A multi-intelligent cooperative method and device for distributed end devices

Abstract

The application provides a multi-intelligent cooperative method and device for distributed terminal equipment. The system comprises: a distributed global network perception module, which is used for collecting and synchronizing global network states and intelligent agent capability information in a distributed autonomous network, and outputting standardized information to a distributed intelligent decision module; the distributed intelligent decision module is used for generating cluster configuration and task execution strategy based on global information and task targets provided by the perception module, and issuing the cluster configuration and the task execution strategy to a multi-intelligent task execution module; the multi-intelligent task execution module is used for feeding back intelligent agent states and task progress to the perception module and the decision module in real time according to the cluster configuration and the strategy issued by the decision module, so as to form a closed-loop cooperative link of "perception-decision-execution-feedback". Cross-layer data interaction and standardized information transmission are realized through a Hongmeng distributed soft bus and Google A2A protocol adaptation and extension between the modules.

Description

Technical Field

[0001] This invention relates to the fields of Internet of Things and artificial intelligence technology, and in particular to a multi-intelligent collaborative method and apparatus for distributed end devices. Background Technology

[0002] With the rapid development of IoT and AI technologies, multi-agent task collaboration on edge intelligent devices (such as drones, unmanned dogs, androids, and mobile tablets) has become a core trend in industrial applications. The relevant supporting technology fields mainly cover the following core directions: Device interconnection and communication, and intelligent agents, are detailed below: 1. Multi-Agent Interaction Protocol: Google's A2A protocol (Multi-Agent Interaction Protocol) provides a standardized communication specification for the interaction of capabilities between multiple agents, enabling capability description and interaction of multiple agents based on a unified protocol, and laying the foundation for multi-agent collaboration.

[0003] 2. Device Interconnection Technology: HarmonyOS's distributed soft bus technology, relying on the characteristics of its distributed architecture, possesses core advantages such as automatic device discovery, low-latency transmission, multi-media adaptation (adapting to various wireless transmission media such as Bluetooth, WiFi, and RF), and self-organizing networking. It can achieve seamless connection and high-speed data transmission between heterogeneous device-side devices. At the same time, by shielding the underlying technical difficulties of cross-terminal collaboration through core capabilities such as a common communication platform and distributed data management, it provides a unified communication interface for upper-layer applications, solving the core pain point of basic communication between multiple device-side devices.

[0004] (a) Insufficient adaptation of A2A protocol to end-side scenarios The A2A protocol lacks practicality in edge-distributed multi-device scenarios, failing to support collaborative decision-making among multiple agents based on network status, and easily leading to task collaboration failures. The reason is that the core design of the A2A protocol focuses on standardized interaction of capabilities between multiple agents, without fully considering the specificities of edge-distributed scenarios. It lacks a standardized description module for the network communication status between peer devices, resulting in the protocol itself lacking network status awareness capabilities and failing to meet the network status awareness requirements during multi-device agent collaboration.

[0005] (ii) Insufficient transmission efficiency of traditional TCP / IP protocol The TCP / IP communication protocol struggles to meet the high-efficiency transmission requirements of multi-media, multi-hop devices on the edge, and cannot support the core capabilities of low latency, self-discovery, self-organizing networks, and self-organization among multiple intelligent agents, thus limiting the flexibility and reliability of collaborative systems. The reason is that the TCP / IP protocol is a general-purpose network communication protocol, and its design is not optimized for complex scenarios involving multiple media (Bluetooth, WiFi, RF, satellite networks, etc.) and multi-hop connections on the edge. It cannot adapt to the dynamic changes in heterogeneous network environments on the edge, thus making it difficult to achieve the core transmission capabilities required for low latency, self-organizing networks, and other aspects of multi-agent collaboration.

[0006] (III) Insufficient integration and collaboration of HarmonyOS distributed soft bus HarmonyOS distributed soft bus technology is designed as a basic communication support between edge devices. Its core objective is to solve the connection and data transmission problems of heterogeneous devices. However, it has not planned multi-agent collaboration as a core application scenario, resulting in a lack of adaptability design between it and multi-agent collaboration technology, and thus failing to achieve deep integration at the technical level. Summary of the Invention

[0007] The purpose of this invention is to provide a multi-agent collaborative construction method based on HarmonyOS distributed soft bus. This method aims to achieve efficient collaborative construction of multiple agents on multiple device nodes such as drones and unmanned aerial vehicles (UAVs). By integrating the A2A protocol with HarmonyOS distributed soft bus, it solves network perception and transmission problems, provides decision-making basis for multi-agent collaborative reasoning, and thus comprehensively improves the overall performance of the collaborative system.

[0008] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution.

[0009] A three-level collaborative system based on distributed terminal devices and multiple agents, the system comprising: The distributed global network perception module is used to collect and synchronize global network status and agent capability information within the distributed autonomous network, and output standardized information to the distributed intelligent decision-making module. The distributed intelligent decision-making module is used to complete the election of intelligent agent nodes and the construction of a multi-agent cluster based on the global information and task objectives provided by the perception module, according to the preset modeling and election logic, and generate cluster configuration and task execution strategy and distribute them to the multi-intelligent task execution module. The multi-intelligent task execution module is used to execute specific tasks based on the cluster configuration and strategy issued by the decision module, relying on the division of labor and cooperation of dual-role (Planning and Working) intelligent agents within the cluster, and to feed back the status of the intelligent agents and the task progress to the perception module and the decision module in real time, forming a closed-loop collaborative link of "perception-decision-execution-feedback". The modules communicate with each other through the HarmonyOS distributed soft bus and the Google A2A protocol adaptation extension to achieve cross-layer data interaction and standardized information transmission.

[0010] Furthermore, the distributed global network awareness module uses the four-layer architecture of HarmonyOS distributed soft bus—"device discovery-connection management-data transmission-status monitoring"—as its underlying communication support, and adopts a "dual-mode" global information synchronization mechanism, including: periodic distributed autonomous broadcast synchronization and instantaneous triggered update synchronization based on preset events; the broadcast period can be dynamically adjusted according to the network scale, with a default value of 8 seconds; instantaneous triggered updates stop periodic broadcasting when a trigger event occurs and push incremental update data packets containing only changed fields to the entire network.

[0011] Furthermore, the distributed global network awareness module periodically broadcasts data packets using the NetworkPerceptionInfo subfield set defined by the A2A protocol extension, and maps the following fields to the HarmonyOS distributed soft bus protocol fields: NetID, MediumType, HopPath, Delay, Bandwidth, and LinkQuality are mapped to core fields of the soft bus such as network ID, connection type, topology information, delay, bandwidth, and link quality. AgentID, AgentCapability, and AliveStatus are combined with DeviceID, DeviceType from the soft bus and the agent's own state to generate the data. During broadcasting, the DeviceID of the soft bus is used as the source identifier, and ConnType specifies the transmission medium; when an event is triggered, only the changed protocol fields are pushed and the corresponding NetID and AgentID identifiers are included to reduce communication overhead.

[0012] Furthermore, the distributed global network awareness module also includes an information verification and integration unit, which is used to first parse the CheckSum verification field of the A2A protocol and combine it with the EncryptType field of the HarmonyOS distributed soft bus to verify data integrity and security; after the verification is passed, the network ownership is confirmed according to NetID and the local global information table is updated according to AgentID, and finally a standardized global information table is output to the distributed intelligent decision-making module.

[0013] Furthermore, the distributed intelligent decision-making module implements the construction and dynamic adaptation of multi-agent clusters using the logic of "modeling and quantification - two-stage election - configuration implementation," specifically including: A mathematical model is constructed to quantify data such as network status and agent capabilities provided by the perception module into a single-node comprehensive fitness score. The single-node score integrates three dimensions: link quality, capability matching, and survival stability, and dynamically allocates weights according to task requirements (the sum of each weight is 1). The mathematical model for the single-node comprehensive fitness score is implemented through the following formula:

[0014] The standardized link quality score Snet,i is calculated using the following formula:

[0015] With the goal of achieving optimal overall cluster collaborative efficiency, the core objective function for achieving optimal overall cluster collaborative efficiency is:

[0016] By setting capability matching constraints, network performance constraints, and scale and redundancy constraints, the election results are ensured to meet the task execution requirements through these four constraints. The core constraints and principles are as follows: Capability matching constraints: Extracting based on the AgentCapability field This ensures that the cluster possesses the core capabilities for the task. Network performance constraints: Verify and perceive the Delay and LinkQuality fields to ensure network compliance; Size and redundancy constraints: The number of core subtasks, This is to reserve resources for redundant nodes and for dynamic re-election.

[0017] Based on the constraints, during the initial election phase, a candidate set is selected from all sensing nodes, sorted by score, and nodes that meet the size and redundancy requirements are selected to build a cluster. In the initial election, based on the scoring results, 1 to 2 nodes with the highest overall score and excellent network stability (e.g., LinkQuality≥90) are selected as Planning agents, and the rest are selected as Working agents. Each Working agent is then bound to a backup node with the second highest score in the candidate set as redundancy.

[0018] Furthermore, the distributed intelligent decision-making module also includes a dynamic re-election mechanism: When any of the abnormal triggering conditions are met, the Planning agent will trigger a re-election. The abnormal triggering conditions include, but are not limited to: the node's AliveStatus field is "abnormal" or the soft bus heartbeat packet is lost for more than 3 cycles; the node's Delay exceeds the threshold for 3 consecutive synchronization cycles or LinkQuality≤60 and cannot be recovered by link switching; the node's AgentCapability is updated to "partial failure". The re-election process includes marking abnormal nodes and pausing their subtasks, broadcasting an abnormal notification to the cluster via the A2A protocol ClusterSyncInfo, and selecting redundant nodes from the initial election cache or candidate set that match the capabilities, meet network requirements, and are currently idle (e.g., TaskLoad≤10%) for replacement. If there are no preset backup nodes, the candidate set is recalculated and the best one is selected. After activation, the task progress and execution parameters are sent out and the status is synchronized. To improve response speed, a pre-calculation and caching strategy is adopted during the initial election. During dynamic re-election, only the real-time network status is checked, so that the re-election response time can be controlled within about 2 seconds.

[0019] Furthermore, the Planning agent in the multi-intelligent task execution module is equipped with a task planning engine, which is responsible for receiving the initial task strategy issued by the decision module and breaking down the total task into sub-tasks that match the number of Working agents. Sub-tasks are issued according to the capability matching degree and network status of each Working agent. The Working agent is equipped with a task execution engine, which is responsible for receiving and executing sub-tasks and reporting the execution status in real time. The information synchronization within the cluster uses the ClusterSyncInfo extended field of the A2A protocol to transmit the network status and agent status to the local branch node, realizing real-time closed-loop feedback on task progress and node status.

[0020] Furthermore, the implementation method of the system includes the following steps: S1 Perception: Through the HarmonyOS distributed soft bus and A2A protocol extension, it collects and synchronizes the network parameters and capability status of each intelligent agent in the distributed autonomous network by combining periodic broadcasting and event-triggered updates, performs data verification, and generates a local global information table. S2 Decision: Based on the global information of S1, calculate the comprehensive fitness score for each candidate node according to the modeling and quantification method, and perform the initial election according to the capacity, network and scale constraints to generate the cluster configuration and assign the Planning / Working role. S3 execution: The Planning agent breaks down tasks and assigns them to the Working agent for execution. The Working agent executes the tasks according to the assignment and reports the status back in real time through ClusterSyncInfo. S4 Feedback and Dynamic Adaptation: When an abnormal event trigger condition is detected, dynamic re-election is triggered, and redundant nodes are selected to replace abnormal nodes according to the pre-calculated cache or candidate set, completing task handover and state synchronization, forming a dynamic adaptation closed loop.

[0021] In addition, the present invention also discloses a computer-readable storage medium storing computer-executable instructions that, when executed by one or more processors, cause the processors to perform the method of claim 8.

[0022] Based on the above technical solution, the present invention has the following technical effects: 1) This invention achieves efficient collaborative operation of multiple agents in complex distributed networks through a hierarchical three-level collaborative architecture of "perception-decision-execution" and deep integration of HarmonyOS soft bus and Google A2A protocol; 2) The proposed solution solves the technical problems of difficult dynamic construction of multi-agent clusters, unbalanced task scheduling, and slow response to anomalies, significantly improving the intelligence level and task execution efficiency of the system, and has broad application prospects and promotion value. Attached Figure Description

[0023] Figure 1 This is the overall architecture diagram of the multi-agent collaborative construction system based on HarmonyOS distributed soft bus of this invention.

[0024] Figure 2 This is a flowchart of the multi-agent collaborative construction method of the present invention, showing the key steps from device access to task completion.

[0025] Figure 3 This is a schematic diagram of the data structure mapping and extended fields for the distributed global network awareness module.

[0026] Figure 4 This is a schematic diagram of the two-stage election process and mathematical model for the distributed intelligent decision-making module.

[0027] Figure 5 This is a flowchart illustrating the task decomposition and execution feedback closed-loop process of the multi-intelligent task execution module. Detailed Implementation

[0028] To better illustrate the present invention, the following detailed description is provided in conjunction with specific embodiments. However, these specific embodiments are merely for illustrative purposes and are not intended to limit the scope of the invention.

[0029] This invention discloses a distributed multi-agent collaborative construction and task execution method based on a three-level collaborative architecture of "perception-decision-execution". This method achieves accurate construction, dynamic adaptation, and efficient task completion of the agent cluster through the orderly linkage of a distributed global network perception module, a distributed intelligent decision-making module, and a multi-agent task execution module. It possesses significant advantages such as rapid response, reliability, robustness, and optimized resource utilization.

[0030] The technical solution of this invention has the following advantages in terms of inventiveness and practicality: The innovative introduction of HarmonyOS distributed soft bus technology and Google A2A protocol adaptation and extension enables standardized perception and synchronization of global network status and capability information of intelligent agents within a distributed autonomous network, meeting the real-time data interaction needs across devices and levels.

[0031] Design and implement a two-stage election mechanism for the distributed intelligent decision-making module. Through mathematical modeling and quantitative scoring, unstructured network parameters and agent capabilities are transformed into a unified comprehensive fitness index, enabling dynamic division and redundancy replacement of cluster roles, and ensuring task continuity and stability.

[0032] By adopting a dual-role intelligent agent collaboration mode (Planning and Working intelligent agents) and combining a real-time information synchronization mechanism, closed-loop management of task planning, decomposition, allocation and execution is achieved, thereby improving the task execution efficiency and collaborative effectiveness of multi-agent intelligent agents.

[0033] This method is applicable to heterogeneous distributed network environments, dynamically responds to changes in node state and network topology adjustments, and has good scalability and robustness, making it suitable for multi-scenario intelligent agent collaborative applications.

[0034] The core of the technical solution of this invention lies in constructing a three-level collaborative architecture of "perception-decision-execution". Through the orderly linkage and data interaction of the three functional modules, the efficient collaborative task completion of multiple agents on distributed end devices can be achieved.

[0035] I. Distributed Global Network Perception Layer This module serves as the system's data foundation layer. It adopts the four-layer architecture of HarmonyOS distributed soft bus, namely "device discovery, connection management, data transmission, and status monitoring," and combines the adaptation and extension of the Google A2A protocol to realize the collection and synchronization of global network status and agent capability information in a distributed multi-agent environment.

[0036] Specific sensing information synchronization mechanisms include: Distributed autonomous broadcast synchronization: Each agent broadcasts the NetworkPerceptionInfo subfield set according to a preset period (default 8 seconds) using the adapted A2A protocol, transmitting node network and capability information. Fields such as NetID, MediumType, HopPath, and Delay in the broadcast packet directly correspond to network parameters in the HarmonyOS soft bus; AgentID, AgentCapability, and AliveStatus fields are generated using soft bus device information and agent capability status, ensuring the integrity and accuracy of the information.

[0037] Event-triggered update synchronization: When events such as network topology changes, link quality anomalies, or agent state changes occur, periodic broadcasting is immediately stopped, and incremental data push is triggered, carrying the changed fields and node identifiers, which significantly reduces network load and improves information response speed.

[0038] Information Verification and Integration: The receiving node uses the CheckSum field in the A2A protocol and the EncryptType field in the soft bus to verify data integrity and security, filtering out abnormal data. Successfully verified information updates the local global information table based on NetID and AgentID, ensuring consistency and real-time synchronization of information across the entire network.

[0039] II. Distributed Intelligent Decision-Making Layer Based on the standardized global information obtained from the perception module, this module designs a two-stage election mechanism for the construction of multi-agent clusters, realizing cluster construction, role division and dynamic adaptation.

[0040] Modeling and Quantization: A multi-dimensional quantitative mathematical model based on unstructured perceptual data is constructed to transform network state, agent capabilities, and survivability into a comprehensive fitness score. The fitness score integrates three core dimensions: link quality, capability matching, and survival stability, and its weights can be dynamically adjusted to meet different task requirements.

[0041] Initial election (task planning phase): A candidate set is formed through step filtering and sorting. The node with the highest comprehensive score is selected first to form a cluster. The cluster is divided into Planning agents (responsible for task planning and scheduling) and Working agents (execute sub-tasks). At the same time, a spare redundant node is assigned to each Working agent to ensure cluster redundancy.

[0042] Dynamic re-election (triggered by execution anomaly): When a node anomaly is detected (such as abnormal liveness status, degraded link performance, or capability failure), a fast re-election process is triggered. This process utilizes candidate node scoring data cached during the initial election, verifies only the real-time network status, and completes redundant node replacement and task handover within 2 seconds, ensuring continuous and stable task execution.

[0043] III. Multi-Intelligent Task Execution Module This module relies on the multi-agent cluster built by the aforementioned decision-making module to achieve dual-role division of labor and collaboration: The Planning agent is responsible for receiving the task strategies issued by the decision-making body, breaking down tasks and allocating sub-tasks, and dynamically scheduling Working agent resources.

[0044] The Working agent receives subtasks, performs specific operations, and provides real-time feedback on task status, network conditions, and its own survival information via the A2A protocol.

[0045] Achieve a closed-loop information system for task execution and perception decision-making, ensuring dynamic task adaptation and stable system operation.

[0046] The relationships between the modules are as follows: The distributed global network perception module (Network Perception Layer), serving as the data foundation layer, is responsible for collecting and synchronizing global network status and agent capability information within the distributed autonomous network. It outputs standardized information to the distributed intelligent decision-making module (Decision & Orchestration Layer). Based on the global information and task objectives provided by the perception module, the distributed intelligent decision-making module elects agent nodes and constructs a multi-agent cluster, generating cluster configurations and task execution strategies, which are then distributed to the multi-agent task execution module. The multi-agent task execution module, acting as the execution layer, relies on the division of labor and collaboration among dual-role agents within the cluster to execute specific tasks. Simultaneously, it provides real-time feedback on agent status, task progress, and other information to the perception and decision-making modules, forming a closed-loop collaborative link of "perception-decision-execution-feedback," ensuring dynamic adaptation and stable reliability of task execution.

[0047] Each module achieves cross-layer data interaction with Google's A2A protocol through HarmonyOS distributed soft bus, ensuring the real-time performance and standardization of information transmission.

[0048] Distributed global network perception layer This module adopts a technical approach of "HarmonyOS distributed soft bus architecture + Google A2A protocol adaptation and extension" to achieve global network perception and information synchronization among multiple intelligent agents in a distributed autonomous network. Specific technical implementations include: The HarmonyOS distributed soft bus technology architecture integrates a four-layer architecture of "device discovery - connection management - data transmission - status monitoring" as the underlying communication support. The core protocol fields and functions of each layer are shown in Table 1 below: Table 1 lists the protocol fields and corresponding functions of each layer of the HarmonyOS distributed soft bus architecture, including key parameters such as device unique identifier (DeviceID), connection type (ConnType), network topology information (TopoInfo), link delay (Delay), bandwidth (Bandwidth), and link quality (LinkQuality), providing comprehensive network status data support for upper-layer perception.

[0049] Table 1 HarmonyOS Distributed Soft Bus Technology Architecture Integration

[0050] Table 1 details the function and hierarchical distribution of each protocol field in the HarmonyOS distributed soft bus technology architecture, providing stable and efficient underlying communication support for the information collection and network awareness of this invention. Each field covers the entire process from device discovery to status monitoring, ensuring that the awareness module can accurately capture the global network and node status.

[0051] Adaptation and extension of Google A2A protocol: A new "NetworkPerceptionInfo" subfield set has been added to the reserved extension fields (field identifier "ExtensionField") of the Google A2A protocol to standardize the carrying of network perception information. The "NetworkPerceptionInfo" subfield set maps one-to-one with the perception parameters of the HarmonyOS distributed soft bus, as shown in Table 2 below: Table 2 shows the newly added "NetworkPerceptionInfo" subfield set in the "ExtensionField" reserved field of the Google A2A protocol. This field set is mapped one-to-one with the HarmonyOS soft bus perception parameters to form a standardized information carrier, realizing unified parsing and complete synchronization of perception data across protocols.

[0052] Table 2 shows the "NetworkPerceptionInfo" subfield set and the HarmonyOS distributed soft bus perception parameters.

[0053] The “NetworkPerceptionInfo” subfield set shown in Table 2 realizes the adaptation and extension of the Google A2A protocol to the HarmonyOS soft bus perception parameters. By mapping the fields one by one, it ensures the standardized transmission and parsing of network perception information, providing a unified, complete and real-time updated perception data foundation for subsequent decision-making modules.

[0054] Global information synchronization mechanism: A dual-mode synchronization strategy of "distributed autonomous broadcasting + event-triggered update" is adopted. The entire synchronization process relies on the HarmonyOS distributed soft bus protocol fields and A2A protocol extension fields to achieve standardized information transmission and verification, as specifically implemented as follows: 1) Distributed Autonomous Broadcast Synchronization: Within the distributed autonomous network, each agent node broadcasts to the entire network at a preset period (which can be dynamically adjusted according to the network size, with a default period of 8 seconds), using the adapted A2A protocol. The core content of the broadcast data packet is the "NetworkPerceptionInfo" subfield set extended by the A2A protocol. The information in this subfield set is filled with mapping relationships with the fields of the HarmonyOS distributed soft bus protocol: the "NetID", "MediumType", "HopPath", "Delay", "Bandwidth", and "LinkQuality" fields directly map to the core fields of the HarmonyOS distributed soft bus, such as network ID, connection type, topology information, delay, and bandwidth; the "AgentID", "AgentCapability", and "AliveStatus" fields are generated by combining the soft bus "DeviceID" and "DeviceType" fields with the agent's own capability status.

[0055] During the broadcast, the "DeviceID" field of the HarmonyOS distributed soft bus serves as the source identifier, ensuring that the receiving end can accurately identify the information source node. The "ConnType" field specifies the transmission medium used for the current broadcast, ensuring synchronization compatibility under different media.

[0056] 2) Event-Triggered Update Synchronization: When a preset event is triggered, periodic broadcasting is immediately terminated, and real-time update push is initiated. The triggering event is determined based on the state changes of the HarmonyOS Distributed Soft Bus protocol field and the A2A extension field, specifically including: ① Network topology change event: The soft bus “TopoInfo” field is updated (such as adding / removing devices, link switching), which synchronously triggers the A2A protocol “HopPath” and “MediumType” fields to be updated; ② Link quality degradation / recovery event: The soft bus “Delay”, “Bandwidth”, “LinkQuality” and “PacketLossRate” fields exceed / regress to preset thresholds (such as Delay>500ms, LinkQuality<60); ③ Agent state change event: The A2A extended fields “AgentCapability” (capability update) and “AliveStatus” (liveness status change) change.

[0057] After the event is triggered, the source node pushes incremental update data packets to the entire network via the A2A protocol. The data packets only contain the changed protocol fields and the corresponding "NetID" and "AgentID" identifier fields, reducing network transmission overhead.

[0058] 3) Information Verification and Integration: Each receiving node, through its built-in information verification unit, prioritizes parsing the "CheckSum" verification field of the A2A protocol and combines it with the "EncryptType" field of the HarmonyOS distributed soft bus to verify the security and integrity of data transmission, filtering out tampered or redundant data. After verification, the network affiliation is confirmed based on the "NetID", the corresponding intelligent agent is matched based on the "AgentID", and the synchronized protocol field information is updated to the local global information table to achieve accurate synchronization and consistency maintenance of information across the entire network. Finally, a standardized global information table is output to the distributed intelligent decision-making module.

[0059] Figure 1 The system architecture diagram of the multi-agent collaborative method and device based on HarmonyOS distributed soft bus clearly shows the four-layer hierarchical relationship of the present invention: "device layer - communication bearer layer - agent layer - collaborative control layer", as well as the core components and data flow of the three-level collaborative architecture of "perception-decision-execution".

[0060] Device layer: Includes heterogeneous physical nodes such as drones, unmanned vehicles, androids, and mobile tablets, which connect to the system via WiFi / satellite, radio frequency / Bluetooth, etc.

[0061] Communication bearer layer: Composed of HarmonyOS distributed soft bus architecture and Google A2A protocol adaptation layer, it realizes device discovery, connection management, data transmission and status monitoring, and extends network-aware information to A2A protocol fields through standardized mapping.

[0062] The intelligent agent layer, serving as the core collaborative hub, comprises a distributed global network perception module, a distributed intelligent decision-making module, and a multi-agent task execution module, forming a closed loop of "perception-decision-execution": the perception module generates a standardized global information table based on soft bus data; the decision-making module completes cluster configuration and task orchestration through mathematical modeling and election algorithms; and the execution module, through the collaboration of planning agents and working agents, achieves task decomposition, dynamic scheduling, and real-time feedback.

[0063] Collaborative Control Layer: Responsible for task assignment and result feedback, supporting specific business scenarios such as outdoor inspections, and enabling task parameter configuration, dynamic planning instruction assignment, task result aggregation, and anomaly alarms.

[0064] This architecture diagram intuitively presents the end-to-end hierarchical design of this invention, from the access of heterogeneous devices at the bottom layer to the collaboration of services at the upper layer, highlighting the core technology of the integration of HarmonyOS distributed soft bus and A2A protocol.

[0065] Figure 2 This is a flowchart of the multi-agent collaborative construction method of the present invention, which fully demonstrates the entire process logic and key technical nodes from device access to task completion. The process unfolds in five stages: initialization, synchronization, decision-making, execution, and termination. Device access and awareness initialization: Device access is completed through soft bus device discovery, identity authentication, link negotiation and parameter collection, and fault isolation is triggered when the multi-node network deteriorates.

[0066] Global information synchronization: Real-time synchronization of global network status is achieved through two methods: 8-second periodic broadcast and event-triggered update. Based on the A2A protocol, the NetworkPerceptionInfo field is extended and verified using CheckSum.

[0067] Decision cluster construction: Through steps such as constraint filtering, scoring and ranking, and redundancy supplementation, combined with the weights of core parameters, comprehensive scores and constraints, the decision cluster is dynamically constructed.

[0068] Dual-role collaborative execution: The Planning agent decomposes and schedules tasks globally, while the Working agent executes sub-tasks and reports resource status. When the network deteriorates, the load is too high, or the task changes, dynamic adjustments are triggered, and strategies such as task handover, link switching, or load balancing are executed. If necessary, the system is replanned.

[0069] Task completion: Summarize results, disband the cluster and switch states, and finally output structured data and logs, and feed the results back to the initial process to form a closed loop.

[0070] This flowchart clearly presents the end-to-end collaborative logic of this invention from device access to task completion, highlighting the core characteristics of dynamic decision-making and adaptive execution.

[0071] (ii) Distributed Intelligent Decision-Making Module (Decision & Orchestration Layer) The core of this module relies on standardized protocol fields (such as AgentCapability, Delay, etc.) synchronized by the distributed global network perception module. Through a progressive logic of "modeling quantification - two-stage election - configuration deployment", it achieves accurate construction and dynamic adaptation of multi-agent clusters. The core focus is on the quantitative support for modeling design and the process logic of the election algorithm. The specific implementation is as follows: The core of this module is based on standardized data (including protocol fields such as network status and agent capabilities) synchronized by a distributed global network perception module. With the goal of "optimal overall collaborative efficiency of the cluster," it completes the accurate construction, role allocation, and dynamic adaptation of a multi-agent cluster through a two-stage election mechanism: "initial election (task planning phase) - dynamic re-election (execution exception triggering)." Simultaneously, it generates standardized cluster configurations and ensures real-time synchronization of these configurations. The specific technical implementation is as follows: 2.1 Modeling Design: The Quantitative Core of Two-Stage Election A multi-dimensional decision-making mathematical model is constructed to transform unstructured data such as the network state of the perception module and the capabilities of the agent into quantifiable fitness indicators, providing a unified quantitative decision-making basis for initial election and dynamic re-election. The core formulas and principles are as follows: 2.1.1 Core Objective Function (Common Quantification Objective of Two-Stage Election) With the core objective of optimizing overall cluster collaboration efficiency, the goal is to maximize the sum of the overall suitability scores of the selected nodes, balancing task requirements with fault tolerance and redundancy. The formula and principle are as follows:

[0072] Principle explanation: Nodes are selected through decision variables, and the overall cluster score is calculated by combining the comprehensive adaptability of a single node to ensure that the selected cluster is the best in terms of capability, network, and stability. Here, represents the total number of candidate nodes (initially all nodes are aware, and dynamically the set of redundant nodes is re-elected), represents selected (initially added to the cluster, dynamically replaced abnormal nodes), and represents not selected.

[0073] 2.1.2 Single-Node Comprehensive Adaptability Scoring Formula (Quantifying Core Dimensions) By integrating three core dimensions—link quality, capability matching, and survival stability—and dynamically allocating weights according to task objectives, we achieve precise quantitative mapping of perceived data. The formula and principle are as follows:

[0074] Explanation of the principle: , , Weights for the three dimensions (satisfying a sum of 1, for real-time tasks) These are the standardized scores for link quality, capability matching, and survival stability (values ​​[0,1]). This is a fine-tuning coefficient; the core of this formula is to transform protocol fields such as Delay, AgentCapability, and AliveStatus of the perception module into unified quantitative indicators.

[0075] 2.1.3 Link Quality Quantification Formula (Detailed Network Dimensions) A standardized scoring formula is constructed for the network parameters of the sensing module to accurately quantify link reliability. The formula is:

[0076] 2.1.4 Constraints (Ensuring Cluster Validity) Four constraints are used to ensure that the election results meet the requirements for task execution. The core constraints and principles are as follows: Capability matching constraints: Extracting based on the AgentCapability field This ensures that the cluster possesses the core capabilities for the task. Network performance constraints: Verify and perceive the Delay and LinkQuality fields to ensure network compliance; Size and redundancy constraints: The number of core subtasks, This is to reserve resources for redundant nodes and for dynamic re-election.

[0077] 2.2.1 First Phase: Initial Election (Task Planning Phase) - Cluster Construction and Role Assignment The core objective of the initial election is to build a multi-agent cluster that meets the requirements of capability, network, and scale from all agent nodes synchronized by the perception module, based on task requirements, and to complete the role election of Planning and Working agents. This process involves the entire perception module's core protocol fields such as NetworkPerceptionInfo and AgentCapability. The specific steps are as follows: Step 1: Single-node capability modeling and data preprocessing. Extract all node protocol fields synchronized by the perception module (including AgentID, DeviceID, AgentCapability, network parameters, liveness status, etc.), and complete the modeling of each node based on a mathematical model. The (comprehensive fit score) is calculated to generate a "full node fit score table" as the basis for election data.

[0078] Step 2: Constraint Filtering to Generate a Candidate Set. Based on the constraints of the mathematical model (capability matching, network performance), all nodes are filtered: nodes whose AgentCapability does not meet the core capability requirements, whose Delay exceeds the threshold, or whose LinkQuality is below the minimum value are removed. Nodes that meet the conditions are retained to form a candidate set, and the candidate set information (including score, capability, and network status) is cached for subsequent dynamic re-election.

[0079] Step 3: Score Ranking and Node Selection. Ranking the candidate set nodes by... Sort the values ​​from highest to lowest, and select nodes sequentially until the condition is met. "Scale requirements (ensuring core task execution while reserving redundant nodes); if the candidate set size is insufficient, trigger an alert and relax the threshold for non-core networks (e.g.,..." (Appropriately improve), and re-select and supplement.

[0080] Step 4: Role Election and Division of Labor. According to... The scoring results are used to assign roles in the cluster: ① Select the 1-2 nodes with the highest overall scores as Planning agents, requiring them to have optimal network stability (LinkQuality≥90) and full capability adaptability. ① The selected nodes are responsible for task planning, decomposition, and dynamic scheduling; ② The remaining selected nodes are used as Working agents, and the core sub-tasks are matched according to the AgentCapability field to clarify the responsibilities of each node; ③ Each Working agent is bound to the second highest-scoring node in the candidate set as a redundant backup node, and the AgentID of the backup node is recorded and synchronized to the perception module for filing.

[0081] 2.2.2 Second Phase: Dynamic Re-election (Triggered by Abnormal Execution Phase) - Redundancy Replacement and Task Continuation Dynamic re-election is triggered by node anomaly events during the execution phase. The core objective is to quickly select redundant nodes to replace abnormal nodes, ensuring continuous task execution. Both the triggering conditions and the re-election process are based on protocol field data synchronized in real time by the perception module. The specific implementation is as follows: Abnormal Triggering Conditions: The Planning agent monitors the synchronization data of each node in the cluster in real time through the local branch node of the perception module. When any of the following abnormalities occurs (based on protocol fields), dynamic re-election is triggered: ① Abnormal liveness status: The node's AliveStatus field becomes "abnormal", or the soft bus heartbeat packet is lost for more than 3 cycles (timeout threshold); ② Network performance degradation: The node's Delay exceeds the threshold for 3 consecutive synchronization cycles, LinkQuality≤60 and cannot be recovered by link switching; ③ Capability failure: The node's AgentCapability field is updated to "partial failure" and cannot complete the current subtask.

[0082] Re-election Process: ① Abnormal Node Marking and Task Suspension: The Planning agent immediately marks the abnormal node's AgentID, suspends its current subtask, and broadcasts the abnormal notification to the entire cluster via the ClusterSyncInfo field of the A2A protocol; ② Redundant Node Screening: From the candidate set in the initial election cache, select backup nodes that meet the following conditions: a. Match the capabilities of the abnormal node. b. Network status meets standards , c. Currently in an idle state (TaskLoad≤10%); if there is no preset backup node, recalculate from the candidate set. Scoring and selecting the optimal node.

[0083] Alternate Execution and Task Handover: ① Redundant Node Activation: Activation commands are sent to selected redundant nodes via the ClusterSyncInfo field to synchronize the task progress and execution parameters (such as subtask objectives and data format requirements) of abnormal nodes; ② Task Handover: After the redundant node confirms activation, it updates its AgentCapability→TaskLoad field, takes over the subtasks of the abnormal node, and continues execution; ③ State Synchronization: The Planning agent synchronizes the re-election results (new node AgentID, role, and task allocation) to the perception module (updating the global information table) and all nodes in the cluster, completing the dynamic adaptation closed loop.

[0084] Algorithm optimization: To improve the efficiency of re-election, a "pre-computation + caching" strategy is adopted: During the initial election, the values ​​of all nodes in the candidate set are pre-computed. The scoring and capability matching matrix is ​​cached locally; during dynamic re-election, the cached data is directly called, and only the real-time network status (Delay, LinkQuality) is checked, so that the re-election response time is controlled within 2 seconds to avoid task execution interruption.

[0085] (II) Multi-intelligent task execution module This module is based on a multi-agent cluster built from a distributed intelligent decision-making module. Through the division of labor and cooperation between dual-role intelligent agents and a dynamic adaptation mechanism, it achieves efficient task execution. Specific technical implementations include: Cluster role configuration and division of labor: Based on the comprehensive scores and capabilities of each agent within the cluster, the decision module designates 1-2 nodes with the highest comprehensive scores and optimal network status as Planning agents, and the remaining nodes as Working agents. The Planning agent is equipped with a task planning engine, responsible for receiving the initial task strategy from the decision module, breaking down the overall task (into subtasks matching the number of Working agents), and allocating subtasks (based on the capability matching degree and network status of the Working agents). The Working agent is equipped with a task execution engine, responsible for receiving the subtasks from the Planning agent, executing specific operations, and providing real-time feedback on the execution status.

[0086] Real-time information synchronization mechanism: Within the cluster, each agent synchronizes two types of information in real time through local branch nodes of the distributed global network perception module: ① Network status information: including link latency, bandwidth, and communication quality between itself and other nodes in the cluster; ② Agent status information: including its own capacity utilization, task execution progress, and survival status (achieved through the heartbeat packet field "HeartBeat"). Synchronization information is transmitted through the extended field "ClusterSyncInfo" of the A2A protocol, ensuring the real-time nature and specificity of information synchronization.

[0087] Dynamic task planning and fault tolerance adaptation: The Planning agent monitors synchronization information in real time and triggers dynamic task planning when the following conditions occur: ① The working agent's network status deteriorates (e.g., latency exceeds the threshold, link interruption); ② The working agent's capacity utilization is too high (≥80%) or it experiences a failure (heartbeat packet loss exceeds 3 cycles); ③ Task execution progress fails to reach the preset threshold. At this point, the planning agent, based on the latest global awareness information, reallocates subtasks: for nodes with deteriorated network / failures, its unfinished subtasks are transferred to other working agents within the cluster that meet the score requirements; for nodes with excessive capacity utilization, some of their subtasks are split and moved to idle nodes to ensure load balancing and stable reliability of cluster task execution. The dynamic adjustment strategy is simultaneously fed back to the decision-making module to complete the update and record of the cluster status.

[0088] Figure 3This diagram illustrates the data structure mapping and extended fields of the distributed global network perception module, fully showcasing the core design logic of the integration and adaptation of the HarmonyOS distributed soft bus and the Google A2A protocol, as well as the end-to-end data processing flow of the perception module.

[0089] Figure 3 The left side vertically arranges the four-layer architecture of HarmonyOS distributed soft bus: device discovery layer, connection management layer, data transmission layer, and status monitoring layer, with the core protocol fields corresponding to each layer listed simultaneously. The right side shows the Google A2A protocol extension architecture, highlighting the newly added NetworkPerceptionInfo subfield set in the reserved ExtensionField field of the protocol. The one-to-one mapping relationship between the protocol fields of each layer of the soft bus and the A2A extension subfield is clearly marked by arrows, clarifying the standardized data carrier for dual protocol integration.

[0090] Figure 3 The lower center section uses arrows to connect and form a complete data processing chain, sequentially presenting the entire process from HarmonyOS soft bus data acquisition, A2A protocol extension mapping, information verification based on CheckSum and EncryptType fields, to the final output of a standardized global information table. This intuitively demonstrates the complete implementation path of the perception module from raw data acquisition to standardized decision data output.

[0091] Figure 4 This diagram illustrates the two-stage election process and mathematical model of the distributed intelligent decision-making module, fully demonstrating the core decision-making logic of this module: "modeling and quantification - two-stage election". Figure 4 The upper center section contains the mathematical model area, which includes four core quantitative models: the cluster comprehensive collaborative efficiency objective function, the single node comprehensive adaptability scoring formula, the link quality quantitative formula, and four major constraints: capability matching, network performance, scale, and redundancy. These provide a unified quantitative basis for the election process. The lower section outlines the two-stage election process, divided into left and right parts. Initial election process: The initial construction and role assignment of the cluster are completed by following the steps of "data preprocessing → constraint filtering to generate candidate set → scoring and sorting node selection → role election and division of labor → outputting cluster configuration".

[0092] Dynamic re-election process: Triggered when network degradation, excessive load or other anomalies are detected. Through the steps of "anomaly marking and broadcasting → redundant node screening → alternative execution and task handover → state synchronization", the cluster can be dynamically adapted. At the same time, the "pre-computation + caching" mechanism ensures that the response time is ≤2s.

[0093] The two-stage process enables data reuse through "candidate set caching," providing an intuitive overview of the complete decision-making chain from cluster construction to anomaly adaptation.

[0094] Figure 5 The flowchart for the task decomposition and execution feedback closed loop of the multi-intelligent task execution module fully demonstrates the internal operating logic and cross-module linkage of this module. Figure 5 The document clarifies the upstream input link from which the distributed intelligent decision-making module issues cluster configuration and task execution strategies to the Planning agent within this module; it clearly presents the dual-role division of labor between the Planning agent and multiple Working agents, where the Planning agent completes the overall task decomposition and sub-task allocation, and the Working agents execute the corresponding sub-tasks; it also demonstrates the bidirectional feedback link where the Working agents use the ClusterSyncInfo field of the A2A protocol to provide real-time feedback on execution status to the Planning agent and synchronize node status with the distributed global network perception module, as well as the dynamic task planning and adjustment mechanism of the Planning agent, fully presenting the closed-loop operation logic of the execution layer: "task issuance - decomposition and execution - status feedback - dynamic adaptation".

[0095] Innovative method for constructing multi-agent systems in distributed autonomous network environments: Breaking through the limitations of traditional multi-agent construction relying on centralized control, this invention constructs a three-level distributed collaborative architecture of "perception-decision-execution". It achieves autonomous synchronization of distributed node information through global network perception and completes the autonomous election and construction of the agent cluster based on distributed decision-making algorithms. It can realize the collaborative networking and task collaboration of multi-device agents without the need for centralized node management and control. It is adapted to the core characteristics of distributed autonomous networks, which are decentralized and have dynamic topology, and fills the technical gap in the collaborative construction of multi-agents in the distributed autonomous network environment.

[0096] HarmonyOS's distributed soft bus and Google's A2A protocol innovatively integrate to break through the bottleneck of heterogeneous device networking: This innovative approach combines the self-discovery, self-connection, and multi-media adaptation capabilities of HarmonyOS's distributed soft bus with the standardized interaction advantages of Google's A2A protocol. By extending the reserved fields of the A2A protocol to carry network-aware information, it achieves complementary advantages between the two technologies. This effectively solves the problems of poor adaptability and complex connections of traditional TCP / IP protocols in heterogeneous device networking, such as drones and autonomous vehicles. It realizes dynamic adaptation and standardized information exchange across multi-media transmission links, significantly improving the discovery efficiency and network stability of intelligent agents among heterogeneous devices.

[0097] Dynamic fault-tolerant adaptation mechanism improves task execution reliability: By employing a dual-role division of labor between Planning and Working agents, combined with a real-time network and state synchronization mechanism within the cluster, the system can dynamically detect network fluctuations, device failures, and changes in capacity utilization of the Working agent, quickly triggering task reallocation, link switching, or load balancing adjustments. Compared to traditional fixed task allocation modes, this mechanism effectively avoids the impact of device failures and network degradation on task execution, significantly improving the anti-interference capability and continuous execution capability of multi-agent collaborative tasks in complex edge environments.

[0098] The following are two detailed application test examples to verify the effectiveness of the invention technology, which are designed for the "distributed global network perception module" and the "distributed intelligent decision-making module + multi-intelligent task execution module" respectively, to ensure that the effectiveness and advantages of the proposed technical solution can be fully verified in the multi-agent collaborative construction system.

[0099] Experiment Example 1: Performance Verification of Distributed Global Network Perception Module in Multi-Agent Complex Network Environment Objective: To verify the core technical performance of a distributed global network awareness module adapted and extended with the HarmonyOS distributed soft bus architecture and Google A2A protocol in a complex heterogeneous multi-agent network environment, specifically including: The effectiveness of the dual-mode global information synchronization mechanism (periodic distributed autonomous broadcasting + event-triggered incremental update); the accuracy of the mapping and matching between the A2A protocol extended fields and the core fields of the HarmonyOS distributed soft bus; the data integrity and security verification capabilities of the information verification and integration unit; the module's perception accuracy, synchronization real-time performance, transmission overhead optimization effect, and long-term operational stability under dynamic heterogeneous networks.

[0100] Experimental environment and conditions: Number of multi-agent nodes: 50, covering heterogeneous devices such as mobile robots, drones, and sensor nodes.

[0101] Network environment: Includes three different transmission media: Wi-Fi, Bluetooth, and 5G. The network topology changes dynamically, with a maximum link latency of 800ms and a link quality range of 10~100.

[0102] Software environment: Deploy the HarmonyOS distributed soft bus communication framework and integrate the adapted Google A2A protocol extension.

[0103] Experimental tools: A network state simulator is used to inject network topology changes and link degradation events; an agent state monitor is used to simulate survival status and capability fluctuations.

[0104] Experimental steps: 1) Normal steady-state synchronization test: Maintain a stable network environment (no topology changes, no drastic fluctuations in link parameters), with all nodes running continuously for 30 minutes; during each broadcast cycle, collect the local global information tables of all nodes across the network and verify the following core indicators: ① Information synchronization accuracy: Calculate the consistency of NetID, AgentID, AgentCapability, and core network parameter fields across all network nodes, and calculate the ratio of the number of consistent fields to the total number of fields; ② Average synchronization latency: Based on the broadcast source node's sending timestamp, calculate the average time taken for all nodes across the network to complete receiving, verifying, and updating their local global information tables; ③ Network occupancy rate: Calculate the proportion of total traffic generated during broadcast synchronization to available network bandwidth, and record the average network occupancy rate per unit time; 3. During the test, record whether any abnormal situations such as broadcast interruption, data parsing errors, or information asynchrony occur.

[0105] 2) Event-triggered incremental update test: Network topology change event test: ① Randomly inject one node addition event (adding 2 heterogeneous nodes) and one node removal event (taking 2 online nodes offline), triggering the TopoInfo and HopPath fields to be updated; ② Record the total time from the time the event occurs to the completion of the local global information table update by all nodes in the network, i.e., the event trigger update response time; ③ Verify the content of the incremental update data packet to confirm that it only contains the changed fields and the corresponding NetID and AgentID identifiers, and calculate the ratio of the incremental data packet size to the full broadcast data packet size; ④ Verify the consistency of information after the update of all nodes in the network, and calculate the incremental update accuracy.

[0106] Link quality degradation event test: ① Randomly select 5 online nodes, inject a link degradation event, make its one-way Delay > 500ms and LinkQuality < 60, and continue for 3 broadcast cycles to trigger the link status update event described in the patent; ② Record the event trigger update response time and verify the synchronization accuracy of the link degradation information of all nodes in the network; ③ After the link degradation is resolved, verify the effectiveness of the event trigger update mechanism after the node link parameters are restored, and record the response time of the restoration synchronization.

[0107] Agent state change event test: ① Randomly select 3 online nodes, modify their AgentCapability field (simulate capability update / partial failure) and AliveStatus field (simulate abnormal survival status) to trigger the agent state update event described in the patent; ② Record the event trigger update response time and verify the synchronization accuracy of agent state information of all network nodes; ③ Verify whether the system automatically stops periodic broadcasts and pushes incremental updates when the event is triggered to confirm the effectiveness of the mechanism.

[0108] 3) Security and data integrity testing: Data packet tampering test: For the broadcast data packets of the test group, the core parameters in the NetworkPerceptionInfo subfield set are randomly tampered with and the CheckSum verification field is modified. The tampered data packets are then sent to the entire network. Replay attack test: Capture historical valid broadcast data packets and replay them to the entire network after 3 broadcast cycles to simulate a replay attack; To address the two types of attacks mentioned above, the abnormal identification rate and malicious data packet filtering rate of the statistical node information verification unit are used to verify whether the joint verification mechanism of the CheckSum field and EncryptType field described in the patent is effective and whether abnormal data is written to the local global information table. The average time taken during the verification process is recorded synchronously to verify the efficiency of the verification mechanism.

[0109] Initialization Phase: All nodes power on and start up, completing hardware initialization and protocol stack loading; nodes start the HarmonyOS distributed soft bus, and the four-layer architecture completes automatic device discovery, multi-media connection management, and encryption authentication, completing the basic networking of all nodes in the network; nodes load the A2A protocol extension adapter program, complete the mapping and binding of the NetworkPerceptionInfo subfield set with the HarmonyOS soft bus core fields, configure dual-mode synchronization parameters (default broadcast period of 8 seconds), and enable the information verification and integration unit; all nodes start periodic broadcasting, sending extended A2A protocol broadcast data packets to the entire network, completing the synchronous construction of the initial global information table.

[0110] Evaluation metrics: Information synchronization accuracy: ≥99.8%; Average synchronization latency of periodic broadcasts: ≤8 seconds (meeting the default broadcast cycle); Event-triggered update response time: ≤500ms; Network occupancy reduction: ≥15% (compared to the non-layered dual-mode synchronization scheme); Security filtering accuracy: 100%, all malicious or redundant data must be discarded in a timely manner.

[0111] Expected test results: All evaluation indicators have reached the above-mentioned preset thresholds, the dual-mode synchronization mechanism is operating stably, the mapping accuracy between the A2A extended field and the HarmonyOS soft bus field is 100%, the information verification and integration unit can completely intercept abnormal data, and achieve standardized, high-accuracy, and low-latency synchronization of global network status and agent capability information in heterogeneous distributed networks.

[0112] This experiment verifies that the "HarmonyOS distributed soft bus + A2A protocol extension and adaptation" technical path described in this patent can stably achieve real-time synchronization of network status and node capability information among multiple agents, while taking into account rapid update response to event triggers, significantly optimizing network resource consumption, ensuring data security and accuracy, and fully meeting the core underlying perception requirements of multi-agent collaboration.

[0113] Expected scenario adaptability: This module can stably achieve global information synchronization in complex end-side scenarios with multi-media heterogeneous networks and dynamic topology changes, fully meeting the underlying perception requirements of multi-agent collaboration of distributed end devices such as drones and unmanned aerial vehicles.

[0114] Experiment Example 2: Dynamic Construction and Task Execution Verification of a Multi-Agent Cluster Based on a Distributed Intelligent Decision-Making Module Objective: To verify the core technical performance of the distributed intelligent decision-making module and its supporting multi-intelligent task execution module, specifically including: the ability to accurately construct, role-based and dynamically adapt multi-agent clusters based on the "two-stage election" algorithm; the task collaborative execution effect of the Planning and Working dual-role intelligent agents; and the system's response speed, task continuation capability and operational stability in the face of sudden node anomalies.

[0115] Experimental Environment and Conditions: Multi-agent Scale: 40 nodes, including a robot swarm with varying planning and execution capabilities; moderate task complexity. Task Type: Warehouse picking and delivery task, involving path planning, loading, and transfer sub-tasks. The task is decomposed into multiple sub-tasks corresponding to multiple working nodes. Simulation of dynamic changes in network and sensing data, such as link latency jitter, changes in node liveness status, and fluctuations in node capabilities. Implementation Environment: Fully integrated distributed global network sensing module, intelligent decision-making module, and multi-agent task execution module.

[0116] Evaluation metrics: Initial election completion time: ≤15 seconds; Cluster role assignment accuracy: ≥99%; Task execution success rate: ≥98%; Dynamic re-election replacement latency: ≤2 seconds; Task continuation success rate: ≥97%; Redundant node resource utilization improvement: ≥25%.

[0117] Expected test results: Phase 1: Initial Election and Cluster Construction Verification System initialization: All nodes are powered on and started. The distributed global network perception module completes the synchronization of information of all nodes in the network and outputs a standardized global information table to the decision module.

[0118] Scoring and candidate set selection: The decision module calculates a comprehensive suitability score for each node based on the global information table and a three-dimensional quantification model. It filters nodes based on capability matching and network performance constraints, retains nodes that meet the conditions to form a candidate set, and caches the relevant data. Node selection and role election: Sort the candidate set in descending order of comprehensive score, select nodes that meet the scale and redundancy requirements to build a cluster, select 1-2 nodes with the highest comprehensive score and LinkQuality≥90 as Planning agents, and the remaining selected nodes as Working agents, and bind 1 redundant spare node to each Working agent. Task decomposition and allocation: The Planning agent receives the task strategy, decomposes the total task into sub-tasks that match the number of Working agents, and allocates the sub-tasks according to their capability matching degree. Data monitoring: Record the total time from the output of the global information table by the perception module to the completion of cluster configuration and task allocation, i.e. the initial election completion time, and verify the cluster role division and constraint conditions are met.

[0119] Phase Two: Dual-role Collaborative Task Execution and Closed-Loop Feedback Testing Task execution: The Working agent receives subtasks and performs specific operations. Through the ClusterSyncInfo extended field of the A2A protocol, it transmits task execution progress, node load, network status, and survival status information back to the Planning agent and the perception module in real time, forming the closed-loop feedback link described in the patent. Dynamic scheduling: Based on real-time feedback information, the Planning agent splits and redistributes subtasks for the Working agent with a load rate ≥80%, thereby achieving cluster load balancing. Data monitoring: Record the entire process of task execution progress, subtask completion rate, node load balancing rate, and status feedback closed-loop delay until the overall task is completed.

[0120] Phase 3: Verification of Dynamic Re-election and Redundancy Replacement Mechanism Abnormal event injection: According to the abnormal triggering conditions described in the original patent, three types of node abnormalities are injected respectively: ① Abnormal survival status: randomly select 2 Working agents and cause their heartbeat packets to be lost for more than 3 cycles; ② Network performance degradation: randomly select 2 Working agents and cause their Delay to exceed the threshold for 3 consecutive cycles, LinkQuality≤60 and cannot be recovered; ③ Capability failure: randomly select 1 Working agent and update its AgentCapability field to "partial failure"; Anomaly identification and response: Record the total time from the occurrence of an anomaly event to the Planning agent marking the anomaly node, pausing the subtask, and broadcasting the anomaly notification to the entire network; Re-election and Task Continuation: Verify that the Planning agent triggers a dynamic re-election process, selects redundant nodes that meet the conditions from the pre-cached candidate set to complete the replacement, records the total time from abnormal triggering to redundant node activation, synchronization of task progress, and takeover of subtasks, and verifies whether it meets the requirement of ≤2 seconds as described in the patent. Data monitoring: Record task continuation success rate, task interruption duration, and total task completion rate.

[0121] Phase 4: Redundancy Mechanism and System Robustness Verification Continuous anomaly scenario test: During task execution, five node anomaly events (covering three types of anomalies: survival, network, and capability) are injected in three batches, with a one-minute interval between each batch, to simulate continuous failure of multiple nodes under complex dynamic scenarios. Robustness verification: Monitor whether the system can stably complete dynamic re-election and task continuation under continuous abnormal scenarios, and record the cluster configuration update accuracy, continuous runtime without task interruption, and total task completion rate; Resource utilization verification: Statistical analysis of redundant node call rate, task takeover success rate, and idle rate to verify the resource utilization effect of the redundancy mechanism described in the patent.

[0122] All evaluation indicators have reached the above-mentioned preset thresholds: the three-dimensional quantitative model can accurately achieve standardized scoring of node adaptability; the two-stage election mechanism can quickly complete the construction and role division of the cluster that meets the constraints; the dynamic re-election mechanism can complete the redundant replacement of abnormal nodes within 2 seconds; and the dual-role collaboration mechanism of Planning and Working can form a stable closed-loop collaborative link of "perception-decision-execution-feedback".

[0123] This experiment verifies that the distributed intelligent decision-making module can achieve optimal cluster construction and role allocation based on global perception data, and has excellent dynamic adaptability and efficient anomaly response capabilities during task execution. The redundancy mechanism can effectively improve the system robustness, ensure the stable operation of the multi-agent system and the continuous progress of tasks, and fully meet the core application requirements of multi-agent collaboration of distributed end devices.

[0124] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A multi-intelligent collaborative device for distributed terminal devices, characterized in that, The system includes: The distributed global network perception module is used to collect and synchronize global network status and agent capability information within the distributed autonomous network, and output standardized information to the distributed intelligent decision-making module. The distributed intelligent decision-making module is used to complete the election of intelligent agent nodes and the construction of a multi-agent cluster based on the global information and task objectives provided by the perception module, according to the preset modeling and election logic, and generate cluster configuration and task execution strategy and distribute them to the multi-intelligent task execution module. The multi-intelligent task execution module is used to execute specific tasks based on the cluster configuration and strategy issued by the decision module, relying on the division of labor and cooperation of dual-role (Planning and Working) intelligent agents within the cluster, and to feed back the status of the intelligent agents and the task progress to the perception module and the decision module in real time, forming a closed-loop collaborative link of "perception-decision-execution-feedback". The modules communicate with each other through the HarmonyOS distributed soft bus and the Google A2A protocol adaptation extension to achieve cross-layer data interaction and standardized information transmission.

2. The multi-intelligent collaborative device for distributed terminal devices as described in claim 1, characterized in that: The distributed global network awareness module uses the HarmonyOS distributed soft bus's four-layer architecture of "device discovery - connection management - data transmission - status monitoring" as its underlying communication support, and adopts a "dual-mode" global information synchronization mechanism, including: periodic distributed autonomous broadcast synchronization and instant-triggered update synchronization based on preset events; the broadcast period can be dynamically adjusted according to the network scale, with a default value of 8 seconds; instant-triggered updates stop periodic broadcasts when a trigger event occurs and push incremental update data packets containing only changed fields to the entire network.

3. The multi-intelligent collaborative device for distributed terminal devices as described in claim 1, characterized in that: The distributed global network awareness module periodically broadcasts data packets using the NetworkPerceptionInfo subfield set defined by the A2A protocol extension, and maps the following fields to the HarmonyOS distributed soft bus protocol fields: NetID, MediumType, HopPath, Delay, Bandwidth, and LinkQuality are mapped to core fields of the soft bus such as network ID, connection type, topology information, delay, bandwidth, and link quality. AgentID, AgentCapability, and AliveStatus are combined with DeviceID, DeviceType from the soft bus and the agent's own state to generate the data. During broadcasting, the DeviceID of the soft bus is used as the source identifier, and ConnType specifies the transmission medium; When an event is triggered, only the changed protocol fields are pushed, along with the corresponding NetID and AgentID identifiers, to reduce communication overhead.

4. The multi-intelligent collaborative device for distributed terminal devices as described in claim 1, characterized in that: The distributed global network awareness module also includes an information verification and integration unit, which is used to first parse the CheckSum verification field of the A2A protocol and combine it with the EncryptType field of the HarmonyOS distributed soft bus to verify data integrity and security. After the verification is passed, the network ownership is confirmed according to NetID and the local global information table is updated according to AgentID. Finally, a standardized global information table is output to the distributed intelligent decision-making module.

5. The multi-intelligent collaborative device for distributed terminal devices as described in claim 1, characterized in that: The distributed intelligent decision-making module implements the construction and dynamic adaptation of multi-agent clusters using the logic of "modeling and quantification - two-stage election - configuration deployment", specifically including: A mathematical model is constructed to quantify data such as network status and agent capabilities provided by the perception module into a single-node comprehensive fitness score. The single-node score integrates three dimensions: link quality, capability matching, and survival stability, and dynamically allocates weights according to task requirements (the sum of each weight is 1). The mathematical model for the single-node comprehensive fitness score is implemented through the following formula: The standardized link quality score Snet,i is calculated using the following formula: With the goal of achieving optimal overall cluster collaborative efficiency, the core objective function for achieving optimal overall cluster collaborative efficiency is: By setting capability matching constraints, network performance constraints, and scale and redundancy constraints, the election results are ensured to meet the task execution requirements through these four constraints. The core constraints and principles are as follows: Capability matching constraints: Extracting based on the AgentCapability field This ensures that the cluster possesses the core capabilities for the task. Network performance constraints: Verify and perceive the Delay and LinkQuality fields to ensure network compliance; Size and redundancy constraints: The number of core subtasks, This is to reserve resources for redundant nodes and for dynamic re-election. Based on the constraints, during the initial election phase, a candidate set is selected from all sensing nodes, sorted by score, and nodes that meet the size and redundancy requirements are selected to build a cluster. In the initial election, based on the scoring results, 1 to 2 nodes with the highest overall score and excellent network stability (e.g., LinkQuality≥90) are selected as Planning agents, and the rest are selected as Working agents. Each Working agent is then bound to a backup node with the second highest score in the candidate set as redundancy.

6. The multi-intelligent collaborative device for distributed terminal devices as described in claim 5, characterized in that: The distributed intelligent decision-making module also includes a dynamic re-election mechanism: When any of the abnormal triggering conditions are met, the Planning agent will trigger a re-election. The abnormal triggering conditions include, but are not limited to: the node's AliveStatus field is "abnormal" or the soft bus heartbeat packet is lost for more than 3 cycles; the node's Delay exceeds the threshold for 3 consecutive synchronization cycles or LinkQuality≤60 and cannot be recovered by link switching; the node's AgentCapability is updated to "partial failure". The re-election process includes marking abnormal nodes and pausing their subtasks, broadcasting an abnormal notification to the cluster via the A2A protocol ClusterSyncInfo, and selecting redundant nodes from the initial election cache or candidate set that match the capabilities, meet network requirements, and are currently idle (e.g., TaskLoad≤10%) for replacement. If there are no preset backup nodes, the candidate set is recalculated and the best one is selected. After activation, the task progress and execution parameters are sent out and the status is synchronized. To improve response speed, a pre-calculation and caching strategy is adopted during the initial election. During dynamic re-election, only the real-time network status is checked, so that the re-election response time can be controlled within about 2 seconds.

7. The multi-intelligent collaborative device for distributed terminal devices as described in claim 1, characterized in that: The Planning agent in the multi-intelligent task execution module is equipped with a task planning engine, which is responsible for receiving the initial task strategy issued by the decision module and breaking down the total task into sub-tasks that match the number of Working agents, and issuing sub-tasks according to the capability matching degree and network status of each Working agent. The Working agent is equipped with a task execution engine, which is responsible for receiving and executing subtasks and reporting the execution status in real time. The ClusterSyncInfo extended field of the A2A protocol is used to transmit network status and agent status to local branch nodes, realizing real-time closed-loop feedback on task progress and node status.

8. A method for implementing the system according to any one of claims 1 to 7, characterized in that, Includes the following steps: S1 Perception: Through the HarmonyOS distributed soft bus and A2A protocol extension, it collects and synchronizes the network parameters and capability status of each intelligent agent in the distributed autonomous network by combining periodic broadcasting and event-triggered updates, performs data verification, and generates a local global information table. S2 Decision: Based on the global information of S1, calculate the comprehensive fitness score for each candidate node according to the modeling and quantification method, and perform the initial election according to the capacity, network and scale constraints to generate the cluster configuration and assign the Planning / Working role. S3 execution: The Planning agent breaks down tasks and assigns them to the Working agent for execution. The Working agent executes the tasks according to the assignment and reports the status back in real time through ClusterSyncInfo. S4 Feedback and Dynamic Adaptation: When an abnormal event trigger condition is detected, dynamic re-election is triggered, and redundant nodes are selected to replace abnormal nodes according to the pre-calculated cache or candidate set, completing task handover and state synchronization, forming a dynamic adaptation closed loop.

9. A computer-readable storage medium storing computer-executable instructions that, when executed by one or more processors, cause the processors to perform the method of claim 8.

Images