A method for resolving cooperation rule conflicts of a UAV cluster

Abstract

The application belongs to the technical field of unmanned aerial vehicles, and specifically discloses a cooperation rule conflict resolution method for an unmanned aerial vehicle cluster, which comprises the following steps: based on a cooperation rule library of the unmanned aerial vehicle cluster and a newly added rule, a conflict event is obtained through rule atomization, rule dependency graph construction and update, and conflict edge detection; based on the conflict event, a recommended resolution action is obtained through processing and confidence evaluation fusion by a double-channel resolution engine, the double-channel resolution engine is constructed based on a static channel and a dynamic channel, the static channel is used for generating a resolution action based on an existing cooperation rule conflict resolution action library, and the dynamic channel is used for generating a resolution action based on reinforcement learning; based on the recommended resolution action, a target resolution action is obtained and the target resolution action is executed on the cooperation rule library of the unmanned aerial vehicle cluster through simulation execution in a virtual environment and dynamic priority reassignment. Through the application, cooperation rule conflicts can be efficiently detected and resolved.

Description

Technical Field

[0001] This application belongs to the field of unmanned aerial vehicle (UAV) technology, and more specifically, relates to a method for resolving conflicts in the cooperation rules of UAV swarms. Background Technology

[0002] With the rapid development of drone technology, drone swarms are increasingly widely used in fields such as express delivery and logistics, agricultural and forestry plant protection, power line inspection, and disaster search and rescue. The cooperation rules of a drone swarm are the logical foundation guiding the drone entities within the swarm to cooperate, avoid obstacles, and allocate tasks. The drone swarm cooperation rule base (the rule library composed of the cooperation rules of the drone swarm) stores existing cooperation rules, which are usually pre-set by domain experts or learned and accumulated from historical tasks. However, in complex and ever-changing open environments, swarms often need to dynamically introduce new cooperation rules to adapt to new tasks or unexpected situations. When new cooperation rules emerge, they may conflict with existing rules, such as competition for the same airspace or charging station resources, or producing mutually exclusive execution conclusions (e.g., simultaneously requiring "hovering and waiting" and "accelerating to the destination"), which can lead to swarm decision-making confusion or even mission failure. Therefore, how to efficiently detect and resolve cooperation rule conflicts is a pressing technical problem to be solved in this field. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the purpose of this application is to achieve efficient detection and resolution of collaborative rule conflicts.

[0004] To achieve the above objectives, in a first aspect, this application provides a method for resolving cooperation rule conflicts in a drone swarm, comprising: Based on the drone swarm collaboration rule base and newly added rules, conflict events are obtained through rule atomization, rule dependency graph construction and updating, and conflict edge detection. Based on conflict events, a dual-channel resolution engine is used to process and fuse with confidence assessment to obtain recommended resolution actions. The dual-channel resolution engine is built on static and dynamic channels. The static channel is used to generate resolution actions based on the existing collaborative rule conflict resolution action library, and the dynamic channel is used to generate resolution actions based on reinforcement learning. Based on the recommended resolution actions, the target resolution actions are obtained by simulating execution in a virtual environment and dynamically reassigning priorities, and then the target resolution actions are executed on the drone swarm collaboration rule base.

[0005] Understandably, the method proposed in this application first transforms the UAV swarm collaboration rule base into a weighted directed rule dependency graph, achieving localized and accurate conflict detection. Then, it introduces a static and dynamic dual-channel resolution engine, combining historical experience and a reinforcement learning model to generate optimal recommended resolution actions within a confidence evaluation framework. Finally, through simulated execution in a virtual environment, supplemented by a dynamic priority reallocation mechanism, it effectively avoids cascading conflicts introduced by resolution actions. This method not only maintains the logical matching of global rules but also significantly reduces the need for manual intervention, thus achieving efficient and secure collaborative rule conflict detection and resolution in complex and ever-changing open environments.

[0006] In one possible implementation, the above-mentioned UAV swarm collaboration rule base and newly added rules are used to obtain conflict events through rule atomization, rule dependency graph construction and updating, and conflict edge detection, including: Based on the condition and conclusion parts of each rule in the drone swarm collaboration rule base, a semantic parser is used to extract the condition and conclusion items (rule atomization) to obtain the rule dependency graph; Based on the rule dependency graph and the condition and conclusion terms of the newly added rules, the graph node set of the rule dependency graph is expanded by semantic similarity comparison to obtain the updated rule dependency graph. Based on the updated rule dependency graph, conflict events are obtained by checking whether there is mutual exclusion or resource competition between the edges connected to the newly added rule conclusion nodes. The conflict events include the identifiers of the two rule conclusion nodes that are in conflict, the conflict type, and the weights of the two rule conclusion nodes that are in conflict.

[0007] Specifically, a semantic parser is used to atomically process each rule in the rule base, extracting condition terms and conclusion terms as graph nodes to construct an initial rule dependency graph. When a new rule is introduced, the system extracts its condition terms and conclusion terms, identifies and expands new nodes that have not yet appeared through semantic similarity comparison, and completes the update of the rule dependency graph. Subsequently, in the updated rule dependency graph, the system checks the edges connected to the conclusion nodes of the new rule. If two directly connected or shared condition conclusion nodes are found to have mutually exclusive conclusions (such as "dangerous" and "safe") or resource competition (such as conflicting operations on the same physical resource), and their weights are close, they are determined to be conflicting edges, and a conflict event containing node identifiers, conflict types, and weights is generated accordingly.

[0008] As can be seen, the process of obtaining conflict events described above aims to transform the rule base into a weighted directed graph to achieve localized conflict detection.

[0009] In one possible implementation, the above-mentioned conflict-based processing and confidence evaluation are fused through a dual-channel resolution engine to obtain recommended resolution actions, including: Based on the collision event, a hash index is performed through the static channel to obtain the static resolution action and the static channel confidence. Based on conflict events, reinforcement learning model inference is performed through dynamic channels to obtain dynamic resolution actions and dynamic channel confidence. Based on the confidence scores of static and dynamic channels, the difference is calculated and compared with a preset threshold (confidence assessment fusion) to obtain recommended resolution actions.

[0010] Understandably, the process of obtaining recommended conflict resolution actions employs a dual-channel collaborative decision-making mechanism involving both static and dynamic approaches. Specifically, the static channel relies on an existing collaborative rule-based conflict resolution action library, using multi-level hash indexes to quickly match highly similar resolution actions from historical experience and output the static channel confidence score. The dynamic channel encodes conflict events as state vectors, utilizes a deep decision network to output action probability distributions, and combines sandbox pre-testing to verify the output of dynamic resolution actions and dynamic channel confidence scores. Finally, the system calculates the difference between the static and dynamic channel confidence scores and performs decision fusion based on preset threshold rules: when the difference is greater than or equal to a set positive threshold, a static action is adopted; when it is less than or equal to a set negative threshold, a dynamic action is adopted; and when it falls between the two, a human-machine collaborative process is triggered. This ensures both resolution efficiency and the reliability and security of the decision-making process.

[0011] In one possible implementation, the above-mentioned method, based on collision events, uses static channels for hash indexing to obtain static resolution actions and static channel confidence levels, including: Based on the conflict type in the conflict event, partitioning is performed using a first-order hash index to obtain the conflict type partition; Based on the weights of the two conflicting rule conclusion nodes in a conflict event, logarithmic-scale bucketing is performed using a second-order hash index to obtain the target bucket within the conflict type partition. The calculation formula is as follows: ; in, This represents the bucket index of the target bucket. This indicates a round-down operation. This indicates the weight ratio of conflict rules. Indicates the number is The weights corresponding to the rule conclusions, Indicates the number is The weights corresponding to the rule conclusions are numbered as follows: The rule conclusions and numbers are as follows The rule conclusion is two conflicting rule conclusion nodes in a conflict event; The logarithmic scale bucket width represents the partitioning width for conflict types; This represents the minimum weight ratio threshold; The cosine similarity is calculated using a three-order hash index to obtain the static resolution actions and static channel confidence within the target bucket. The calculation formula is as follows: ; in, Indicates the confidence level of the static channel. This represents the cosine similarity corresponding to the static resolution action. This indicates the success rate of the static resolution action.

[0012] Specifically, the system performs first-order hash partitioning based on the type of conflict event (mutually exclusive conclusions or resource competition) to narrow the search scope. Second, it extracts the weights of the two rule conclusion nodes that conflict, calculates their weight ratio, and uses a logarithmic scale bucketing formula to perform second-order hash indexing, mapping the weight ratio to specific target buckets to match historical conflict scenarios with similar priority differences. Finally, within the target bucket, the system performs third-order indexing by calculating cosine similarity, filters out the static resolution action closest to the current conflict, and multiplies the cosine similarity of the action with its preset resolution success rate to calculate the static channel confidence, thus providing a reliable historical experience reference for subsequent dual-channel fusion.

[0013] Therefore, the above process of obtaining static resolution actions achieves an efficient mapping from collision events to resolution actions by constructing a three-level hash index.

[0014] In one possible implementation, the above-mentioned reinforcement learning model inference based on conflict events and dynamic channels is performed to obtain dynamic resolution actions and dynamic channel confidence, including: Based on the conflict event, a state encoder is used to encode and obtain the state vector; Based on the state vector, features are extracted and fused through a deep decision network to obtain the action probability distribution. The deep decision network includes two parallel sub-networks (one sub-network uses a CNN convolutional neural network to extract data features, and the other sub-network uses a bidirectional LSTM network to process environmental information), a fusion layer, and a nonlinear function layer. Based on the action probability distribution, action selection, sandbox pre-visualization verification, and risk verification are performed through online decision-making. If the risk verification passes, the dynamic resolution action and dynamic channel confidence are obtained, calculated using the following formula: ; in, Indicates the confidence level of the dynamic channel. This represents the variance of the action probability distribution. This is a constant used to avoid the denominator being zero.

[0015] Specifically, the state encoder transforms the input conflict events into a fixed-dimensional state vector. This vector is then fed into a deep decision network comprising a CNN and a bidirectional LSTM dual-branch structure. After feature extraction and fusion layers, the vector is concatenated, and a nonlinear function layer outputs a probability distribution covering four types of atomic actions: adding rule conditions, priority boosting, priority debuffing, and manual intervention requests. During the online decision-making phase, the system employs a greedy strategy to select actions and places them in a sandbox pre-trial environment for risk verification. If the risk change is below a set threshold, the action is output as a dynamic resolution action, and the dynamic channel confidence is calculated based on the variance of the action probability distribution (the smaller the variance, the higher the confidence). Simultaneously, the system acquires rewards, stores experience, and updates network parameters based on the pre-trial results, forming a closed-loop online reinforcement learning mechanism.

[0016] Therefore, the process of obtaining dynamic resolution actions described above is based on a reinforcement learning framework to achieve intelligent exploration and resolution of unknown conflicts.

[0017] In one possible implementation, the aforementioned recommended resolution actions are obtained by simulating execution in a virtual environment and dynamically reassigning priorities, including: Based on the recommended resolution actions, the weights of changing rule conditions and / or rule conclusions are simulated in a virtual environment to obtain the changed rules, and based on the changed rules, the rule dependency graph is simulated to be updated in the virtual environment. Based on the updated rule dependency graph obtained from the simulation, the system detects whether there are mutually exclusive conclusions or resource competition among the rule conclusion nodes and obtains the conflict detection results. If the conflict detection results indicate the existence of a chain of conflicts (conflicts caused by rule changes), dynamic priority reallocation and simulation in a virtual environment are continuously performed until the chain of conflicts are eliminated and a conflict-free target resolution action is obtained. The following formula is used to dynamically reallocate the rule conclusions with conflicts: ; in, This indicates the weight of the recalculated rule result (the larger the weight value of the rule conclusion, the higher the priority of the rule conclusion). Indicates the weight of historical rule results. Indicates credibility. Indicates the time factor, Indicates conflict penalty items, , , and These are preset coefficients.

[0018] It should be noted that after the dual-channel engine outputs the recommended resolution action, the system does not immediately apply it to the real rule base. Instead, it first simulates the execution of the action in a virtual environment (such as adjusting rule weights or adding conditions) and updates the rule dependency graph accordingly. Subsequently, the system performs conflict detection on the updated rule dependency graph obtained from the simulation. If no new mutually exclusive conclusions or resource competition are found, the recommended resolution action is confirmed as a safe target resolution action. If a chain of conflicts is detected, the system triggers a dynamic priority reallocation mechanism, comprehensively considering historical weights, credibility, timeliness factors, and conflict penalty terms to recalculate the rule weights. The resolution action is then adjusted with the new weights and the simulation is performed again for verification. This process is iterated until all potential chain conflicts are eliminated, and finally, the safe and reliable target resolution action is applied to the UAV swarm collaboration rule base.

[0019] It is evident that the aforementioned process of obtaining the target resolution action aims to ensure, through a forward-looking simulation verification mechanism, that the execution of the resolution action will not trigger new systemic risks.

[0020] Secondly, this application provides a device for resolving cooperation rule conflicts in a drone swarm, comprising: The conflict detection module is used to obtain conflict events based on the drone swarm collaboration rule base and newly added rules, through rule atomization, rule dependency graph construction and update, and conflict edge detection. The conflict resolution action recommendation module is used to obtain recommended conflict resolution actions based on conflict events by processing and fusing confidence evaluation through a dual-channel conflict resolution engine. The dual-channel conflict resolution engine is built based on static and dynamic channels. The static channel is used to generate conflict resolution actions based on the existing collaborative rule conflict resolution action library, and the dynamic channel is used to generate conflict resolution actions based on reinforcement learning. The resolution action simulation execution module is used to obtain the target resolution action based on the recommended resolution action by simulating execution in a virtual environment and dynamically reassigning priorities, and then execute the target resolution action on the UAV swarm collaboration rule base.

[0021] Thirdly, this application provides an electronic device, including: a memory and one or more processors; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions; the one or more processors invoke the computer instructions to cause the electronic device to perform the method described in the first aspect or any possible implementation of the first aspect.

[0022] Fourthly, this application provides a computer-readable storage medium including instructions that, when executed on an electronic device, cause the electronic device to perform the method described in the first aspect or any possible implementation thereof.

[0023] Fifthly, this application provides a computer program product, including a computer program or instructions that, when run on an electronic device, cause the electronic device to perform the method described in the first aspect or any possible implementation thereof.

[0024] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0025] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: This application achieves localized conflict detection of rules and resolves rule conflicts by using conflict detection based on rule dependency graph construction, generating collaborative rule conflict resolution actions, and simulating the execution of resolution actions to avoid chain conflicts. It maintains the global rule logic matching, reduces the need for manual intervention, and achieves efficient conflict detection and resolution. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the method for resolving collaboration rule conflicts in a drone swarm provided in this application embodiment; Figure 2 This is a schematic diagram of the dynamic channel decision-making process provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the drone swarm collaboration rule conflict resolution device provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0028] In this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A existing alone, A and B existing simultaneously, and B existing alone. In this application, the symbol " / " indicates that the related objects are in an "or" relationship, for example, A / B means A or B.

[0029] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0030] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple processing units means two or more processing units, multiple elements means two or more elements, etc.

[0031] First, the meaning of the parameters in the formulas involved in the embodiments of this application will be explained.

[0032] (1) ; Parameter explanation: : Indicates the drone swarm collaboration rule base (or rule base for short); : Represents the specific collaboration rules stored in the rule base; : Indicates the total number of collaborative rules in the rule base; ; Parameter explanation: : indicates the first rule in the rule base Rules; : indicates the first The condition part of the rule; : indicates the first The conclusion of the rule.

[0033] (2) ; Parameter explanation: : indicates the first Atomic expressions for the rules; : indicates the first The rule conditions section contains One condition item; : Indicates the number of condition terms; : indicates the first The conclusion item in the conclusion section of the rule.

[0034] (3) ; Parameter explanation: : Represents the constructed rule dependency graph (the mathematical representation of a weighted graph); : Represents the vertex set (graph node set), which represents the set of nodes in the graph (corresponding to condition terms and conclusion terms); : Represents an edge set, which is a set of connections (such as paths or relationships) that connect nodes; : Represents the weight set, which records the connection strength of these edges or the priority weight of the rules.

[0035] (4) ; Parameter explanation: : Represents the atomic expression of the newly added rule; : Indicates a newly added collaboration rule; : Indicates the condition item of the newly added rule; : Indicates the conclusion of the newly added rule.

[0036] (5) ; Parameter explanation: : Indicates the bucket index (number) calculated by the rule weight ratio for interval bucketing; : indicates that the number is The weight corresponding to the rule conclusion (representing the priority of the rule conclusion); : indicates that the number is The weights corresponding to the rule conclusions; : Indicates the weight ratio of conflict rules; : Indicates the minimum weight ratio threshold; : Indicates the width of the logarithmic scale bins (default value is 0.5); : Indicates the round-down operation.

[0037] (6) ; Parameter explanation: : Represents the reward function (reward value) in a reinforcement learning model; : Indicates the number of new conflicts in the next 5 minutes (when (Maximize reward value at the time) : Indicates the time taken for the decision-making process; : Indicates whether to request human intervention (1 if human intervention is required, 0 otherwise); : Represents the task risk vector in the sandbox simulation; : Represents the L2 norm of a vector.

[0038] (7) ; Parameter explanation: : Indicates the confidence level of the static channel; : Represents the cosine similarity corresponding to the resolution action determined by the static channel; : Indicates the success rate (preset value) of the resolution action determined by the static channel.

[0039] (8) ; Parameter explanation: : Indicates the confidence level of the dynamic channel; : Represents the variance of the action probability output by the dynamic channel (reflecting the uncertainty of action selection); : indicates avoiding the minimum value where the denominator is 0 (default is 0). ).

[0040] (9) ; Parameter explanation: : Represents the decision fusion result (the difference between the confidence levels of the static channel and the dynamic channel); : Indicates the confidence level of the static channel; : Indicates the confidence level of the dynamic channel.

[0041] (10) ; Parameter explanation: : Represents the new weight of conflicting rule conclusions after recalculation; : Represents the historical weighting coefficient; : Represents the historical weight of the rule's conclusion; : Represents the credibility coefficient; : Indicates credibility (preset value); : Represents the timeliness factor coefficient; : Represents the timeliness factor (calculated using the formula). ,in This is the time when the rules have not been updated. (This is the attenuation coefficient). : Indicates the coefficient of the conflict penalty term; : Indicates a conflict penalty.

[0042] (11) ; Parameter explanation: : Indicates a conflict penalty item; : Indicates the number of conflicts triggered by this rule during the most recent simulation of the resolution action; : Indicates the conflict coefficient (preset value).

[0043] The embodiments of this application are described below with reference to the accompanying drawings.

[0044] The steps of the method for resolving cooperation rule conflicts in a drone swarm provided in this application are as follows: Figure 1 As shown, it mainly includes three stages: First, conflict detection is based on the rule dependency graph, which transforms the rule base into a weighted directed graph to achieve localized conflict detection; second, collaborative rule conflict resolution actions are generated, which are generated and merged through a static and dynamic dual-channel engine; and finally, the resolution actions are simulated to avoid chain conflicts, which verifies the safety of the resolution actions and performs dynamic priority reallocation through simulation in a virtual environment (such as sandbox pre-play).

[0045] 1. Conflict detection based on rule dependency graph; The purpose of this step is to convert the drone swarm collaboration rule base (hereinafter referred to as the rule base) into a weighted directed graph to achieve localized conflict detection. It mainly includes three parts: rule atomization, rule dependency graph update, and conflict edge detection.

[0046] 1.1 Rule atomization and rule dependency graph construction; For the rule base Each rule is represented as Extract the conditional part using a semantic parser. condition items (The conditions section contains one or more conditional items) and the conclusion section Conclusion (The conclusion section contains one or more conclusion items). Each condition / conclusion item corresponds to a node in the rule dependency graph. The rule atomic expression is as follows: .

[0047] Construct a rule dependency graph based on the condition and conclusion items corresponding to each rule in the rule base. . It is the mathematical representation of a weighted graph, defined by three sets. (Graph node set) represents the set of nodes in the graph, and E (edge ​​set) represents the set of connections (such as paths or relationships) that link the nodes. The weight set assigns a numerical value to each edge to quantify the strength of the connection.

[0048] For example, suppose there is a rule in the rule base. "If the drone's battery level is below 20% (condition 1) and it is more than 5 kilometers away from the base (condition 2), then immediately return to base (conclusion)." This rule contains two conditions. (Battery level <20%) and (Distance > 5km), the conclusion section contains one conclusion item. (Immediately return to base). In the constructed rule dependency graph In this context, these three items will be mapped to graph node sets respectively. Three independent nodes in the edge set It will generate directed edges from the condition node to the conclusion node, and a weight set. Then record the connection strength of these edges or the priority weight of the rules.

[0049] 1.2 Rule dependency graph update; For the new rules Dependency graph with existing rules First, extract the expression for the new rule. In the existing graph node set In the graph node set, find nodes that are semantically similar to the newly added rule, and analyze which condition / conclusion terms in the newly added rule have not appeared in the graph node set (nodes with no semantically similar meaning were not found). Add the condition / conclusion terms that have not appeared to the graph node set. To expand the number of nodes in the graph.

[0050] 1.3 Conflict edge detection; For the updated graph node set This requires checking all edges connected to the nodes leading to the conclusions of the newly added rules. Conflicts are categorized into two types (conflict types): (a) The conclusions are mutually exclusive; There exist two rule conclusion nodes that are directly connected, have similar weights, and are mutually exclusive. For example, if one rule conclusion is "dangerous" and the other is "safe," then they are mutually exclusive.

[0051] (b) Resource competition; There are two rule conclusion nodes with similar weights, both requiring the same resource, but their operations conflict. For example, one rule conclusion requires updating a resource, while the other requires deleting it; this constitutes an operation conflict.

[0052] When a conflicting edge is detected, a conflict event is generated. The specific content of the conflict event includes: the identifiers of the two conflicting rule conclusion nodes, the conflict type, and the weights of the two conflicting rule conclusion nodes (the larger the weight value of the rule conclusion, the higher the priority of the rule conclusion).

[0053] For example, suppose a logistics drone swarm already has a rule with the conclusion node "fly to delivery point A" (weight 0.8). A new rule is introduced with the conclusion node "avoid and fly to alternate landing point B" (weight 0.85). During conflict edge detection, if these two conclusion nodes are triggered by the same condition (e.g., "a large obstacle was detected ahead of the flight path"), since "continue flying" and "avoid and make an alternate landing" are mutually exclusive (conflict type a) and their weights are close (0.8 and 0.85), the system will generate a conflict event. This conflict event includes: node identifiers (fly to area A, fly to alternate landing point B), conflict type (mutually exclusive conclusions), and their weight values ​​(0.8 and 0.85). Similarly, in a farmland plant protection scenario, an existing rule requires "plant protection drone #1 to occupy pesticide application point #1," and a new rule requires "plant protection drone #2 to occupy pesticide application point #1." This is competition for the same physical resource (conflict type b), and will also trigger a conflict event.

[0054] 2. Generation of actions to resolve collaboration rule conflicts; This step employs a dual-channel resolution engine: a static channel (generating resolution actions based on an existing collaborative rule conflict resolution action library) and a dynamic channel (generating resolution actions based on reinforcement learning). The dual channels output recommended resolution actions after confidence evaluation.

[0055] 2.1 Static Channel; Based on the existing collaborative rule-based conflict resolution action library, a three-level hash index is used to map conflict events to resolution actions. For an input conflict event, the conflict type partition is first located, then buckets with matching weight ratios are selected, and finally, the cosine similarity is used to find the closest resolution action.

[0056] First-order index: Hash partitioning for conflict types (mutually exclusive conclusions or resource contention); Second-order index: The rule weight ratio is calculated by dividing the interval into buckets, and the calculation formula is as follows: ; in: This indicates a round-down operation. This indicates the weight ratio of conflict rules; Indicates the width of the bins on a logarithmic scale; This represents the minimum weight ratio threshold. Indicates the number is The weights corresponding to the rule conclusions; Indicates the number is The weight corresponding to the rule conclusion; this weight represents the priority of the rule conclusion.

[0057] Third-order indexing: using cosine similarity to find the closest resolution action.

[0058] 2.2 Dynamic Channel; (a) Construction of reinforcement learning models; First, the state space is encoded as a fixed-dimensional numerical vector, translating conflict events into numerical vectors. For its action space, four types of atomic actions are defined: adding rule conditions, priority boosting (increasing the weight of rule conclusions), priority debuffing (debuffing the weight of rule conclusions), and manual intervention requests. These four types of atomic actions cover all conflict resolution scenarios. In essence, adding rule conditions updates the rule dependency graph, thereby promoting the elimination of edges between two conflicting rule conclusion nodes. If this edge is eliminated, the conflict is resolved. Priority boosting or debuffing increases the priority difference between two conflicting rule conclusion nodes. If the priority difference exceeds a certain threshold, simultaneous execution of both conflicting rule conclusion nodes is avoided, thus resolving the conflict. Manual intervention updates the rule dependency graph or boosts or debuffs priorities, further promoting conflict resolution.

[0059] Based on the collaboration rules of drone swarms, the following examples illustrate the four types of atomic actions: 1) Adding conditions to rules: In disaster search and rescue scenarios, if rule A (returning home when low battery) conflicts with rule B (flying to newly discovered signs of life), the condition "in non-emergency search and rescue mission status" can be added to rule A to eliminate the conflict edge in the logic graph; 2) Priority enhancement: In power line inspection scenarios, if rule C (avoiding areas with severe weather) conflicts with rule D (completing inspection according to the predetermined route), the system can enhance the conclusion weight of rule C, making its priority much higher than D, ensuring that the drone prioritizes its own safety; 3) Priority reduction: In the above situations, the weight of rule D can also be reduced to achieve the same resolution purpose; 4) Manual intervention request: When the conflict involves the delivery of important goods (such as emergency medical supplies), and the system assesses that the risk of automatic resolution is too high, a manual intervention action is triggered, and the dispatcher manually adjusts the rule dependency graph or reassigns weights.

[0060] A reward mechanism is added to the model. There are three types of reward mechanisms, as detailed below: The fewer new conflicts that occur in the next 5 minutes, the higher the reward. Decision-making time The shorter the time, the higher the reward; Task risks in sandbox rehearsals ( The reduction rate is converted into a reward; reward function Represented as: ; in, This indicates a new conflict within 5 minutes. The maximum reward value is 8. This is expressed as the time spent on the decision-making process; Indicate whether to request human intervention; if human intervention is requested, It is 1 if it is true, otherwise it is 0; This represents the risk vector.

[0061] (b) Network architecture; Figure 2 The deep decision network in the model consists of two parallel sub-networks, a fusion layer, and a nonlinear function layer. Specifically, the state vector (generated by encoding conflict events through a state encoder) is fed into two parallel sub-networks. One branch uses a CNN convolutional neural network to extract data features, while the other branch uses a bidirectional LSTM network to process environmental information. The features from the two branches are concatenated by the fusion layer and then output as an action probability distribution (the probability of each of the four types of atomic actions) by the nonlinear function layer.

[0062] (c) Online decision-making; Online decision-making includes real-time decision-making (using a greedy selection strategy to choose the highest probability action or a random action), sandbox pre-testing and verification, and judgment based on the results of the sandbox pre-testing and verification—based on the risk vector provided by the sandbox pre-testing and verification. The decision is to either output a resolution action or re-execute the decision through the deep decision network, and then learn from the feedback (execute actions to obtain rewards, store experience, and update network parameters).

[0063] For example, such as Figure 2 As shown, the execution flow of the dynamic channel—the process of dynamically generating and resolving actions based on reinforcement learning—is as follows: The input "conflict event" is transformed into a fixed-dimensional state vector by the "state encoder"; subsequently, the state vector is input to the "deep decision network" (containing a CNN and a bidirectional LSTM dual-branch structure and a fusion layer), which outputs the "action probability distribution" of four types of atomic actions; then, a greedy selection strategy is adopted, selecting the "highest probability action" with a 90% probability and a "random action" with a 10% probability, thereby determining the "selected action"; the selected action then enters the "sandbox pre-playing" stage for "risk verification," and if the risk changes... If the value is less than the set threshold (yes), then the system will "output the resolution action", and at the same time "acquire rewards", "store experience" and "update network parameters" to form a closed-loop learning; if the risk verification fails (no), then the system will directly return to the status encoder to make a new decision.

[0064] 2.3 Dual-channel collaboration; Static channel confidence : ; in, This represents the cosine similarity corresponding to the resolution action determined through the static channel. This is the resolution success rate (preset value) corresponding to the resolution action determined by the static channel. Dynamic channel confidence : ; in, The variance of the action probability output by the dynamic channel (calculated based on the aforementioned action probability distribution, which reflects the uncertainty of action selection). To avoid the denominator being 0.

[0065] The decision fusion rules are as follows: ; if If so, the elimination action determined by the static channel is adopted as the recommended elimination action; if If the resolution action is determined by the dynamic channel, then the resolution action is adopted as the recommended resolution action. if If the human-machine collaboration process is activated, dual-channel suggestions and reasoning are pushed to designated personnel, who then manually select resolution actions and feed them back to the reinforcement learning model. The manually selected resolution actions are then used as recommended resolution actions.

[0066] 3. Simulate the execution of resolution actions to avoid chain conflicts; When a recommended resolution action is generated (such as adjusting rule priority or adding conditions to a rule), the action does not take effect immediately. Instead, the recommended resolution action is first simulated—in a virtual environment, the rules in the rule base are adjusted based on the resolution action, and potential conflicts are detected. If no new conflicts (chain conflicts) are found during the simulation, the recommended resolution action is used as the target resolution action. If a chain conflict is detected, a dynamic priority adjustment mechanism is triggered, adopting a new priority (new weight) to adjust the recommended resolution action. After determining that no chain conflicts occur, the adjusted resolution action is used as the target resolution action. Executing the target resolution action against the rule base eliminates conflicts between new and existing rules and also avoids chain conflicts.

[0067] 3.1 Simulate the execution of the resolution action; By simulating the execution of resolution actions, the rules (weights of rule conditions and rule conclusions) are modified, resulting in the generation of modified rules and the updating of the graph node set. Based on graph node sets The system checks if the rule conclusion nodes have the following conflicts: mutually exclusive conclusions or resource contention. If a conflict is found, a dynamic priority reallocation is performed. If no conflict is found, the recommended resolution action mentioned above is used as the target resolution action, and the target resolution action is executed against the rule base.

[0068] 3.2 Dynamic priority reallocation; Recalculate the weights of conflicting rule conclusions using the following formula (the larger the weight of a rule conclusion, the higher its priority): ; in, Indicates historical weight, Indicates credibility (preset value). The timeliness factor (calculated using the formula is) , This is the time when the rules have not been updated. (It is the attenuation coefficient) Indicates the conflict penalty (related to the number of conflicts and the conflict coefficient); The coefficients satisfy: ; Conflict penalty terms are represented as follows: ; This indicates the number of conflicts triggered by the rule during the most recent simulation of the resolution action; This indicates the conflict coefficient (preset value).

[0069] After dynamic priority reallocation, the recommended resolution actions are adjusted with new priorities (new weights). The adjusted resolution actions are then re-simulated and executed. If a conflict is found, dynamic priority reallocation continues, and so on until no chain conflict occurs. The adjusted resolution actions are then used as the target resolution actions, and the target resolution actions are executed against the rule base.

[0070] In summary, this application achieves localized conflict detection of rules and resolves rule conflicts by using conflict detection based on rule dependency graphs, generating collaborative rule conflict resolution actions, and simulating the execution of resolution actions to avoid chain conflicts. It also maintains the global rule logic matching, reduces the need for manual intervention, and achieves efficient conflict detection and resolution.

[0071] The following describes the device for resolving conflicts in the cooperation rules of a drone swarm provided in this application. The device for resolving conflicts in the cooperation rules of a drone swarm described below can be referred to in correspondence with the method for resolving conflicts in the cooperation rules of a drone swarm described above.

[0072] like Figure 3 As shown, this application also provides a device for resolving cooperation rule conflicts in a drone swarm, comprising: The conflict detection module 10 is used to obtain conflict events based on the UAV cluster collaboration rule base and newly added rules, through rule atomization, rule dependency graph construction and update, and conflict edge detection. The conflict resolution action recommendation module 20 is used to obtain recommended conflict resolution actions based on conflict events by processing and fusing confidence evaluation through a dual-channel conflict resolution engine. The dual-channel conflict resolution engine is built based on static and dynamic channels. The static channel is used to generate conflict resolution actions based on the existing collaborative rule conflict resolution action library, and the dynamic channel is used to generate conflict resolution actions based on reinforcement learning. The resolution action simulation execution module 30 is used to obtain the target resolution action based on the recommended resolution action by simulating execution in a virtual environment and dynamically reassigning priorities, and then execute the target resolution action on the UAV cluster collaboration rule base.

[0073] It is understood that the detailed functional implementation of each of the above units / modules can be found in the description in the aforementioned method embodiments, and will not be repeated here.

[0074] It should be understood that the above-described device is used to execute the methods in the above embodiments. The implementation principle and technical effect of the corresponding program modules in the device are similar to those described in the above methods. The working process of the device can be referred to the corresponding process in the above methods, and will not be repeated here.

[0075] Based on the methods in the above embodiments, this application provides an electronic device, such as... Figure 4 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute the methods in the above embodiments.

[0076] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0077] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0078] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0079] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0080] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0081] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0082] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0083] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for resolving a conflict of cooperation rules of a UAV cluster, characterized in that, include: Based on the drone swarm collaboration rule base and newly added rules, conflict events are obtained through rule atomization, rule dependency graph construction and updating, and conflict edge detection. Based on conflict events, a dual-channel resolution engine is used to process and fuse with confidence assessment to obtain recommended resolution actions. The dual-channel resolution engine is built on static and dynamic channels. The static channel is used to generate resolution actions based on the existing collaborative rule conflict resolution action library, and the dynamic channel is used to generate resolution actions based on reinforcement learning. Based on the recommended resolution actions, the target resolution actions are obtained by simulating execution in a virtual environment and dynamically reassigning priorities, and then the target resolution actions are executed on the drone swarm collaboration rule base.

2. The method of claim 1, wherein, The method, based on the UAV swarm collaboration rule base and newly added rules, obtains conflict events through rule atomization, rule dependency graph construction and updating, and conflict edge detection, including: Based on the condition and conclusion parts of each rule in the drone swarm collaboration rule base, a semantic parser is used to extract the condition and conclusion items to obtain the rule dependency graph; Based on the rule dependency graph and the condition and conclusion terms of the newly added rules, the graph node set of the rule dependency graph is expanded by semantic similarity comparison to obtain the updated rule dependency graph. Based on the updated rule dependency graph, conflict events are obtained by checking whether there is mutual exclusion or resource competition between the edges connected to the newly added rule conclusion nodes. The conflict events include the identifiers of the two rule conclusion nodes that are in conflict, the conflict type, and the weights of the two rule conclusion nodes that are in conflict.

3. The method of claim 1, wherein, The process, based on conflict events, involves processing and fusing the conflict events with a dual-channel resolution engine and confidence assessment to obtain recommended resolution actions, including: Based on the collision event, a hash index is performed through the static channel to obtain the static resolution action and the static channel confidence. Based on conflict events, reinforcement learning model inference is performed through dynamic channels to obtain dynamic resolution actions and dynamic channel confidence. Based on the confidence scores of static and dynamic channels, the difference is calculated and compared with a preset threshold to obtain the recommended resolution action.

4. The method of claim 3, wherein, The method of obtaining static resolution actions and static channel confidence levels based on collision events through static channel hash indexing includes: Based on the conflict type in the conflict event, partitioning is performed using a first-order hash index to obtain the conflict type partition; Based on the weights of the two conflicting rule conclusion nodes in a conflict event, logarithmic-scale bucketing is performed using a second-order hash index to obtain the target bucket within the conflict type partition. The calculation formula is as follows: ； wherein, denotes a bucketing index of a target bucket, denotes a floor operation, denotes a conflict rule weight ratio, denotes a weight corresponding to a rule conclusion numbered denotes a weight corresponding to a rule conclusion numbered denotes a rule conclusion numbered denotes a log-scale bucket width of a conflict type partition; denotes a minimum weight ratio threshold;​​​ The cosine similarity is calculated using a three-order hash index to obtain the static resolution actions and static channel confidence within the target bucket. The calculation formula is as follows: ； wherein, represents a static channel confidence, represents a cosine similarity corresponding to the static resolution action, represents a resolution success rate corresponding to the static resolution action.

5. The method of claim 3, wherein the conflict resolution of the cooperation rules of the UAV swarm is based on a priority of the UAVs. The process of using a reinforcement learning model for inference based on conflict events and dynamic channels to obtain dynamic resolution actions and dynamic channel confidence includes: Based on the conflict event, a state encoder is used to encode and obtain the state vector; Based on the state vector, features are extracted and fused through a deep decision network to obtain the action probability distribution. The deep decision network includes two parallel sub-networks, a fusion layer, and a nonlinear function layer. Based on the action probability distribution, action selection, sandbox pre-visualization verification, and risk verification are performed through online decision-making. If the risk verification passes, the dynamic resolution action and dynamic channel confidence are obtained, calculated using the following formula: ； wherein, denotes the dynamic channel confidence, denotes the action probability variance corresponding to the action probability distribution, is a constant for avoiding zero denominator.

6. The method of claim 1, wherein, The recommended resolution action, obtained through simulated execution in a virtual environment and dynamic priority reallocation, includes: Based on the recommended resolution actions, the weights of changing rule conditions and / or rule conclusions are simulated in a virtual environment to obtain the changed rules, and based on the changed rules, the rule dependency graph is simulated to be updated in the virtual environment. Based on the updated rule dependency graph obtained from the simulation, the system detects whether there are mutually exclusive conclusions or resource competition among the rule conclusion nodes and obtains the conflict detection results. If the conflict detection results indicate the existence of cascading conflicts, dynamic priority reallocation and simulation in a virtual environment are continuously performed until the cascading conflicts are eliminated and a conflict-free target resolution action is obtained. The following formula is used to dynamically reallocate the rule conclusions with conflicts: ； wherein, denotes the recalculated rule result weight, denotes the historical rule result weight, denotes the confidence, denotes the timeliness factor, denotes the conflict penalty term, , , and is a preset coefficient.

7. An apparatus for resolving a cooperation rule conflict of a UAV cluster, characterized in that, include: The conflict detection module is used to obtain conflict events based on the drone swarm collaboration rule base and newly added rules, through rule atomization, rule dependency graph construction and update, and conflict edge detection. The conflict resolution action recommendation module is used to obtain recommended conflict resolution actions based on conflict events by processing and fusing confidence evaluation through a dual-channel conflict resolution engine. The dual-channel conflict resolution engine is built based on static and dynamic channels. The static channel is used to generate conflict resolution actions based on the existing collaborative rule conflict resolution action library, and the dynamic channel is used to generate conflict resolution actions based on reinforcement learning. The resolution action simulation execution module is used to obtain the target resolution action based on the recommended resolution action by simulating execution in a virtual environment and dynamically reassigning priorities, and then execute the target resolution action on the UAV swarm collaboration rule base.

8. An electronic device, comprising: include: Memory and one or more processors; The memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions; The one or more processors invoke the computer instructions to cause the electronic device to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium comprising instructions, characterized in that: When the instructions are executed on an electronic device, the electronic device causes the electronic device to perform the method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program or instructions, characterized in that: When the computer program or instructions are run on an electronic device, the electronic device causes the electronic device to perform the method as described in any one of claims 1-6.

Images