cut off the path across the risk-driven reconstruction boundary. Furthermore, the task dependency edges connecting high-risk nodes and low-risk nodes are preserved, and the task dependency edges within the locked node area are retained, thus enabling the compensation nodes to... Establish update task dependency edges with adjacent locked nodes or migrateable nodes to generate update subtask chains; The inheritance allocation module is used to allocate the updated subtask chain to the target UAV for execution based on the UAV's remaining energy, payload capacity, spatial reachability, link reachability, and task inheritance mismatch cost. The task inheritance mismatch cost represents the target UAV's accessibility to the compensation node. The degree of mismatch in payload continuity, trajectory continuity, data context continuity, and link continuity.

Images