Communication interruption time friend behavior prediction and multi-platform autonomous decision-making method and system

Abstract

The application relates to the field of avionics, and discloses a method and system for friendly behavior prediction and multi-platform autonomous decision-making in communication interruption, which comprises the following steps: constructing a multi-level fuzzy reasoning tree, fuzzy processing multi-source heterogeneous input information to form multiple fuzzy sets; decomposing complex input relations into several subsystems based on a hierarchical fuzzy reasoning mechanism, and obtaining a subtask response benefit value through layer-by-layer reasoning; and fuzzy processing input variables based on a triangular membership function, and predicting a potential task allocation scheme of a friendly formation in combination with a hierarchical rule reasoning mode. The application can realize multi-platform autonomous task allocation without a center, with low dependence and high robustness, and can improve the cooperative decision-making ability and task execution efficiency of formation members under non-perfect information conditions in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of avionics technology, and in particular to a method and system for predicting friendly behavior and making autonomous decisions across multiple platforms during communication interruptions. Background Technology

[0002] In complex and dynamic environments, large-scale distributed intelligent agent platform systems (such as drone swarms, mobile robot clusters, and edge computing node networks) face a significant contradiction between real-time performance and robustness in collaborative task allocation. On the one hand, task allocation needs to be completed within milliseconds to seconds to adapt to rapidly changing environmental conditions and task requirements. On the other hand, the system needs to continuously generate reasonable, feasible, and efficient allocation schemes under complex conditions such as heterogeneous agent capabilities (e.g., differences in perception range, computing power, energy capacity, and motion performance), a large number of agents, diverse task types and concurrent execution, and dynamic fluctuations in communication link quality (e.g., high latency, high packet loss rate, and intermittent interruptions).

[0003] Current mainstream task allocation methods are mostly based on centralized or semi-centralized architectures, where a central node collects state information, calculates the optimal solution, and distributes instructions. These methods exhibit good performance under ideal conditions of stable communication and fixed topology, but they have significant limitations: First, the central node easily becomes a system bottleneck, struggling to handle high-concurrency requests from a large number of nodes; second, in scenarios with unstable or partially interrupted communication links, the central node cannot obtain the global state in a timely manner, leading to delayed or ineffective decision-making; third, centralized architectures lack elastic scalability, making it difficult to adapt to the dynamic addition or removal of nodes.

[0004] To enhance the decentralization and fault tolerance of systems, distributed task allocation mechanisms have gradually become a research hotspot. Among them, the negotiation-based allocation method based on the Contract Net Protocol (CNP) achieves autonomous resource matching through interactive processes such as task publication, bid response, and bid confirmation, exhibiting good scalability and distributed characteristics. However, the Contract Net mechanism is highly dependent on a stable, low-latency communication environment. In practical applications, due to factors such as wireless channel interference, link jitter caused by node movement, and network congestion, the negotiation process is often prolonged due to message loss or delay, causing the task allocation cycle to exceed the available time window, making it difficult to meet real-time requirements.

[0005] In the event of communication interruption or severe degradation, when multiple distributed platform nodes are unable to exchange task status, allocation results, or intent information, each node becomes an information silo. In this situation, if the contract network mechanism, which relies on information exchange, is still used, duplicate task allocation (resource redundancy) or unattended tasks (task omissions) will inevitably occur, thereby damaging the overall task completion rate and resource utilization efficiency of the system. Especially in scenarios with multiple concurrent tasks and significant differences in task priorities, such problems can trigger a chain reaction, leading to a sharp decline in system performance.

[0006] Patent document CN117240400A provides a robust cooperative sensing method and system to resist random communication interruptions. The method includes: at each time-series keyframe, extracting features from the raw observation data input to the agent using a feature encoding model; when a communication interruption occurs, the agent, based on its stored historical features, sequentially uses a historical feature completion model and a prediction model to predict possible fusion features at the current timestamp; treating the fusion features as features of a virtual agent, performing information sharing and feature fusion, fusing the virtual agent's features with the agent's current sensing features and features of other currently communicating agents to obtain a fusion result; and using a feature decoding model to decode the fusion result to obtain the sensing result.

[0007] While the robust cooperative perception method described above attempts to address some of the problems caused by communication interruptions, it still has significant shortcomings and deficiencies. On one hand, this method focuses solely on feature completion and fusion at the perception level, failing to provide effective support for task allocation decision-making in communication interruption scenarios, and thus struggling to fundamentally resolve task conflicts and omissions. On the other hand, this method lacks comprehensive integration and efficient processing capabilities for multi-source heterogeneous information. When faced with scenarios involving significant differences in agent capabilities and complex task types, its adaptability and robustness are insufficient. It cannot accurately predict the potential behaviors of other agents and the global allocation situation, making it difficult for the system to maintain a stable and efficient operating state when communication is restricted or interrupted. Summary of the Invention

[0008] To address the problems of information sharing difficulties, lack of global coordination in task allocation decisions, and easy occurrence of task conflicts and omissions among distributed space-based platforms under extreme communication interruption environments, this invention proposes a method and system for predicting friendly behavior and autonomous decision-making across multiple platforms during communication interruptions. This method enables decentralized, low-dependency, and highly robust autonomous task allocation across multiple platforms, improving the collaborative decision-making capabilities and task execution efficiency of formation members under incomplete information conditions in complex environments.

[0009] The technical solution adopted in this invention is as follows:

[0010] A method for predicting friendly behavior during communication interruption includes:

[0011] Construct a multi-level fuzzy inference tree to fuzzify multi-source heterogeneous input information and form multiple fuzzy sets;

[0012] The hierarchical fuzzy reasoning mechanism decomposes complex input relationships into several subsystems and obtains the sub-task response benefit value through hierarchical reasoning.

[0013] The input variables are fuzzified based on the triangular membership function, and the potential task allocation scheme of the friendly formation is predicted by combining hierarchical rule reasoning.

[0014] Furthermore, the fuzzy inference tree includes an input layer, a multi-subtask response layer, and an output layer, and the multi-source heterogeneous input information includes the formation platform composition, real-time location, and subtask information.

[0015] Furthermore, the step of obtaining the subtask response benefit value through layer-by-layer reasoning includes: performing distance demand matching degree reasoning, task load degree reasoning, and demand matching degree reasoning in layers based on the multi-source heterogeneous input information after fuzzification processing, and finally outputting the subtask response benefit value.

[0016] Furthermore, it also includes encoding and learning optimization of the parameters of the fuzzy inference tree:

[0017] The membership function parameters of the fuzzy inference tree are encoded with real numbers, and the fuzzy rule base of the fuzzy inference tree is encoded with integer strings, forming a hybrid chromosome structure;

[0018] The direction of mutation is determined by the fitness difference vector between superior and inferior individuals in the population. The best individual in each generation is subjected to directional mutation separately, and the amount of directional mutation gradually decreases with the number of iterations, while retaining individuals with improved fitness.

[0019] A system for predicting friendly behavior during communication interruption includes:

[0020] The fuzzification processing module is configured to construct a multi-level fuzzy inference tree and perform fuzzification processing on multi-source heterogeneous input information to form multiple fuzzy sets;

[0021] The layer-by-layer reasoning module is configured to decompose complex input relationships into several subsystems based on a hierarchical fuzzy reasoning mechanism, and obtain the sub-task response benefit value through layer-by-layer reasoning.

[0022] The friendly behavior prediction module is configured to fuzzify the input variables based on the triangular membership function and predict the potential task allocation scheme of the friendly formation by combining hierarchical rule reasoning.

[0023] A multi-platform autonomous decision-making method for communication interruption includes:

[0024] In the event of a communication interruption, each formation, based on the last communication status information, invokes the already trained friendly behavior prediction model to predict the response benefits of other formations to the sudden mission.

[0025] Compare the effectiveness of the current formation's task allocation plan with the predicted effectiveness of the friendly formation's task allocation plan, and comprehensively evaluate whether to respond to the current task; if so, distribute the task allocation plan to the execution platform within the current formation to complete the autonomous task allocation.

[0026] Furthermore, the friendly behavior prediction model is trained through a distributed parallel real-time simulation training mechanism. The training mechanism adopts a hierarchical architecture of scheduling master node, simulation sub-node, message queue and simulation system monitoring. The master node is responsible for parameter distribution and algorithm scheduling, and the sub-nodes process chromosomes in parallel and return the fitness values.

[0027] Furthermore, the training mechanism of the distributed parallel real-time simulation also includes a fault tolerance mechanism: when a simulation sub-node crashes, the master node copies the corresponding number of chromosomes and fitness values ​​from the remaining nodes to replace the lost individuals, based on the number of non-crash nodes and the original proportion of individuals; if the crashed node does not recover, tasks are continuously distributed to other nodes until training is completed.

[0028] Furthermore, the comprehensive assessment of whether to respond to the current task includes: after generating and decomposing the emergency task, calculating the sub-tasks that the formation can complete based on the task load and real-time status of the formation; if the formation can complete all the sub-tasks corresponding to the emergency task, then performing the prediction step of the effectiveness of the task allocation scheme for friendly formations.

[0029] A multi-platform autonomous decision-making system in the event of communication interruption includes:

[0030] The response benefit prediction module is configured to, in the event of a communication interruption, each formation, based on the last communication status information, invoke the trained friendly behavior prediction model to predict the response benefits of other formations to the sudden task.

[0031] The autonomous decision-making module is configured to compare the benefits of the current formation's task allocation plan with the predicted benefits of the friendly formation's task allocation plan, and comprehensively evaluate whether to respond to the current task; if so, it will send the task allocation plan to the execution platform within the current formation to complete the autonomous task allocation.

[0032] The beneficial effects of this invention are as follows:

[0033] 1. This invention constructs a multi-level fuzzy inference tree to comprehensively integrate and fuzzify multi-source heterogeneous input information. Combined with a hierarchical inference mechanism, complex input relationships are rationally decomposed into several subsystems. Then, through layer-by-layer inference, the response benefit values ​​of sub-tasks are accurately obtained. Simultaneously, by leveraging the triangular membership function for efficient adaptation of input variables, the ability to handle uncertain information is significantly improved, making the prediction of potential task allocation schemes for friendly formations more reliable and accurate. Based on this, in scenarios where communication is limited or interrupted, each platform can, based on the last communication status information, invoke the trained friendly behavior prediction model. By comparing the benefits of its own task allocation scheme with those of friendly formations, it conducts a comprehensive evaluation and autonomously decides whether to respond to the current task and complete internal task distribution, forming a complete closed loop from information processing and behavior prediction to decision execution.

[0034] 2. This invention not only effectively breaks down information silos during communication interruptions, enabling autonomous task allocation without relying on real-time interaction and fundamentally avoiding task conflicts and omissions; it also fully integrates prior knowledge such as historical behavior patterns, task characteristics, and environmental constraints of various platforms, optimizing resource allocation based on the heterogeneous capabilities of intelligent agents, significantly improving the overall task completion rate and resource utilization efficiency of the system. Simultaneously, the multi-level inference architecture and flexible decision-making mechanism enable the system to flexibly adapt to complex scenarios with a large number of intelligent agents, diverse task types, and concurrent execution. Even in the face of communication link quality fluctuations and dynamic additions or removals of nodes, it can maintain a stable and efficient operating state, greatly enhancing the robustness and adaptability of large-scale distributed intelligent agent platform systems in dynamic and complex environments. Furthermore, this invention does not rely on centralized node scheduling support, possesses good scalability, and can easily adapt to the collaborative needs of intelligent agent systems of different sizes, providing a strong guarantee for the long-term stable operation of large-scale intelligent agent systems.

[0035] 3. Compared to CN117240400A, this invention addresses the critical need for task allocation by deeply integrating information processing and behavior prediction into the decision-making process. This achieves efficient transformation from perceived information to task allocation schemes, representing a leap in capability from basic perception to decision execution. Furthermore, the multi-level fuzzy reasoning mechanism of this invention provides more comprehensive processing of multi-source heterogeneous information and more efficient decomposition of complex relationships. Compared to single feature completion and fusion methods, it better addresses the practical needs of varying agent capabilities and complex task scenarios, exhibiting stronger environmental adaptability and decision-making accuracy. Moreover, through its benefit-comparison-based autonomous decision-making mode, this invention achieves precise matching between tasks and platform capabilities. Compared to solutions that only focus on the perception level, it directly improves the system's task completion quality and resource utilization efficiency, possessing broader practical application scenarios and higher practical value. Attached Figure Description

[0036] Figure 1 This is a flowchart of the method for predicting friendly behavior during communication interruption in Embodiment 1 of the present invention.

[0037] Figure 2 This is a flowchart of the friendly behavior prediction method in Embodiment 4 of the present invention.

[0038] Figure 3 This is a flowchart of the multi-platform autonomous task decision-making method of Embodiment 4 of the present invention. Detailed Implementation

[0039] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0040] Example 1

[0041] like Figure 1 As shown, this embodiment provides a method for predicting friendly behavior during communication interruption, including:

[0042] Construct a multi-level fuzzy inference tree to fuzzify multi-source heterogeneous input information and form multiple fuzzy sets;

[0043] The hierarchical fuzzy reasoning mechanism decomposes complex input relationships into several subsystems and obtains the sub-task response benefit value through hierarchical reasoning.

[0044] The input variables are fuzzified based on the triangular membership function, and the potential task allocation scheme of the friendly formation is predicted by combining hierarchical rule reasoning.

[0045] Furthermore, the fuzzy inference tree includes an input layer, a multi-subtask response layer, and an output layer, and the multi-source heterogeneous input information includes the formation platform composition, real-time location, and subtask information.

[0046] It should be noted that the method in this embodiment achieves effective integration and processing of multi-source heterogeneous information through the construction of a multi-level fuzzy inference tree. The hierarchical inference mechanism reduces the difficulty of processing complex input relationships, and the triangular membership function ensures the rationality of the fuzzification of input variables. The combination of these three factors improves the accuracy and reliability of predicting friendly behavior and provides an effective basis for decision-making in communication interruption scenarios.

[0047] Preferably, the fuzzy inference tree includes an input layer, a multi-subtask response layer, and an output layer. The multi-source heterogeneous input information includes formation platform composition, real-time location, and subtask information. Specifically, when constructing the fuzzy inference tree, the input layer is explicitly defined as receiving multi-source heterogeneous data such as formation platform composition, real-time location, and subtask information; the multi-subtask response layer processes and infers specific information based on different subtask response logics for each type of input; and the output layer outputs the relevant results obtained after processing and inference. In actual operation, the input layer continuously collects various multi-source heterogeneous input information and transmits it to the multi-subtask response layer. After processing by this layer, the output layer outputs the corresponding results.

[0048] It should be noted that clearly defining the hierarchical structure of the fuzzy inference tree makes the reasoning process more organized and logical. At the same time, limiting the specific types of multi-source heterogeneous input information ensures the relevance and effectiveness of the input data, providing a clear data foundation for subsequent fuzzification processing and reasoning, and improving the overall operability of the method.

[0049] Preferably, the subtask response benefit value is obtained through layer-by-layer reasoning, including: based on the fuzzified multi-source heterogeneous input information, layer-by-layer distance-demand matching degree reasoning, task load degree reasoning, and demand matching degree reasoning are performed, and finally the subtask response benefit value is output. Specifically, after obtaining the fuzzified multi-source heterogeneous input information, distance-demand matching degree reasoning is first performed to analyze the compatibility between each formation and the task execution location; then, task load degree reasoning is performed to assess the saturation of the tasks currently undertaken by each formation; finally, demand matching degree reasoning is implemented to determine the degree of fit between the formation's capabilities and the task requirements. In each layer of reasoning, calculations are performed based on the reasoning results of the previous layer and preset rules, and finally, the results of the three layers of reasoning are integrated to obtain and output the subtask response benefit value.

[0050] It should be noted that performing reasoning in layers across different dimensions allows for a comprehensive consideration of various factors affecting the response efficiency of subtasks, avoiding the limitations of single-dimensional reasoning. This makes the obtained subtask response efficiency values ​​more scientific and comprehensive, providing accurate data support for subsequent task allocation scheme predictions.

[0051] Preferably, the method in this embodiment further includes encoding and learning optimization of the parameters of the fuzzy inference tree: the membership function parameters of the fuzzy inference tree are encoded with real numbers, and the fuzzy rule base of the fuzzy inference tree is encoded with integer strings to form a hybrid chromosome structure; the direction of mutation is determined by the fitness difference vector of superior and inferior individuals in the population, and the optimal individual in each generation is subjected to directional mutation separately, and the amount of directional mutation gradually decreases with the increase of the number of iterations, while retaining individuals with improved fitness values.

[0052] Specifically, firstly, the membership function parameters of the fuzzy inference tree are encoded according to real number encoding rules, while the fuzzy rule base is encoded using integer string encoding. The two encoding results are combined to form a hybrid chromosome structure. During the learning and optimization process, the fitness difference vector between dominant and suboptimal individuals in the population is calculated, and this is used to determine the direction of mutation. For the best individual in each generation of the population, a directional mutation operation is performed separately, and the magnitude of the directional mutation is gradually reduced as the number of iterations increases. After the mutation is completed, the fitness values ​​of the individuals before and after the mutation are compared, and only individuals whose fitness values ​​have improved are retained to enter the next generation of the population.

[0053] It should be noted that the method in this embodiment adopts a hybrid chromosome structure encoding, which takes into account the encoding requirements of membership function parameters and fuzzy rule base, thereby improving the adaptability of the encoding. The directional mutation strategy, combined with the dynamic adjustment of the mutation amount, can accelerate the convergence speed of the algorithm while ensuring population diversity, retain individuals with better fitness values, and continuously improve the performance and inference accuracy of the fuzzy inference tree.

[0054] Accordingly, this embodiment also provides a friendly party behavior prediction system when communication is interrupted, including:

[0055] The fuzzification processing module is configured to construct a multi-level fuzzy inference tree and perform fuzzification processing on multi-source heterogeneous input information to form multiple fuzzy sets;

[0056] The layer-by-layer reasoning module is configured to decompose complex input relationships into several subsystems based on a hierarchical fuzzy reasoning mechanism, and obtain the sub-task response benefit value through layer-by-layer reasoning.

[0057] The friendly behavior prediction module is configured to fuzzify the input variables based on the triangular membership function and predict the potential task allocation scheme of the friendly formation by combining hierarchical rule reasoning.

[0058] Specifically, after the fuzzification processing module is started, it first completes the construction of a multi-level fuzzy inference tree, then receives multi-source heterogeneous input information, processes it according to the preset fuzzification algorithm, generates multiple fuzzy sets, and passes them to the layer-by-layer inference module. After receiving the fuzzy sets, the layer-by-layer inference module decomposes the complex input relationship into several subsystems according to the hierarchical fuzzy inference mechanism, performs inference operations on each subsystem in turn, obtains the sub-task response benefit value, and sends it to the friendly behavior prediction module. The friendly behavior prediction module uses the triangular membership function to fuzzify the input variables, and combines the built-in hierarchical rule inference logic to predict the potential task allocation scheme of the friendly formation and output the result.

[0059] It should be noted that by using a modular design, different functions are broken down, making the responsibilities of each module clear and ensuring their cooperation, which improves the maintainability and scalability of the entire system. Each module operates in an orderly manner according to preset logic, ensuring efficient connection between fuzzification processing, layer-by-layer reasoning, and friendly behavior prediction, thereby improving the overall system's operating efficiency and prediction accuracy.

[0060] Example 2

[0061] This embodiment is based on embodiment 1:

[0062] This embodiment provides a multi-platform autonomous decision-making method when communication is interrupted, including:

[0063] In the event of a communication interruption, each formation, based on the last communication status information, invokes the already trained friendly behavior prediction model to predict the response benefits of other formations to the sudden mission.

[0064] Compare the effectiveness of the current formation's task allocation plan with the predicted effectiveness of the friendly formation's task allocation plan, and comprehensively evaluate whether to respond to the current task; if so, distribute the task allocation plan to the execution platform within the current formation to complete the autonomous task allocation.

[0065] It should be noted that when communication is interrupted and real-time friendly information cannot be obtained, response benefit prediction is made based on the last communication status information and the trained model, ensuring the continuity of decision-making. By comparing the benefits of this formation with those of friendly formations, more reasonable response decisions can be made, avoiding the situation of duplicate task allocation or no response, and improving the efficiency and coordination of handling sudden tasks.

[0066] Preferably, the friendly behavior prediction model is trained through a distributed parallel real-time simulation training mechanism. This training mechanism adopts a hierarchical architecture of scheduling master node, simulation child node, message queue and simulation system monitoring. The master node is responsible for parameter distribution and algorithm scheduling, and the child nodes process chromosomes in parallel and return the fitness value.

[0067] Specifically, when training the friendly behavior prediction model, a hierarchical training environment is built, including a scheduling master node, multiple simulation sub-nodes, a message queue, and simulation system monitoring. The master node first determines the parameters required for model training and distributes them to each simulation sub-node through the message queue. It is also responsible for the overall scheduling of the algorithm and coordinating the work progress of each sub-node. After receiving the parameters, each simulation sub-node processes the assigned chromosome in parallel, including operations such as encoding, mutation, and fitness calculation. After completion, it sends the calculated fitness value back to the master node through the message queue. The simulation system monitors the running status of the master node and each sub-node in real time to ensure the stable progress of the training process.

[0068] It should be noted that the training mechanism of the distributed parallel real-time simulation adopts a layered architecture, which realizes a clear division of labor between the master node and the child nodes. The centralized scheduling of the master node ensures the orderliness of training, while the parallel processing of the child nodes greatly improves the training efficiency and shortens the model training cycle. The introduction of message queues ensures the reliability and real-time performance of data transmission, and the simulation system monitoring reduces the risk of failure during training, ensuring that the model training can be completed smoothly.

[0069] Preferably, the training mechanism of the distributed parallel real-time simulation also includes a fault tolerance mechanism: when a simulation sub-node crashes, the master node copies the corresponding number of chromosomes and fitness values ​​from the remaining nodes to replace the lost individuals, based on the number of non-crash nodes and the original proportion of individuals; if the crashed node does not recover, the task is continuously distributed to other nodes until training is completed.

[0070] Specifically, during the distributed parallel real-time simulation training process, the simulation system continuously monitors the running status of each simulation sub-node. When a simulation sub-node is detected to have crashed, the system immediately feeds this information back to the master node. The master node counts the number of currently running nodes and, based on the proportion of individuals originally allocated to the crashed node, copies the corresponding number of chromosomes and their corresponding fitness values ​​from the remaining running nodes to replace the individual data lost by the crashed node, ensuring the integrity of the training data. If the crashed node does not recover within a certain period of time, the master node adjusts the task allocation strategy, continuously distributing the tasks originally allocated to the crashed node to other running nodes until the entire model training task is completed.

[0071] It should be noted that the introduction of the fault tolerance mechanism effectively addresses issues such as training interruption and data loss that may result from the failure of simulation sub-nodes, ensuring the continuity and stability of the training process. Through reasonable task and data substitution strategies, it avoids training failure or deviation of training results due to individual node failures, thereby improving the reliability and robustness of the distributed parallel real-time simulation training mechanism.

[0072] Preferably, the comprehensive evaluation of whether to respond to the current task includes: after generating and decomposing the emergency task, calculating the sub-tasks that the formation can complete based on the task load and real-time status of the formation; if the formation can complete all the sub-tasks corresponding to the emergency task, then performing the prediction step of the benefit of the task allocation scheme of the friendly formation.

[0073] Specifically, when a sudden task occurs, it is first decomposed into several interrelated sub-tasks according to the nature, requirements, and complexity of the task. Next, the current task load of the formation is statistically analyzed, including the number of tasks already undertaken, progress, and resource usage. Simultaneously, real-time status information of the formation is collected, such as the operating status of the execution platform and remaining resources. Based on the above information and the execution requirements of the sub-tasks, the range of sub-tasks that the formation is capable of completing is calculated. If the calculation results show that the formation can independently complete all sub-tasks corresponding to the sudden task, the friendly behavior prediction model is activated to predict the effectiveness of the friendly formation's task allocation scheme. If the formation cannot complete all sub-tasks, it directly proceeds to other subsequent evaluation processes.

[0074] It should be noted that breaking down the emergency task and assessing the task completion capability of the formation in advance can avoid blindly predicting the benefits of friendly forces and reduce unnecessary computational overhead. Predicting the benefits of friendly forces is only carried out when the formation has the ability to independently complete all sub-tasks. This ensures the pertinence and rationality of subsequent decisions, while also making more efficient use of resources and improving the overall efficiency of handling emergency tasks.

[0075] Accordingly, this embodiment also provides a multi-platform autonomous decision-making system in the event of communication interruption, including:

[0076] The response benefit prediction module is configured to, in the event of a communication interruption, each formation, based on the last communication status information, invoke the trained friendly behavior prediction model to predict the response benefits of other formations to the sudden task.

[0077] The autonomous decision-making module is configured to compare the benefits of the current formation's task allocation plan with the predicted benefits of the friendly formation's task allocation plan, and comprehensively evaluate whether to respond to the current task; if so, it will send the task allocation plan to the execution platform within the current formation to complete the autonomous task allocation.

[0078] Specifically, when a communication interruption occurs, the response benefit prediction module is automatically triggered. It retrieves the status information from the last communication of each formation and calls a pre-trained, mature friendly behavior prediction model through a preset interface. The model uses the input status information to perform calculations and obtain the response benefit prediction results of other formations for the emergency task. This result is then transmitted to the autonomous decision-making module. After receiving the prediction results, the autonomous decision-making module calculates the benefit of its own formation's task allocation plan for the emergency task. It performs a multi-dimensional comparative analysis of the two and, based on preset evaluation criteria and decision-making rules, comprehensively judges whether its own formation should respond to the current emergency task. If the evaluation conclusion is to respond, the autonomous decision-making module generates a task allocation plan and distributes it to each execution platform within the formation through an internal data transmission channel. After receiving the plan, the execution platform initiates relevant operations to complete the autonomous task allocation.

[0079] It should be noted that, through the collaborative work of the response benefit prediction module and the autonomous decision-making module, the system in this embodiment realizes the automation and intelligence of autonomous decision-making by multiple platforms in the case of communication interruption. The response benefit prediction module provides key friendly information reference for decision-making, while the autonomous decision-making module makes reasonable decisions through scientific comparison and evaluation, ensuring the efficiency and accuracy of task allocation and improving the collaborative combat capability and emergency handling capability of multiple platforms in the case of communication interruption.

[0080] Example 3

[0081] In extreme communication disruption environments, distributed space-based platforms suffer from problems such as ineffective task allocation decisions, frequent task conflicts and omissions due to a lack of real-time information exchange, and the inability of traditional centralized or decentralized task allocation models to adapt to dynamic and uncertain environments.

[0082] Based on this, this embodiment provides a method for predicting friendly behavior when communication is interrupted, including: designing a multi-level matching decision tree from task requirements to resource status based on fuzzy mathematics technology and drawing on the hierarchical optimization solution idea, constructing a friendly behavior prediction model, organizing the resource relationships between multiple platforms, and predicting the most likely friendly task allocation scheme.

[0083] Preferably, the method for predicting friendly behavior during communication interruption in this embodiment includes the following steps:

[0084] First, a multi-level fuzzy inference tree structure is constructed, dividing the input information into an input layer, a multi-subtask response layer, and an output layer. The input layer contains heterogeneous data from multiple sources, such as the composition of the formation platform, real-time location, subtask information, winning bid status, and historical bidding response rate. After fuzzification processing, multiple fuzzy sets are formed.

[0085] Then, through a hierarchical fuzzy inference mechanism, the complex input relationship is decomposed into several subsystems, and the distance demand matching degree FIS (Fuzzy Inference System), task load degree FIS, and demand matching degree FIS are used to perform layer-by-layer inference, and finally the sub-task response benefit value is output.

[0086] Finally, the input variables are fuzzified based on the triangular membership function, and a hierarchical rule-based reasoning approach is adopted to reduce the risk of combinatorial explosion, thereby achieving efficient and accurate prediction of potential task allocation schemes for friendly formations.

[0087] Accordingly, this embodiment provides a fast learning and training method for a friendly behavior prediction model that integrates a directed mutation genetic algorithm, including: an efficient learning algorithm for fuzzy inference parameters based on the directed mutation genetic algorithm and a distributed parallel fast training mechanism, to achieve rapid training and generation of optimal model parameters, autonomously make up for missing, correct or improve decision knowledge, and support rapid iterative upgrades of resource self-organization mode.

[0088] Preferably, the fast learning and training method for the friendly behavior prediction model in this embodiment includes the following steps:

[0089] First, the membership function parameters in the fuzzy inference tree and the fuzzy rule base are encoded into a chromosome structure, and a unified representation of the parameters of the multi-subsystem is achieved through segmented encoding.

[0090] Then, a mutation operator based on directed mutation is designed, which uses the fitness difference vector of individuals in the population to determine the mutation direction, guides the search process to converge faster in the direction of optimization, and performs separate tracking and mutation on the best individual in each generation to improve the global optimization efficiency.

[0091] Finally, a distributed parallel ultra-real-time simulation training platform is constructed. It relies on the RabbitMQ message queue to realize chromosome distribution and fitness value backhaul. Combined with hierarchical parallel computing and fault tolerance mechanism, it realizes efficient iterative update and fast convergence of model parameters.

[0092] In addition, this embodiment also provides a multi-platform autonomous task decision-making method based on friendly behavior prediction, including: constructing an autonomous task allocation strategy for formation based on a model based on friendly behavior prediction, so as to realize autonomous decision-making for multi-platform collaborative tasks under communication restrictions or interruptions.

[0093] Preferably, the multi-platform autonomous task decision-making method of this embodiment includes the following steps:

[0094] First, in the event of a communication interruption, each formation, based on the last communication status information, invokes the already trained friendly behavior prediction model to predict the response benefits of other formations to the sudden task.

[0095] Then, by comparing the effectiveness of its own task allocation plan with the predicted effectiveness of the friendly plan, a comprehensive evaluation is made on whether to respond to the current task, thus achieving rational decision-making in a non-cooperative environment.

[0096] Finally, if the decision is successful, the task allocation plan will be sent to the execution platform within the formation to complete the autonomous task allocation, ultimately forming a consistent and conflict-free autonomous decision result.

[0097] Example 4

[0098] This embodiment is based on embodiment 3:

[0099] like Figure 2 The flowchart of the friendly behavior prediction method in this embodiment is shown. Based on the multi-subtask response benefits, a resource self-organizing fuzzy inference tree is designed, which is mainly divided into three layers: input layer, response layer, and output layer. Specifically, the input information includes the formation platform composition (platform type, number of platforms), real-time formation location (formation center location, formation distribution range, formation type), subtask information (task type, task area, task time), successful bid status (task type, task time, successful bid platform information), historical bidding response rate (task type, task response rate, response time), etc.; the output information includes task response benefits, where 0 represents no response, and the interval (0, 1] represents the task response benefit value information. The entire fuzzy tree contains 6 fuzzy input variables, 1 fuzzy output variable, and 4 fuzzy inference systems.

[0100] This embodiment uses the example of a drone formation collaboratively executing multiple sub-tasks to illustrate the specific implementation method. Specifically, the friendly behavior prediction method in this embodiment can adopt the following steps:

[0101] 1. The first layer is the input layer. The input information includes the composition of the formation platform (platform type, number of platforms), the real-time location of the formation (formation center location, formation distribution range, formation type), sub-task information (task type, task area, task time), the status of successful bids (task type, task time, information of successful bidders), and the historical bid response rate (task type, task response rate, response time), etc.

[0102] Preferably, the six fuzzy input variables in the input layer are as follows:

[0103] a) Number of corresponding platforms: Based on the number of members in the subtask receiving formation, it is divided into 5 fuzzy sets: many, relatively many, medium, relatively few, and few.

[0104] b) Formation center position: divided into 5 fuzzy sets: far, relatively far, medium, relatively near, and near;

[0105] c) Formation distribution range: divided into 5 fuzzy sets: far, relatively far, medium, relatively near, and near;

[0106] d) Sub-task region: divided into 5 fuzzy sets: far, relatively far, medium, relatively near, and near;

[0107] e) Number of successful bids: Divided into 5 fuzzy sets: many, relatively many, medium, relatively few, and few;

[0108] f) Historical response rate: divided into 5 fuzzy sets: high, relatively high, medium, relatively low, and low;

[0109] 2. The second layer is the multi-subtask response layer. It primarily infers the subtask response benefits based on the matching degree of each resource to each subtask and the task load. Taking the FT calculation of subtask response benefits as an example, the specific fuzzy tree structure within each layer is explained. It integrates demand information such as the number of corresponding platforms, the formation center location, the formation distribution range, the subtask area, the number of successful bids, and historical response rates. Based on the relationships between the input elements, it constructs the FT calculation for subtask response benefits and finally outputs the subtask response benefit value.

[0110] Preferably, the four fuzzy inference variables in the multi-subtask response layer are as follows:

[0111] a) Distance-to-demand matching degree (FIS);

[0112] b) Task load factor (FIS);

[0113] c) Demand Matching Index (FIS);

[0114] d) Subtask Response Benefits (FIS).

[0115] 3. The third layer is the output layer. The output information includes task response benefits, where 0 represents no response and the interval (0, 1) represents the task response benefit value information.

[0116] Preferably, one fuzzy output variable in the output layer is as follows:

[0117] a) Subtask response benefits: Divided into 5 fuzzy sets: high, relatively high, medium, relatively low, and low.

[0118] Each fuzzy inference system has two inputs and one output. The relationship between the inputs and output can be represented by fuzzy inference rules. Taking the Attack Time Matching System (FIS) as an example, it has 25 rules, which can be represented as follows:

[0119] If the formation center is far and the sub-task area is near, then the distance requirement matching degree is low.

[0120] If the formation center is far away and the sub-task area is far away, then the distance requirement is highly matched.

[0121] In friendly behavior prediction models, the quality of input-output fuzzification and the setting of fuzzy rule sets have a significant impact on inference results. Addressing the issues of slow updates and iterations and difficulty in guaranteeing optimal performance in traditional manual offline modification functions, this embodiment employs a distributed parallel ultra-real-time simulation training mechanism based on the classic genetic algorithm. This mechanism shortens model training time through the rapid parallel learning of multiple models.

[0122] Due to the large size of the parameters (fuzzy membership function and fuzzy inference rule base) of the friendly behavior prediction model, which are difficult to manually modify and adjust, this embodiment designs an adaptive genetic algorithm based on directed mutation to support the efficient learning and training of the friendly behavior prediction model, taking advantage of the fact that the model parameters are easy to encode and represent.

[0123] Preferably, the adaptive genetic algorithm based on directed mutation follows this process:

[0124] a) Determining the mutation direction based on the difference in individual fitness values: The mutation direction reflects the trend of improvement in the fitness function value along a certain direction at a certain point in the search space. If the mutation direction is chosen as the direction for the next step of the genetic algorithm's directed mutation, the objective function can be optimized with a higher probability. Two individuals with different fitness values ​​are selected from the population. By subtracting individuals with poorer fitness from those with better fitness, we can obtain the direction of individual variation. :

[0125]

[0126] Among them, the function f (X) represents the fitness function of individual X, which can be understood as the function from vector X to fitness value. f A mapping of (X). For example, if X is a two-dimensional vector, f (X) = -X·X T If X approaches (0,0), then the fitness value... f The larger (X) is.

[0127] b) If the fitness value improves after targeted mutation, the individual is retained; otherwise, it is discarded.

[0128] c) The magnitude of the directional variation gradually decreases as the number of iterations increases.

[0129] If a feasible solution has been determined The direction of mutation at the location is ,exist The formula for directional mutation is as follows:

[0130]

[0131] In the formula, For genes after directional mutation, For the number of iterations, The maximum number of iterations, These are uniformly distributed random numbers.

[0132] d) Individual mutation of the optimal individual: The optimal individual in each generation has the highest probability of approaching the global optimum relative to other individuals. The optimal individual retained in a certain generation of the population is... The better individual is The direction of mutation is If the fitness value improves after performing targeted mutation according to the targeted mutation formula, then replace the original individual with a new one. Otherwise keep constant.

[0133] This embodiment uses the original model for predicting friendly behavior as an example to illustrate the specific implementation method. Specifically, the fast learning and training method for the friendly behavior prediction model in this embodiment can adopt the following steps:

[0134] 1. Rapid learning of model parameters based on directed mutation genetic algorithm.

[0135] First, the parameters of the fuzzy inference tree (including the membership function vertex position, left and right width, and fuzzy rule base) are encoded into a hybrid chromosome structure, where the membership function parameters are encoded with real numbers and the rule base is encoded with integer strings.

[0136] Then, a fitness value calculation method based on sub-task response benefits is introduced, and the effectiveness of the generated task allocation scheme is evaluated by combining the grey hierarchical analysis method, which serves as the individual fitness.

[0137] Secondly, a mutation operator based on directed mutation is designed. The direction of mutation is determined by the difference vector of fitness values ​​between superior and inferior individuals in the population, which guides the mutation towards the optimization direction. The optimal individual in each generation is subjected to directed mutation separately to improve the convergence speed.

[0138] Finally, efficient learning of model parameters is achieved through iterative optimization.

[0139] 2. Distributed parallel ultra-real-time simulation training mechanism accelerates the learning process.

[0140] First, a hierarchical architecture based on a scheduling master node, simulation child nodes, RabbitMQ message queues, and simulation system monitoring is constructed. The master node is responsible for genetic algorithm scheduling and parameter distribution, while the child nodes run multiple simulation programs to process chromosomes in parallel.

[0141] Then, the chromosomes of each generation are pushed into the queue through the chromosome distribution message queue. Each simulation program, as a consumer, continuously obtains chromosomes and performs fast simulation. After the simulation is completed, the fitness value is sent back to the master node.

[0142] Finally, the master node collects all fitness values ​​to generate a new generation of chromosomes, which advances the algorithm iteration and achieves ultra-real-time parallel training.

[0143] 3. Establish a fault-tolerance mechanism to ensure training stability.

[0144] First, a "at most once" fault tolerance level is adopted to avoid performance loss caused by individual simulations waiting to be recalculated due to node failures.

[0145] Then, when a slave node is detected to have crashed, the master node, based on the number of surviving nodes and the original proportion of individuals, randomly copies the corresponding number of chromosomes and fitness values ​​from the remaining nodes to replace the lost individuals.

[0146] Finally, if the failed node recovers, normal scheduling continues; if it does not recover, tasks are continuously distributed to other nodes until training is complete, thus ensuring the robustness and integrity of the training process.

[0147] like Figure 3 This diagram illustrates the flowchart of the multi-platform autonomous task decision-making method based on friendly behavior prediction in this embodiment. In the event of a formation communication interruption, formation members cannot directly share situational awareness and individual member states via communication. Therefore, it is necessary to predict the behavior of other friendly formation members, forecasting their current possible formation behavior and potential task allocation schemes. Each formation member must autonomously calculate task allocation under incomplete information to determine its required tasks, avoiding task conflicts and omissions. By constructing a formation task autonomous allocation model based on a trained friendly behavior prediction fuzzy logic tree, this model addresses the autonomous decision-making problem of multi-platform task allocation under communication interruption conditions. This formation task autonomous allocation model aims to avoid task conflicts and omissions, improving the overall formation's task execution efficiency and task completion quality.

[0148] Specifically, the multi-platform autonomous task decision-making method of this embodiment can adopt the following steps:

[0149] 1. Based on the task generation and decomposition, calculate the sub-tasks that the formation can complete based on information such as the task load and real-time status of the formation.

[0150] 2. If this formation can complete all the sub-tasks corresponding to the objective, then predict the behavior of friendly formations;

[0151] 3. Based on historical information such as the composition, real-time location, successful bids, and historical bid response rate of friendly fleet platforms, call the friendly behavior prediction model to accurately predict the effectiveness of the task allocation scheme for listening to friendly fleets.

[0152] 4. Compare the effectiveness of our task allocation plan with the predicted effectiveness of the friendly task allocation plan, and decide whether our formation should respond to the emergency.

[0153] 5. If the decision is made to respond to the emergency, the task allocation plan will be issued to the corresponding execution platform within the formation.

[0154] Example 5

[0155] This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method for predicting friendly behavior and making autonomous decisions on multiple platforms when communication is interrupted, as described in Embodiment 1 or Embodiment 3. The computer program can be in the form of source code, object code, executable file, or some intermediate form.

[0156] Example 6

[0157] This embodiment provides a computer-readable storage medium storing a computer program. When executed by a processor, this computer program implements the method for predicting friendly behavior and making autonomous decisions on multiple platforms during communication interruptions, as described in Embodiment 1 or Embodiment 3. The computer program can be in the form of source code, object code, executable file, or some intermediate form. The storage medium includes any entity or device capable of carrying computer program code, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0158] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

[0159] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

Claims

1. A method for predicting friendly behavior and making autonomous decisions across multiple platforms during communication interruptions, characterized in that, include: Construct a multi-level fuzzy inference tree to fuzzify multi-source heterogeneous input information and form multiple fuzzy sets; The fuzzy inference tree includes an input layer, a multi-subtask response layer, and an output layer. The multi-source heterogeneous input information includes the formation platform composition, real-time location, and subtask information. The hierarchical fuzzy reasoning mechanism decomposes complex input relationships into several subsystems and obtains sub-task response benefit values ​​through hierarchical reasoning. The hierarchical reasoning includes performing distance demand matching degree reasoning, task load degree reasoning, and demand matching degree reasoning in layers based on multi-source heterogeneous input information after fuzzification, and finally outputting sub-task response benefit values. The input variables are fuzzified based on the triangular membership function, and the potential task allocation scheme of the friendly formation is predicted by combining hierarchical rule reasoning. In the event of a communication interruption, each formation, based on the last communication status information, invokes the already trained friendly behavior prediction model to predict the response benefits of other formations to the sudden mission. Compare the effectiveness of the current formation's task allocation plan with the predicted effectiveness of the friendly formation's task allocation plan, and comprehensively evaluate whether to respond to the current task. If a response is received, the task allocation plan will be sent to the execution platform within this formation to complete the autonomous task allocation; The learning and training method of the friendly behavior prediction model includes: encoding the membership function parameters and fuzzy rule base in the fuzzy inference tree into a chromosome structure, and realizing the unified representation of multi-subsystem parameters through segmented encoding; designing a mutation operator based on directed mutation, using the fitness difference vector of individuals in the population to determine the mutation direction, guiding the search process towards the optimization direction to accelerate convergence, and implementing separate tracking and mutation for the best individual in each generation; constructing a distributed parallel ultra-real-time simulation training platform, relying on the RabbitMQ message queue to realize chromosome distribution and fitness value backhaul, and combining hierarchical parallel computing and fault tolerance mechanism to realize iterative update and convergence of model parameters.

2. The method for predicting friendly behavior and multi-platform autonomous decision-making during communication interruption as described in claim 1, characterized in that, It also includes encoding and learning optimization of the parameters of the fuzzy inference tree: The membership function parameters of the fuzzy inference tree are encoded with real numbers, and the fuzzy rule base of the fuzzy inference tree is encoded with integer strings, forming a hybrid chromosome structure; The direction of mutation is determined by the difference in fitness values ​​between superior and inferior individuals in the population. The best individual in each generation is subjected to directional mutation separately, and the amount of directional mutation gradually decreases with the number of iterations, while retaining individuals with improved fitness values.

3. The method for predicting friendly behavior and multi-platform autonomous decision-making during communication interruption as described in claim 1, characterized in that, The friendly behavior prediction model is trained through a distributed parallel real-time simulation training mechanism. The training mechanism adopts a hierarchical architecture of scheduling master node, simulation sub-node, message queue and simulation system monitoring. The master node is responsible for parameter distribution and algorithm scheduling, and the sub-nodes process chromosomes in parallel and return the fitness value.

4. The method for predicting friendly behavior and making autonomous decisions across multiple platforms during communication interruptions as described in claim 3, characterized in that, The training mechanism of the distributed parallel real-time simulation also includes a fault tolerance mechanism: when a simulation sub-node crashes, the master node copies the corresponding number of chromosomes and fitness values ​​from the remaining nodes to replace the lost individuals, based on the number of undisturbed nodes and the original proportion of individuals allocated; if the crashed node does not recover, tasks are continuously distributed to other nodes until training is completed.

5. The method for predicting friendly behavior and making autonomous decisions across multiple platforms during communication interruptions as described in claim 1, characterized in that, The comprehensive assessment of whether to respond to the current task includes: after generating and decomposing the emergency task, calculating the sub-tasks that the formation can complete based on the task load and real-time status of the formation; if the formation can complete all the sub-tasks corresponding to the emergency task, then performing the prediction step of the effectiveness of the task allocation scheme of the friendly formation.

6. A system for predicting friendly behavior and making autonomous decisions across multiple platforms during communication interruptions, characterized in that: include: The fuzzification processing module is configured to construct a multi-level fuzzy inference tree and perform fuzzification processing on multi-source heterogeneous input information to form multiple fuzzy sets; The fuzzy inference tree includes an input layer, a multi-subtask response layer, and an output layer. The multi-source heterogeneous input information includes the formation platform composition, real-time location, and subtask information. The layer-by-layer reasoning module is configured to decompose complex input relationships into several subsystems based on a hierarchical fuzzy reasoning mechanism, and obtain the sub-task response benefit value through layer-by-layer reasoning; the layer-by-layer reasoning includes performing distance demand matching degree reasoning, task load degree reasoning and demand matching degree reasoning in layers based on multi-source heterogeneous input information after fuzzification, and finally outputting the sub-task response benefit value. The friendly behavior prediction module is configured to fuzzify the input variables based on the triangular membership function and combine hierarchical rule reasoning to predict the potential task allocation scheme of the friendly formation. The response benefit prediction module is configured to, in the event of a communication interruption, each formation, based on the last communication status information, invoke the trained friendly behavior prediction model to predict the response benefits of other formations to the sudden task. The autonomous decision-making module is configured to compare the effectiveness of the current formation's task allocation plan with the predicted effectiveness of the friendly formation's task allocation plan, and comprehensively evaluate whether to respond to the current task; if it responds, it will send the task allocation plan to the execution platform within the current formation to complete the autonomous task allocation. The learning and training method of the friendly behavior prediction model includes: encoding the membership function parameters and fuzzy rule base in the fuzzy inference tree into a chromosome structure, and realizing the unified representation of multi-subsystem parameters through segmented encoding; designing a mutation operator based on directed mutation, using the fitness difference vector of individuals in the population to determine the mutation direction, guiding the search process towards the optimization direction to accelerate convergence, and implementing separate tracking and mutation for the best individual in each generation; constructing a distributed parallel ultra-real-time simulation training platform, relying on the RabbitMQ message queue to realize chromosome distribution and fitness value backhaul, and combining hierarchical parallel computing and fault tolerance mechanism to realize iterative update and convergence of model parameters.

Images